Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Simultaneous imaging of structural and functional plasticity in the awake brain Podgorski, Kaspar Jan 2015

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


24-ubc_2015_may_podgorski_kaspar.pdf [ 16.73MB ]
JSON: 24-1.0166144.json
JSON-LD: 24-1.0166144-ld.json
RDF/XML (Pretty): 24-1.0166144-rdf.xml
RDF/JSON: 24-1.0166144-rdf.json
Turtle: 24-1.0166144-turtle.txt
N-Triples: 24-1.0166144-rdf-ntriples.txt
Original Record: 24-1.0166144-source.json
Full Text

Full Text

Simultaneous imaging of structural and functionalplasticity in the awake brainbyKaspar Jan PodgorskiB.Sc. Neuroscience, Cognitive Science and Artificial Intelligence,University of Toronto, 2008A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Neuroscience)The University of British Columbia(Vancouver)April 2015c© Kaspar Jan Podgorski, 2015AbstractDuring learning, and particularly during development, neurons in the brain undergostructural and functional changes that are intricately interrelated. This plasticity isguided by patterns of activity that encode information about the environment, al-lowing the brain to adapt to an organism’s specific experiences. Here I developedoptical methods and analysis tools to measure and analyze sensory-evoked activitypatterns in the awake brain, and track how sensory information guides plastic-ity. Several different methods and their applications are presented. I describedmodels and analysis tools for nonlinear decoding of somatic activity patterns inpopulations of neurons, and used them to track functional reorganization of neu-ral circuits during training. I identified a group of ultrabright and stable organicdyes that enable two-photon imaging deep within living tissue, and applied themto produce a sensitive intracellular label for excitatory synapses. I developed arandom access microscope capable of tracking activity at all excitatory synapseson a neuron simultaneously, enabling the first comprehensive measurements of asingle neuron’s dendritic input and firing output within the awake brain. I used thismicroscope to track neurons’ comprehensive activity and structural changes acrossplasticity-inducing training, and identified rules by which somatic and dendritic ac-tivity direct the detailed growth patterns of dendrites, producing spatially clusteredinput patterns along neurons’ dendritic arbor. Throughout this work, I’ve taken ad-vantage of the Xenopus laevis model system to observe rapid experience-dependentplasticity in the awake, developing brain. These results demonstrate ways in whichspecific experiences direct the detailed connectivity of developing neural circuits.iiPrefaceThe work in Chapter 3 has previously been published:Podgorski, Dunfield, and Haas [205]D. Dunfield and I contributed equally to this work. I designed and performedimaging experiments, performed electrophysiology, developed models, performedanalyses, and prepared text and figures. Derek Dunfield designed and performedimaging experiments, developed imaging methods and protocols, and preparedtext. Prof. K. Haas provided advice on experimental design and prepared text.The work in Chapter 5 has previously been published:Podgorski, Terpetschnig, Klochko, Obukhova, and Haas [206]I characterized two-photon cross-sections, synthesized synaptic labels, per-formed imaging experiments, and prepared the text. E. Terpetschnig, O. Klochko,and O. Obukhova synthesized and characterized fluorophores. Prof. K. Haas as-sisted in preparing text.The work in Chapter 6 has been published inPodgorski and Haas [204]I designed and performed all experiments and prepared the text. Prof. K. Haasassisted in preparing text.These works are reproduced here with permission of the copyright holder. Allstudies were performed with the approval of the UBC Animal Care Committee(certificate - A09-0021).iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1.1 Learning in the brain . . . . . . . . . . . . . . . . . . . . 11.1.2 Learning in machines . . . . . . . . . . . . . . . . . . . . 31.2 Monitoring activity patterns in the brain . . . . . . . . . . . . . . 41.2.1 Introduction to linear and nonlinear processes . . . . . . . 41.2.2 Linear and nonlinear biological processes . . . . . . . . . 61.2.3 Population coding . . . . . . . . . . . . . . . . . . . . . 71.2.4 Dendritic integration . . . . . . . . . . . . . . . . . . . . 101.3 Microscope development . . . . . . . . . . . . . . . . . . . . . . 111.3.1 A brief history of imaging . . . . . . . . . . . . . . . . . 111.3.2 Imaging in the awake brain . . . . . . . . . . . . . . . . . 141.3.3 Ultrafast microscopy for synaptic imaging . . . . . . . . . 16iv1.4 Brain development . . . . . . . . . . . . . . . . . . . . . . . . . 201.4.1 The Xenopus laevis model system . . . . . . . . . . . . . 201.4.2 Activity-dependent structural development and early learning 231.4.3 Experience-dependent functional plasticity . . . . . . . . 251.5 Scientific motivations . . . . . . . . . . . . . . . . . . . . . . . . 272 Introduction to Manuscript 1 . . . . . . . . . . . . . . . . . . . . . . 283 Manuscript 1: Functional Clustering in the Tectum . . . . . . . . . 303.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.3.1 Monitoring neuronal firing with calcium imaging . . . . . 333.3.2 Noise correlations . . . . . . . . . . . . . . . . . . . . . 343.3.3 Noise correlations impair network performance . . . . . . 383.3.4 Visual training improves network encoding . . . . . . . . 393.3.5 Effects of NMDAR blockade on basal responses . . . . . 413.3.6 NMDARs mediate experience-driven network plasticity . 423.3.7 Specificity of training-induced plasticity . . . . . . . . . . 433.3.8 Plasticity is spatially structured . . . . . . . . . . . . . . 433.3.9 Network re-organization drives encoding improvements . 453.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.5 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533.5.1 Animal rearing conditions . . . . . . . . . . . . . . . . . 533.5.2 Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . 533.5.3 Visual stimulation . . . . . . . . . . . . . . . . . . . . . 533.5.4 Two-photon-guided patch recording . . . . . . . . . . . . 543.5.5 Fluorescence data processing . . . . . . . . . . . . . . . . 553.5.6 Single-neuron properties . . . . . . . . . . . . . . . . . . 563.5.7 Network properties . . . . . . . . . . . . . . . . . . . . . 573.5.8 Decoding algorithms . . . . . . . . . . . . . . . . . . . . 583.5.9 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . 603.5.10 Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 60v3.6 Supplementary figures . . . . . . . . . . . . . . . . . . . . . . . 614 Introduction to Manuscript 2 . . . . . . . . . . . . . . . . . . . . . . 715 Manuscript 2: Ultra-Bright Two-Photon Fluorophores . . . . . . . . 735.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 755.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805.5 Materials and methods . . . . . . . . . . . . . . . . . . . . . . . 815.5.1 Fluorescent labels . . . . . . . . . . . . . . . . . . . . . . 815.5.2 Animal rearing conditions . . . . . . . . . . . . . . . . . 815.5.3 Two-photon action cross-sections . . . . . . . . . . . . . 815.5.4 Labelled dextran synthesis and single-cell electroporation 835.5.5 In vivo imaging and photobleaching . . . . . . . . . . . . 835.5.6 Targeted electroporation and PSD-95 labelling . . . . . . 845.5.7 Immunohistochemistry . . . . . . . . . . . . . . . . . . . 846 Summary of Manuscript 3 . . . . . . . . . . . . . . . . . . . . . . . 876.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 876.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887 Introduction to Manuscript 4 . . . . . . . . . . . . . . . . . . . . . . 917.1 Relating neuronal structure and function . . . . . . . . . . . . . . 917.2 Summary of results . . . . . . . . . . . . . . . . . . . . . . . . . 927.3 Dynamic morphometrics . . . . . . . . . . . . . . . . . . . . . . 937.4 Random access microscope design . . . . . . . . . . . . . . . . . 948 Manuscript 4: Comprehensive Dendritic Imaging . . . . . . . . . . 968.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 968.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 978.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 988.3.1 Dual in vivo imaging of sensory-driven growth and so-matic firing . . . . . . . . . . . . . . . . . . . . . . . . . 98vi8.3.2 Correlated structural and functional plasticity . . . . . . . 998.3.3 Experience-induced functional plasticity is upstream of struc-tural plasticity . . . . . . . . . . . . . . . . . . . . . . . 1018.3.4 Hotspots of structural plasticity are mediated by intracel-lular signals . . . . . . . . . . . . . . . . . . . . . . . . . 1048.3.5 2D synaptic imaging of NMDA currents . . . . . . . . . . 1048.3.6 Comprehensive imaging . . . . . . . . . . . . . . . . . . 1078.3.7 Sensory-evoked dendritic activity is spatially structured . . 1098.3.8 Local activity cues instruct structural plasticity in dendriticfilopodia . . . . . . . . . . . . . . . . . . . . . . . . . . 1108.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1128.5 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1158.5.1 Animal rearing conditions . . . . . . . . . . . . . . . . . 1158.5.2 Imaging conditions . . . . . . . . . . . . . . . . . . . . . 1158.5.3 Calcium indicator loading and tectal infusions . . . . . . . 1158.5.4 Targeted neuronal silencing . . . . . . . . . . . . . . . . 1168.5.5 Target neurons . . . . . . . . . . . . . . . . . . . . . . . 1168.5.6 Selection of stimuli . . . . . . . . . . . . . . . . . . . . . 1168.5.7 Conventional two-photon imaging . . . . . . . . . . . . . 1178.5.8 Visual stimulation protocol . . . . . . . . . . . . . . . . . 1178.5.9 Processing of calcium imaging data . . . . . . . . . . . . 1188.5.10 Targeted electroporation . . . . . . . . . . . . . . . . . . 1198.5.11 Morphological imaging and dynamic morphometrics . . . 1208.5.12 Spatial clustering . . . . . . . . . . . . . . . . . . . . . . 1208.5.13 2D synaptic imaging . . . . . . . . . . . . . . . . . . . . 1218.5.14 Random access microscope design . . . . . . . . . . . . . 1218.5.15 ARAMiS - a random access microscopy suite . . . . . . . 1238.5.16 Processing of random access imaging data . . . . . . . . . 1248.5.17 Compensation of spreading calcium signals . . . . . . . . 1248.5.18 Spontaneous activity . . . . . . . . . . . . . . . . . . . . 1248.5.19 Measurement of correlations . . . . . . . . . . . . . . . . 1258.5.20 Spatial clustering of local dendritic responses . . . . . . . 1258.5.21 Image processing . . . . . . . . . . . . . . . . . . . . . . 126vii8.5.22 Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 1268.6 Supplementary figures . . . . . . . . . . . . . . . . . . . . . . . 1269 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1369.1 Learning mechanisms . . . . . . . . . . . . . . . . . . . . . . . . 1369.2 Strengths, weaknesses, and future directions . . . . . . . . . . . . 139Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141viiiList of FiguresFigure 3.1 In vivo imaging of evoked network activity in the unanesthetizeddeveloping brain. . . . . . . . . . . . . . . . . . . . . . . . . 35Figure 3.2 Orientation and direction responses in optic tectum . . . . . . 36Figure 3.3 Tectal noise correlations influence network decoding . . . . . 37Figure 3.4 Effects of visual training on single-neuron response properties 40Figure 3.5 Training induces NMDAR-dependent improvement of whole-network encoding . . . . . . . . . . . . . . . . . . . . . . . . 41Figure 3.6 Training-induced changes are stimulus-specific . . . . . . . . 44Figure 3.7 Training strengthens clustering of receptive fields and networkcorrelations. . . . . . . . . . . . . . . . . . . . . . . . . . . . 46Figure 3.8 NMDAR-dependent coordination between clusters supports net-work encoding improvement . . . . . . . . . . . . . . . . . . 48Figure 3.9 Schematic of receptive field and noise correlation plasticity fortrained and untrained stimuli. . . . . . . . . . . . . . . . . . . 50Figure 3.10 Methods for fluorescence data processing . . . . . . . . . . . 62Figure 3.11 Correlation of optical and electrophysiological firing rate mea-surements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63Figure 3.12 Optical measures of firing rate are correlated with electrophys-iological measurements . . . . . . . . . . . . . . . . . . . . . 64Figure 3.13 Noise correlations differ across stimuli . . . . . . . . . . . . . 65Figure 3.14 Noise correlation encoding . . . . . . . . . . . . . . . . . . . 66Figure 3.15 Receptive field similarity and noise correlation are associated . 67Figure 3.16 MK-801 blocks NMDAR currents . . . . . . . . . . . . . . . 68ixFigure 3.17 Neuron receptive fields are spatially clustered . . . . . . . . . 69Figure 3.18 Performance of shuffled decoders does not change with training 70Figure 5.1 Squaraine derivatives with large two-photon cross-sections . . 76Figure 5.2 Brightness and photostability of SeTau-647 . . . . . . . . . . 77Figure 5.3 Cellular and subcellular labelling with SeTau-647 . . . . . . . 79Figure 5.4 Multi-day cellular labeling with SeTau-647 . . . . . . . . . . 86Figure 8.1 Targeted Single-Cell Electroporation . . . . . . . . . . . . . . 100Figure 8.2 Neuronal activity determines experience-induced structural plas-ticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102Figure 8.3 Filopodial motility is clustered along dendrites by intracellulardistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105Figure 8.4 2D imaging of synaptic calcium transients . . . . . . . . . . . 106Figure 8.5 Comprehensive imaging of spatially-structured activity . . . . 108Figure 8.6 Noise Correlations reflect spreading dendritic signals . . . . . 111Figure 8.7 Sensory-evoked activity determines structural changes . . . . 113Figure 8.8 Neurons electroporated by TSCE show normal patterns of ac-tivity and growth . . . . . . . . . . . . . . . . . . . . . . . . 127Figure 8.9 Neurons of different activity profiles show similar dendritic ar-bor sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128Figure 8.10 Motility of filopodia in neurons of different evoked activityprofiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129Figure 8.11 Neuronal activity profiles in APV-treated tecta . . . . . . . . 130Figure 8.12 Filopodia are clustered along tectal neuron dendrites . . . . . 130Figure 8.13 Schematic of RAMP microscope . . . . . . . . . . . . . . . . 131Figure 8.14 Random Access Imaging Protocol . . . . . . . . . . . . . . . 132Figure 8.15 Effect of biasing diffusion compensation mixing coefficients . 133Figure 8.16 Spontaneous activity and evoked activity show different pat-terns across filopodia . . . . . . . . . . . . . . . . . . . . . . 134Figure 8.17 Correlations in activity are not associated with training-inducedmotility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135Figure 8.18 Dendritic shaft activity at sites of filopodial retraction . . . . . 135xAcknowledgmentsMy sincerest thanks to Professor Kurt Haas, who gave me every opportunity to be-come a mature and independent scientist, trusted me to pursue challenging projects,and was a great running partner throughout six years.My sincerest thanks also to my supervisory committee for insightful critiqueof my work. I’ve been very lucky to have you guiding me.Thank you also to the members of the Haas Lab, particularly Serhiy Opush-nyev, Janaina Brusco, DK Kim, Kelly Sakaki, Ana Niciforovic, Sharmin Hossain,Derek Dunfield, Simon Chen, Philip Edgcumbe, Steffen Kaiser, Tobias Horn, Bas-tian Hablitzel, Chris Conway, Dano Morrison, Blair Duncan, Xuefeng Liu, AnaRist, Lara Thompson, and many others.My sincerest thanks to the community of neuroscientists, cell biologists, andphysicists I have dealt with at UBC, Cold Spring Harbor, and Stanford, for every-thing you taught me.Thanks also to all my friends - scientists, climbers, housemates, all of theabove, or none of the above - who listened to my problems and often solved them.The greatest insights in my work have come from times spent with you.I am very grateful for public funding that made my work possible, and the workof many other scientists, through the Natural Sciences and Engineering ResearchCouncil, the Michael Smith Foundation for Health Research, and the CanadianInstitutes for Health Research.Most importantly, thanks to my family for understanding, and for supportingme even when I was sometimes too far away to support you.xiDedicated to my mother.I will always aspire to be more like you.xiiChapter 1Introduction1.1 Learning1.1.1 Learning in the brainLearning is the process by which activity passing through a neural network changesthat network to improve future processing. Different forms of learning underlievirtually all human abilities, ranging from playing a musical instrument to solvinga math problem, or simply looking at our environment and carving it in our mindsinto distinct objects, people, and places. Learning allows us to adapt to our specificenvironment. When we learn, our brain observes patterns in the world that areat first meaningless and somehow changes itself to extract meaningful featuresfrom those patterns. For example, consider how we learn our first language byobserving others speaking it. Without explicit instruction [203], our brains acquirethe ability to carve an unsegmented utterance into morphemes, phonemes, words,and phrases. A child learns that first language equally well whether it is Finnish orFrench, English or Urdu. The brain adapts to whatever environment it encounters.How the human brain performs such feats remains unexplained. Over a cen-tury of inquiry has described many of the psychological phenomena of learning:broad strategies, at times irrational, that humans and other animals use to makeassociations, ignore irrelevant details, and otherwise take advantage of regularitiesin their environment [196, 251]. At the same time, fundamental cellular mecha-1nisms of learning have been discovered. The list of genes and proteins known tobe required for the proper expression of learning and memory is large and growing[36, 82, 127]. These include activity-dependent kinases and phosphatases, sig-nal transduction proteins, structural proteins, and transcription factors that triggerconcerted genetic programs. Many of these mechanisms mediate changes in thestrength of synapses, the formation of new synapses, or the elimination of exist-ing ones. The birth and maturation of entirely new neurons has also been tied tolearning in two specific brain regions [231, 233]. There is now a consensus thatchanges in patterns of activity in neurons, which depend on the strength or presenceof synaptic connections and how those inputs are integrated, underlie all learningin the brain. These changes are themselves cued by patterns of activity. One sim-ple such pattern is conveyed by the well-worn summary of Donald Hebb’s famouspostulate [101]: ’Neurons that fire together, wire together’.Surprisingly, however, most of what we now know about learning has beendiscovered without directly measuring patterns of activity or synaptic connections.We have only recently developed the tools to simultaneously read out the activity ofmore than a handful of discrete neurons, and neuroscientists lack methods to readout large numbers of synapses, or to accurately determine the connectivity of morethan a handful of neurons within a given circuit. As a result, changes in these pat-terns of connectivity or patterns of synaptic strengths have only been directly tiedto changes in activity and learning in very specific cases. For example, one studytook advantage of the topographic separation of structurally plastic and non-plasticinputs in the auditory system of the barn owl, and its known topographic functionalorganization, to demonstrate that experience drives the clustering of functional in-puts along specific dendritic branches [169].Neuroscientists currently have an excellent understanding of psychological phe-nomena and microscopic elements such as proteins and single synapses, and agrowing understanding of mesoscale phenomena, such as patterns of activity oversmall number of neurons, or small numbers of synapses. However, we lack anexplanation of what is happening in the brain during learning that connects thesedifferent scales. Such an explanation must account for how patterns of activity car-rying information about specific experiences cause changes in subsequent process-ing, itself represented by patterns of brain activity. Moreover, such an explanation2must be detailed enough for us to reproduce learning abilities in simulations orphysical machines. For many researchers, the real test of whether we understandthe brain will be to build one.1.1.2 Learning in machinesIn many ways, the field of machine learning originates from pioneering work ofthe 1940s. Alan Turing formalized the modern theory of computation in 1946 (fora discussion and historical account, see [200]), inspiring generations of researchersin the construction of increasingly adaptable, intelligent machines. Claude Shan-non formalized the theory of how patterns convey information in 1948 [243]. Thequantification of information transmission has enabled study of how neurons en-code and decode information (discussed further in Chapter 2 and Chapter 3), andinsights into the efficiency of our nervous system [e.g., 15, 41, 72, 249]Researchers have long sought to develop intelligent machines, and much of thiswork was inspired by the structure of the human brain. Among the earliest artifi-cial neural networks, described in 1954, were a type of linear classifier known asthe perceptron [225]. Computation by early instantiations of perceptrons was rel-atively simple, as these networks were composed of a single layer of input nodes,each multiplied by a weight and summed to produce a single output. However,the important insight of these networks was not in their computational power, butin their ability to learn from experience. Neural networks like perceptrons canbe trained through experience, by presenting them a set of training examples andupdating their connection weights according to a ’learning rule’, a formula that de-pends on the activity levels of neurons, their inputs, and the existing weights. Thisprocedure mimics how neural networks in the brain learn from experience. Learn-ing rules are typically chosen so as to progressively decrease some measure of error(known as an objective function) with successive rounds of training, allowing themto, for example, classify inputs.The linear properties of perceptrons made it easy to define learning rules forthem and analyze resulting networks. This simplicity, however, came at a cost.Marvin Minsky demonstrated that networks made up of a single layer of percep-trons are incapable of computing even simple logical operations, such as XOR (i.e.,3A or B, but not both) [178]. A network composed entirely of linear componentsis only able to compute, in total, a linear function. Of course, the vast majority ofmeaningful computations performed by our brains are not linear, and this demon-stration discouraged many researchers at the time. The field of artificial neuralnetworks has advanced greatly since the perceptron. Methods were developed fortraining multiple layers of neurons, each computing their own decision boundary,allowing for nonlinear computations. For an excellent introduction to this topic,see Rumelhart et al. [228].Many other architectures have been described for artificial neural networks[85, 111, 149, 209]. Some of these architectures use learning rules or computa-tional mechanisms that less directly mimic the structure of neurons in the brain,while others are more biologically plausible. Perhaps most notably, significantadvances have recently been made in our ability to train large, many-layered non-linear neural networks, known as ’deep learning’, using general purpose learningrules that rely largely on locally-measurable statistics. Such networks are able toperform impressive tasks, and represent the state of the art in challenging taskssuch as natural image recognition and speech recognition [106, 137]Perceptrons and modern neural networks exemplify an important distinctionbetween linear and nonlinear computations. This distinction has recently becomethe focus of intense investigation, not in artificial neural networks, but in neuronsinside the awake brain.1.2 Monitoring activity patterns in the brain1.2.1 Introduction to linear and nonlinear processesIn a nonlinear system, the effect of any element depends on the states of manyother elements. This makes nonlinear phenomena in the brain difficult to study,because it requires measurement of many things simultaneously. Measuring onlysmall amounts of a pattern at a time make it impossible to detect higher-level struc-ture, which can carry important information in a nonlinear system. To illustratethese issues I will briefly focus on information encoding, which is the topic of themanuscript presented in Chapter 3. Understanding encoding requires a model, ei-4ther explicit or implicit, of the probability of detecting different neural activity pat-terns in response to different possible stimuli. When neurons encode informationlinearly, these probabilities can be computed by treating each neuron as indepen-dent and multiplying their individual probabilities, which can then be measured andanalyzed separately. In contrast, nonlinear encoding demands more difficult mea-surements, not only because many neurons have to be measured simultaneously,but because more samples are required to accurately measure the probabilities ofvarious patterns. Consider a network composed of 10 neurons, each of which canhave two states, ON or OFF. If neurons are independent, the probability of anyglobal state over all 10 neurons depends only on the probability of each neuronbeing ON, so we have to estimate a set of 10 numbers. If neurons interact, but onlyin pairs, we have to estimate a 2 x 2 table of states for each of 45 pairs of neurons.Some of these parameters are redundant, so we are left estimating 55 numbers. Ifwe make no assumptions about how sets of neurons can interact, we must directlyobserve each possible network state, leaving us estimating 210− 1 = 1023 num-bers. How many samples would be required to have a reasonable estimate of theseprobabilities? In the last case, if each state is equally likely, having a 50% chanceof observing each state at least once requires over 9,800 samples 1. In the first(linear) case, we achieve this level of precision with just 4 samples. In general, asthe complexity of interactions in a system increases, the number of samples neededto characterize them increases much more rapidly, a problem that has been termed’combinatorial explosion’.When performing experiments in vivo, obtaining many samples is difficult. Inour case, a sample corresponds to the presentation of a stimulus. In most experi-ments, the fastest that stimuli can be presented to animals is a few times per minute,and experiments can last at most a few hours. Over time as stimuli are presented,neuronal responses can change due to accommodation, habituation, or long-lastingplasticity. As a result, it is in many cases impossible to collect more than a few1For smaller numbers, this can be calculated precisely by computing the Stirling number of thesecond kind, but for larger numbers this approach becomes computationally costly. Here I haveproduced a lower bound on the required number of samples by assuming that counts for each cell inthe table are independent Poisson processes, so the probability of observing every state at least oncetakes the form (1− e−k/n)k. This calculation itself demonstrates how assumptions of independencecan simplify analyses!5hundred samples from a given set of neurons. In most cases, the situation is muchworse than I have described here. Neural circuits consist of many more than 10neurons, and each can show graded levels of activity, not just ON or OFF states.Measurements contain noise, and samples can be lost, complicating analyses.Analysis of interacting systems can be even more difficult than measurement.For the same reasons of combinatorial explosion, it is difficult to build and solvemodels that contain many high-order interactions. An important line of researchhas been to develop models that can capture higher-order interactions while havingfitting algorithms that are efficient in memory and computation time [201, 202,241]. One such model for a circuit of pairwise-interacting neurons is presented inChapter Linear and nonlinear biological processesGiven the challenges of working with nonlinear systems, there has been a lot ofinterest in whether or not the brain uses nonlinear interactions to encode informa-tion and perform computation. It is now known that many nonlinear phenomenaoccur within neural circuits as they process information, and this fact is perhapsthe greatest motivation to study patterns of activity in the awake brain.Perhaps the best known nonlinearity in the brain is the generation of action po-tentials. Neuronal membranes contain voltage-gated ion channels which are non-linear, showing different permeability to ions depending on the recent history of themembrane voltage. Neurons express several types of voltage gated channels, withexcitatory sodium and inhibitory potassium channels playing the largest role in theaction potential of most neurons [108], along with voltage-gated calcium channels.These channels result in a threshold potential, which, when crossed, produces abrief stereotyped spike in the membrane voltage. Inputs that do not drive the cellabove the threshold produce no spikes, so spiking is more sensitive to inputs ar-riving simultaneously. NMDA receptors have voltage-dependent conductance andare also gated by the presence of extracellular glutamate, meaning that these chan-nels conduct when presynaptic activity coincides with postsynaptic activity. Unlikemany other receptors, NMDA receptors conduct calcium, an important intracellularindicator of activity that triggers plasticity and other activity-dependent processes6[160], but also have metabotropic effects independent of conduction [186]. Thedetection of calcium signals inside the cell involves further cooperative, nonlin-ear processes. Some calcium-dependent proteins possess multiple binding sitesthat must be simultaneously occupied to trigger physiological effects, resulting inthreshold-like phenomena [e.g., 261]. Calcium transients of sufficient size triggerfurther calcium release from intracellular stores, while even higher concentrationsblock release [75, 292]. Further interactions arise from the geometry of neuronaldendrite. How an input to a distal dendrite travels to the soma, where the actionpotential is often generated, depends on intervening conductances which can shuntcurrents out of the cell [262].Given the wealth of nonlinear mechanisms within cells, it may seem obviousto conclude that nonlinear mechanisms are involved in, and important for, mostaspects of neural information processing. However, many phenomena are wellapproximated by strictly linear interactions, including the generation of receptivefields of ’simple cells’ in visual cortex [118], and indeed it seems that certain partsof the brain are exquisitely adapted to maintain linearity throughout their entire dy-namic range [11, 124, 276]. A fundamental insight appears to be that higher-orderprocessing centers in the brain contain many neurons that encode complex stimu-lus features in a low dimensional, or linear, way, such as neurons with responses tothe presence of specific people (e.g., Jennifer Aniston [215]). As methods continueto be developed to measure patterns of activity within neurons, the importance oflinearity, and the contribution of nonlinearities, are becoming increasingly popularlines of research.1.2.3 Population codingIn some cases, the properties of sensory systems are arranged such that sensoryneurons provide roughly independent information, resulting in linear populationcoding. One such example is the cercal system of crickets for the sensation ofwind direction, in which different neurons are tuned toward nonoverlapping com-binations of wind direction and speed [176]. Even when the activity of neurons isindependent, it can be important to measure population activity. Of course, measur-ing population activity is typically necessary to demonstrate that neurons are in fact7independent. Moreover, even in independent populations of neurons, the questionremains how independence is established, and how effiently these independent re-ceptive fields are organized to encode the space of possible environmental stimuli.These questions are explored in Chapter 3.Information can also be conveyed by nonlinear combinations of neuronal ac-tivity [202], a fact that becomes obvious if we consider how complex stimuli acton early stages of sensory transduction. For example, the presence of a specificperson X in the image on our retina is not a linear function of the activation of pho-toreceptors. However, the information is undoubtedly there when we recognizeX, meaning not only that the retina encodes information nonlinearly but that ourbrain is able to decode this information. To compute nonlinear features, neuronsmust use nonlinear mechanisms to detect patterns in their input. In some cases,researchers have been able to identify how neurons detect nonlinear features [290],but a great deal of work in this field remains to be done. When we build modelsto decode population activity, these models typically only show one way that in-formation can be extracted from that activity, which does not necessarily representthe way downstream neurons decode that or any other information.Bearing this complication in mind, the structure of activity across neuronalpopulations indicates a wide variety of ways in which information is encoded inneuronal activity. Interesting phenomena include signals conveyed by the precisetiming of spikes relative to oscillating population activity [120], and populationsof ’grid cells’ in the entorhinal cortex which encode an animal’s location withina specific environment using a basis set of hexagonal lattices of various spatialfrequencies [180].These and other phenomena were identified using a variety of electrical meth-ods. Early studies used single sharp electrodes to penetrate the brain and recordactivity in axon bundles or individual neurons, and local field potentials repre-senting synaptic transmission of many neurons. The patch clamp [234] enabledrecording and manipulation of intracellular potentials or membrane currents bysealing an insulating glass pipette to the membrane of a single neuron. More re-cently, these methods have been greatly parallelized, allowing recording of activityin large neuronal populations. Patch recordings are routinely performed in mul-tiple neurons simultaneously, and systems under robotic control have been devel-8oped [133]. Multi-electrode arrays have been used to densely sample neuronalpopulations, such as every cell in a patch of retina in vitro [202]. Tetrodes andoctrodes, composed of thin wires wound around each other to form a single probe,allow identification of up to dozens of spiking neurons simultaneously based onthe waveform and amplitude of spikes recorded at each site. Silicon wafer probescan similarly contain hundreds of recording sites, allowing recording of even largerneuronal populations [33].Electrophysiological methods have provided unparalleled insights into howneurons and neuronal populations encode information. Nevertheless, such meth-ods have fundamental limitations, as inserting a probe into the brain necessarilydamages neurons and severs their connections. This damage, unfortunately, is pro-portional to the number of neurons sampled. Recording from larger numbers ofneurons results in greater damage. To address this problem, and enable completesampling of activity in neuronal populations in awake animals, neuroscientists haveturned to optical methods. Light can pass through living tissue without damage, al-lowing neuronal populations to be sampled remotely thanks to the developmentof activity-sensing optical probes. Activity-sensing probes can be used to denselylabel neurons, in transgenic animals, by viral labelling, or bulk loading with cell-permeable organic dyes. Imaging is fundamentally parallel, and every pixel thatis resolved by an imaging system represents a distinct channel of recording. Ac-tivity sensors can by imaged in vivo through a growing assortment of methods,reviewed thoroughly in Section 1.3. The most common of these methods is fluo-rescence imaging. Fluorescent probes have been developed to detect several kindsof activity in neurons, including membrane voltage [302], neurotransmitters [167],and intracellular ions such as Chloride [298] and Calcium [45]. Calcium sensorsin particular have several advantages as probes of neuronal firing. Somatic actionpotentials are accompanied by a calcium current that can be used to detect firing inmany types of neurons. The cytosol has a very low resting calcium concentration,so this current can result in orders of magnitude changes in intracellular calciumconcentration that are easily detected. Calcium currents are slow, which is a dis-tinct advantage for scanning methods or cameras with lower framerates. An actionpotential lasts for milliseconds, while the associated calcium current typically lastsfor well over 100 ms, and fluorescent sensors may be slower still. Calcium imag-9ing also has disadvantages. Because these fluctuations are slow, very fast eventsmight not be distinguishable. This is less of an issue than one might think, becauseuntil recently methods did not exist to image probes at very high frame rates invivo over long periods of time. Calcium transients also do not perfectly report ac-tion potentials in all cell types, and often do not report subthreshold inputs at all.Nevertheless, the distinct advantages of calcium imaging have led to widespreadadoption of this technique for monitoring of population activity. The developmentof reliable reporters of intracellular calcium also enabled imaging of activity at asmaller spatial scale: the synapse.1.2.4 Dendritic integrationIn addition to the soma, calcium transients can also reflect activity within neuronalprocesses [25, 123, 185]. Presynaptic calcium is the fundamental trigger for re-lease of vesicles containing neurotransmitter. In the postsynaptic compartment,NMDA Receptors, calcium-permeable AMPA receptors, and voltage-gated cal-cium channels deliver calcium currents reporting their activity. Calcium transientsin dendritic spines of voltage-clamped cortical neurons almost entirely representglutamatergic synaptic transmission [123].The detection of calcium currents is even more effective in dendrites than it is inthe soma. Because dendritic compartments are small, depolarizations and calciumtransients driven by membrane currents are amplified compared to similar currentsat the soma. Dendritic spines, the sites of many excitatory inputs, compartmental-ize calcium and allow readout of distinct inputs to the same neuron. Researchershave taken advantage of these properties to map the spatial organization of synapticinputs to dendrites [123] and investigate coincident activity among neighbouringinputs [265]. Calcium imaging has been used to measure the depolarization of in-dividual dendritic spines, a compartment too small to be accessible to patch clamp[98]. Calcium imaging has also recently been used to directly investigate linearand nonlinear modes of integration of synaptic input in layer IV cortical neuronsin vivo [124].The spatial organization of synaptic inputs affects how they are integrated bya neuron. Active conductances can amplify coincident depolarizations and affect10signal propagation, and even passively propagated signals depend on the location ofsynapses and membrane conductances. As a result, the contribution of a synapseto a neuron’s activity depends on the simultaneous activity of other, potentiallydistant, synapses. Such nonlinearities potentially allow neurons to be sensitive tocorrelations in their input, responding to specific features only when certain otherfeatures are present.An important question to many neuroscientists is therefore whether inputs toneurons are spatially and temporally organized to take advantage of nonlinear in-tegration. For example, the clustering of inputs to dendritic branches could allowbranches to act as independent computational units [262], summing local inputsand applying a local nonlinear transformation. This could increase the complexityof computations performed by a given circuit, as computations that would other-wise require a circuit might instead be computed within a single neuron.Consistent with such possibilities, both in vitro and in vivo studies have shownthat coincident synaptic activity and plasticity can be spatially structured alongdendrites in various contexts [45, 99, 131, 162]. Moreover, recent studies haveshown that nonlinearities play an important role in certain computations in vivo[146, 252, 290], though these studies did not measure nonlinearities at the scaleof individual dendritic spines. Despite a great deal of interest in nonlinear inte-gration, it remains unknown to what extent neurons actually use interactions insynaptic activity to perform behaviorally-relevant computations, and what sorts oflearning rules would allow them to do so. This lack of research has largely beendue to the absence of techniques for measuring large numbers of synaptic inputssimultaneously, in an awake brain, as it performs a computation.1.3 Microscope development1.3.1 A brief history of imagingAdvances in our understanding of the brain have for a long time been coupled toadvances in microscopy. The work of prominent neuroanatomist Santiago Ramony Cajal was made possible by advances in microscopy by Ernst Abbe, Carl Zeiss,and Otto Schott, including the development of the condenser and compensation11for spherical aberration [29], and Camillo Golgi’s discovery of the sparse silverstain technique that bears his name [212], in the latter part of the 19th century.These techniques allowed Ramon y Cajal to label and trace individual neurons andglia in fixed brain tissue, identify morphological classes of cells, and form earlyhypotheses on the nature of synapses and information transmission in the brain[18], among many other contributions. Ramon y Cajal hypothesized that neuronssignal to each other across a minute extracellular gap at the synapse, as opposed tobeing directly connected to form a reticulum. This hypothesis, however, could notbe directly proven through optical observation, which is limited by diffraction to aresolution of several hundred nanometers (the wavelength of visible light), a scalelarger than the hypothesized synaptic cleft. It was not until the development of theelectron microscope that this prediction was verified by direct observation.Electron microscopy enabled a new wave of anatomical study at a much finerscale, thanks to its increased resolution over optical techniques due to the muchhigher momentum (and shorter wavelength) of electrons compared to photons.The synapse, once an invisible structure, could be subdivided into numerous pre-and post- synaptic features clearly visible in electron micrographs. Developmentsin electron microscopy continue to advance neuroscience. Serial sectioning andblockface imaging have allowed the volumetric reconstruction of nanoscale brainstructures, and methods continue to be developed to image larger volumes at thisresolution in shorter times, with fewer errors [31]. These detailed imaging tech-niques, paired with automated computational 3D reconstruction and annotation,promise to produce complete maps of connectivity within a given volume of braintissue, a field recently termed ’connectomics’. Very large projects have begunaround the world to develop the necessary software for reconstructing such vol-umes, and hardware, such as multi-beam scanning electron microscopes and in-chamber milling techniques [31, 122, 126, 132]. Projects of this type have al-ready produced detailed structural maps of the nervous system of the roundwormCaenorhabditis elegans (with manual reconstruction) [95, 284], and more recentlythe mammalian retina [104]. The advent of these monumental connectomics ef-forts leaves several questions to be addressed. First, electron microscopy is in-herently monochromatic, and it is unclear whether all relevant aspects of cell orsynapse identities could be identified with these techniques. Second, it is unclear12how to relate purely structural information to the activity patterns and functionalrelationships between neurons that subserve information processing. Finally, elec-tron microscopy produces a fundamentally static image of an organism in time,leaving questions of how these circuits are formed - or rather, form themselves - indevelopment and maturity.Here again, microscopy techniques continue to be developed to answer thesetypes of questions, enabling imaging of structures, molecular identities, and ongo-ing processes that could not previously be observed. Perhaps surprisingly, mostof these highest-resolution techniques are optical, but take advantage of variousphysical phenomena to overcome limits of resolution, imaging depth, speed, orcontrast. Array Tomography involves labeling of samples sectioned to the samethickness as electron microscopy [174] which are labelled with fluorescent anti-bodies to provide desired contrast, and imaged optically. Antibody probes allowdifferent targets to be imaged in different colours, and ultrathin sections enablethe highest possible optical resolutions in the imaging plane and even better axialresolution.Recent technologies have pushed the resolution limits of optical systems toenable ’super-resolution’ imaging of features smaller than the diffraction limit oflight. Structured illumination [e.g., 39] can double the resolution of an optical sys-tem, along each axis, by mathematically combining multiple images under differ-ent known patterns of illumination. Each image and illumination pattern is limitedby the diffraction of light, but the interference of the illumination pattern with thestructure of the sample recodes high-spatial-frequency features of the sample intolower-frequency components of the image, which can then be recovered. Whilethe diffraction limit on the resolution of optical systems has been understood sincebefore astronomer George Biddel Airy formalized it in 1835 [1], this limit onlyapplies to light waves propagating multiple wavelengths through a medium, theso-called far field. A number of techniques place collection optics very close to thesource of emission, the near field, to break the diffraction limit [e.g., 10]. Morever,the diffraction limit only applies to linear optical phenomena. By taking advan-tage of nonlinear phenomena, such as stimulated emission depletion [102], opticaltechniques have achieved resolutions of dozens of nanometers. Finally, researchershave used the knowledge that light sources comprise single molecules [229], and13the ability to manipulate their photophysical properties [105], to localize thosemolecules on spatial scales even smaller than the pixels of their images.All of these super-resolution techniques have been applied to the study ofsynapses. Higher resolution allows the visualization of apposed pre- and post-synaptic elements, and subcellular structures such as cytoskeletal proteins [e.g.,79]. Array tomography has been used to catalogue patterns of protein expressionat synapses, to identify variations in synapses across a single neuron and whethersynapses can be classified into types, analogous to established methods of classi-fying cells [175, 254]. Both electron microscopy and these high-resolution opticaltechniques have enabled new experiments in neuroscience, but they are largelylimited to thin samples or slices of fixed brain tissue. Many important questions inneuroscience, particularly those relating to how information is processed, requirethe monitoring of ongoing processes deep in the awake brain.1.3.2 Imaging in the awake brainMarvin Minsky, introduced here for his work on perceptrons, made another seminalcontribution to neuroscience as the inventor of the confocal microscope in 1955[177]. Imaging of structures embedded in tissue, such as neurons within the brain,is strongly limited by the scattering of light. The mean scattering length of visiblephotons is roughly 100 microns in brain tissue, meaning that much of the lightarriving at a detector from this depth or below traces an unknown path throughthe sample and does not produce a useful image in a traditional imaging system.The confocal microscope addresses this issue by placing a ’pinhole’ aperture in theimaging path at a point conjugate to a single point in the focal plane. Light thatpasses through the pinhole largely represents unscattered light originating fromthe focal point in the sample. By scanning the illumination point and the pinholeacross the sample, an image can be assembled (in practice the pinhole is fixedand collected light is ’descanned’ using a moving mirror, which also scans theillumination). The confocal microscope therefore enabled imaging of 3D samples,albeit at a high cost in time required to scan illumination across the sample, and theconcomitant damage induced by the illumination of large volumes of the samplerejected at any given time by the pinhole.14The next major advance in the field of in vivo imaging, which remains the dom-inant technology today, was two photon microscopy, introduced by Winfried Denkand Watt Webb [61]. Two Photon microscopy achieves the optical sectioning ca-pability of the confocal microscope by relying on nonlinear two-photon absorptionthat depends on the square of the light intensity. Because the integrated squaredlight intensity is greatly increased at the focus of a cone of light produced by anobjective, meaningful levels of excitation occur only at the focal point. All ofthe emitted light can be collected by a single detector, without a pinhole, and as-signed to the location of the focal point, which is scanned to produce an image.An important advantage of two-photon imaging is that excitation light for optimaltwo-photon absorption is often near-infrared, at wavelengths to which water (andhence brain tissue) is most transparent. This allows a tight focus of excitation lightto be produced at depths of hundreds of microns in awake brain tissue [60]. Twophoton imaging is the most common technique currently used to monitor activitypatterns in groups of neurons in awake animals in vivo.The depth achieved by two-photon imaging has been greatly extended by theuse of adaptive optics, an idea co-opted from astronomy, to measure and correctfor wavefront aberrations produced by the sample [e.g., 282]. Three-photon imag-ing, relying on third-order absorption, has demonstrated extremely deep imagingby taking advantage of another ’window’ of wavelengths at which tissue is highlytransparent [e.g., 114]. Nevertheless, even with extensive refinements, these meth-ods are limited to depths of roughly 1 mm within tissue.Several alternative methods also show promising applications at extreme depthsinaccessible to multiphoton imaging. Miniature endoscopes can avoid the need toimage at great depth by implanting a thin imaging objective (an optical fiber and/orprism) directly into the brain, though this requires removing overlaying tissue andcould affect function of the region being imaged [301]. Opto-Acoustic imagingcombines infrared excitation of a light-absorbing probe with ultrasonic detectionof thermal phonons emitted by these probes [283]. Sound waves can be used toconstruct images in much the same way as light, albeit with a lower resolution dueto the longer wavelengths involved. Sound travelling through tissue is attenuatedmuch less than light, potentially allowing monitoring of activity throughout thedepth of the brain in mammalian models.15The selection processes that determine action cross-sections of different typesof microscopy are distinct, requiring distinct probes. Opto-acoustic microscopyrelies on absorption of light without fluorescence emission. Ideal fluorophores, onthe other hand, have strong absorption and a high efficiency transition to the emis-sion of a Stokes-shifted (redder) photon. Two-photon absorption depends stronglyon polarizability of the molecule under the absorption event, and strongly favoursasymmetric molecules [281]. Similarly, second-harmonic generation requires themutual alignment of many asymmetric harmonophores [34]. Therefore, an im-portant line of inquiry has been the development of contrast agents that are brightunder specific modes of excitation and detection. Chapter 5 deals with the identifi-cation of novel ultrabright organic molecules for use in two-photon microscopy.1.3.3 Ultrafast microscopy for synaptic imagingAs introduced in the previous sections, a hope for modern microscopy techniquesis the development of methods to simultaneously sample activity throughout a sin-gle neuron’s dendrites in an awake animal, to directly observe dendritic integra-tion. Such a complete picture of how a single neuron integrates information wouldgreatly complement ongoing studies of information processing in groups of neu-rons, and comprehensive maps of neuronal connectivity anticipated from connec-tomics efforts. Ideally, new tools would allow us to directly and completely mea-sure quantities that are relevant to mechanistic models of neural processing, such asconnection strengths, correlations, response filters, and learning rules. This wouldallow neuroscientists to take advantage of methods and theorems developed by themachine learning community, more closely linking these two fields. Neuroscien-tists have so far been unable to simultaneously observe activity in large numbersof synapses distributed in 3D on a single neuron in an awake animal. A majorproject of mine has been to develop a microscope capable of this feat, and apply itto the identification of information processing mechanisms and learning rules usedby neural networks in the brain, described in Chapter 8.Several techniques have been investigated, or are currently being investigated,to allow rapid sampling of activity in 3D samples, and some of these techniquesshow promise for imaging synaptic activity in vivo. These methods fall into two16broad categories: methods that use camera-like detectors to record fluorescencefrom many sites in parallel, and methods that use a single detector to rapidly readout emitted photons serially. The most significant limitation for camera-basedimaging in vivo is scattering. As mentioned earlier, this concern motivated thedevelopment of confocal and two-photon laser-scanning microscopy. However,where scattering can be overcome, cameras allow parallelization of the imagingprocess, with the potential to significantly increase imaging speed over serial laserscanning. Scattering is less of an issue where the sources of light are sparse. Be-cause photons are typically scattered at low angles in tissue [181], scattering canto some degree be approximated by local blurring of an image. If sources are sep-arated by more than the spatial scale of the blurring, they can be resolved. Tech-niques have therefore been developed to illuminate distinct spots across the sampleand produce images or detect brightness fluctuations in resulting fluorescence, us-ing either fixed patterns of focal points [20] or a spatial light modulator to producepatterns of illumination in 3D [194]. In a traditional imaging system (one built ac-cording to the Abbe sine criterion), a camera can collect in-focus fluorescence froma single image plane. As more light is collected (i.e., as the numerical aperture in-creases), the depth of the in-focus image field decreases. This makes it difficultfor a camera to image a 3D volume because light from outside the focal plane ishighly defocused. Methods have been developed to extend depth of field, allowingimaging of a large axial extent while retaining high lateral resolution and a largenumerical aperture. Most prominently, wavefront coding techniques add phasemodulation in the detection path of a traditional microscope to extend the depthof focus, producing an extended, pencil-shaped point spread function that effec-tively collapses the 3D structure of the sample onto the 2D detector [e.g., 67, 214].This obviates the need for axial scanning, improves acquisition speed, and reducesphotodamage relative to confocal imaging by collecting light from all illuminatedpoints simultaneously. Active, adaptable control over spatial phase of emitted lightcan be achieved using a 2D deformable mirror in the microscope collection path,though the resolution of these devices is lower.The speed of such camera-based methods is limited largely by the framer-ate achievable by the detector, the current state-of-the-art being roughly 1kHz for1024x1024 acquisition, though methods which rely on axial scanning are slower17by a factor of the number of image planes sampled.One of the most promising serial scanning methods is random access mi-croscopy [218], though it has heretofore been little used in vivo. The main chal-lenge of applying laser-scanning methods such as traditional two-photon microscopyto imaging synapses is the time taken to scan all the pixels in an area or volume[123]. Even a modest 3D volume (100x100x100 microns) imaged at a resolutionsufficient to detect synapses will contain at least several million pixels. Even ifeach pixel is dwelled upon for only 1 µs, volumes will require many seconds toacquire, making it impossible to monitor rapid activity signals. Random access mi-croscopy ameliorates this problem by selecting only a subset of pixels to acquire,and using hardware capable of scanning arbitrary sets of pixels in 2 or 3 dimen-sions. When relevant imaging sites are sparsely scattered in space, random accessmicroscopy promises to greatly improve imaging speed compared to full volumeraster scanning. 1 mm of dendrite can be densely sampled at 1 micron intervalswith only 1000 pixels, many thousand times faster than a raster scan. Random ac-cess scanning can also be faster than camera-based methods for smaller numbersof points, because of the limited read-out speed of camera detectors. If a targetpoint can be acquired in 10 µs, for example, an image plane with less than 100points can be scanned by random access imaging faster than a camera can acquirean image of that plane.However, random access imaging has a number of practical limitations. Thelaser beam is scanned with a series of acousto-optic deflectors (AODs), which splitthe beam into a series of diffracted orders. The diffracted beams (but not the un-diffracted 0th order beam) are deflected by an angle dependent on the spacing ofthe crystal planes. This spacing can be altered periodically by passing a soundwave through the crystal, altering the diffraction angle by a small amount. If nec-essary, this scan angle can be magnified using a telescope, though this reduces thediameter of the excitation beam (and increases the size of the focal point). Theconsequence of this is that the number of spots resolvable by an AOD system isdetermined by the aperture diameter of the crystal. Large optical-quality crystalsare difficult to grow, and it takes longer for sound waves to cross a larger aperture.This second limitation, the ’access time’ of the deflectors, determines how longit takes to move the laser focus between two arbitrary points, resulting in a fun-18damental trade-off between imaging speed and resolution. Acousto-optic crystalsare also highly dispersive, acting differently on the different wavelengths of lightwithin ultrafast laser pulses used for two-photon imaging. AOD crystals introduceboth temporal and angular dispersion, which must be compensated for to achievehigh resolution and effective two-photon excitation. The diffraction efficiency ofeach AOD varies across its angular range, requiring active compensation.The greatest difficulty associated with in vivo random access imaging is sam-ple motion. Raster scans or camera frames acquired at high enough rates can bealigned to compensate for a moving sample. However, even a small amount of sam-ple movement will cause a random access scan to miss its target points, with nopossibility to realign the data post-hoc. Several studies have investigated rapid on-line detection and correction of sample motion [e.g., 42], but real-time submicronmotion correction deep within a living animal has not been demonstrated. Motionarises from voluntary movement of the animal, if it is unanaesthetized, but alsofrom breathing and blood circulation, and is therefore unavoidable in these prepa-rations. Very close to the surface of the cortex, brain tissue is more firmly heldto the dura, and motion is roughly rigid. Several hundred microns deeper, wherethe cell bodies of cortical neurons lie, this rigidity breaks down, posing seriousproblems for AOD-based imaging. Therefore, random access scanning is currentlyonly effective in preparations where sample motion can be stringently controlled.There are currently no methods that simultaneously allow for rapid high-resolution3D imaging and robustness against sample motion. Speed can be compromised forrobustness to small movements by imaging small patches surrounding every targetsite. Even a relatively small buffer against motion comes at the cost of greatly in-creased scan time; an additional 5 pixels in each direction will take over 100 timeslonger to acquire than a single pixel at the target site. Camera-based methods couldallow for imaging in moving samples by illuminating larger focal points to com-pensate for small movements, though this would come at a great cost to resolution,because resolution in these systems arises from the tight focus of the excitationlight (detection in these systems is strongly limited by scattering).191.4 Brain development1.4.1 The Xenopus laevis model systemWith the aforementioned methods and their respective limitations in mind, it wasclear to me that random access two-photon calcium imaging would be an effectiveapproach to imaging large numbers of excitatory synaptic inputs to a neuron if thatneuron could be kept stationary. While this is not currently possible in rodents,the brain and head of Xenopus laevis tadpoles is sufficiently small that they can befirmly immobilized, restricting movements of structures in the brain to less than 1micron, even while awake. Xenopus laevis tadpoles have a number of other distinctadvantages making them particularly suited for the study of developmental brainplasticity and early learning.Xenopus is a genus of aquatic frogs native to Africa. Xenopus laevis, in par-ticular, has been extensively used as an in vivo model system, particularly in earlydevelopment and neuroscience, and Xenopus oocytes are commonly used for ex-ogenous expression of proteins for in vitro study of signal transduction and cellularphysiology. Being a vertebrate, Xenopus shares many aspects of neuronal andsynaptic physiology [77, 250] with humans and other vertebrate model systems,and has been used in the study of numerous human diseases (for a review, seeHarland and Grainger [97]).Early studies of vertebrate development made use of easily accessible exter-nally developing embryos of amphibians to investigate how various body struc-tures are patterned, including the formation of the vertebrate body axis [256] andthe induction of neural cell fates, by dissecting and transplanting groups of cellswithin developing embryos.Xenopus laevis has more recently enjoyed a great deal of attention as a tractablemodel of activity-dependent vertebrate brain development. Since imaging of braincircuit formation has largely proved untenable in mammals due to poor access tothe developing brain during in utero and early postnatal periods, most of our knowl-edge of in vivo brain circuit formation comes from studies of transparent fish andamphibian embryos. The retina, the light-sensitive organ at the back of the eye,detects and processes visual information and ultimately transmits information to20several brain regions via the firing of retinal ganglion cells (RGCs), the axons ofwhich form the optic nerve. In humans, the primary projection of RGCs is to thelateral geniculate nucleus of the thalamus, which in turn projects to primary vi-sual cortex and other cortical areas [239] subserving most cognitive visual tasks.A second large projection of RGCs innervates the tectum, also termed the supe-rior colliculus. These projections maintain a spatial map of retinotopic space, withretinal coordinates being transformed in an ordered fashion onto the coordinatesof the target structure. Tectal neurons also receive inputs from several other sen-sory modalities [59], and their primary function appears to be to direct orientingbehaviours towards particular regions of space corresponding to their location onthe topographic map [170].The retinotectal projection of RGC axons are an important model for the forma-tion and refinement of other topographic projections such as those in the thalamusand cerebral cortex [77]. In Xenopus, the tectum is the largest target of RGC pro-jections, and has been used extensively to study the establishment and refinementof these functional maps. For example, gradients of innately expressed cues thatestablish the rostro-caudal (ephrins-A and EphA) and medial-lateral(ephrins-B andEphB) axes of the tectal topographic map, predicted by Roger Sperry [257], werefirst identified in Xenopus [47]. The retinotectal projection has been used to studymany other important aspects of early neuronal development, including neuroge-nesis, axonal and dendritic growth, synaptogenesis, and activity-dependent refine-ment.The compartmentalized nature of the eye and ventricle simplifies pharmacolog-ical manipulations and make it easy to label either pre- or post- synaptic neuronsby electroporation [73]. The ease of transplanting early embryonic tissues, and theunilateral projection of RGCs to the opposite hemisphere, have enabled insight-ful manipulations that elucidate the mechanisms of axon growth and refinement[184, 219]. The skin of albino tadpoles is transparent and the retinotectal projec-tion is very superficial in the brain, enabling in vivo imaging without surgery, andfacilitating longitudinal studies across many days of development [e.g., 115].Studies of topographic mapping in the tectum have largely focused on the or-ganization of RGC axons, the major excitatory inputs to tectal neurons. However,it is now clear that tectal dendrites are also highly active in the formation and re-21finement of connections, using numerous intracellular signalling mechanisms toselectively alter growth dynamics. As with most excitatory neurons, the principalcells of the tectum are produced from radial glia lining the ventricle, and migrateprogressively outward (laterally) over many days as they mature. As they mature,tectal neurons form dendrites and extend filopodia, which receive synaptic connec-tions from the retina, other tectal neurons, and other brain regions. The maturestructure of tectal neurons, and all neurons in the nervous system, is the cumula-tive product of a vast number of individual extensions and retractions of dendriticand axonal processes. The initial structural development of tectal neurons has beencaptured by time-lapse 3D microscopy, imaging the full volume of neurons at rel-atively short intervals to track the dynamic formation, elimination, growth, andshortening of these processes [43, 115]. Such studies during early stages of de-velopment have revealed that most new filopodia are rapidly pruned, with somebecoming stable, and some elongating to form new branches which then extendfilopodia of their own [115]. Similarly, the development of presynaptic axons ischaracterized by the early formation of relatively large axonal arbors which thendecrease in size with maturation. The relative size of arbors, as compared to theextent of the tectum, decreases strongly with development. Because incoming ax-ons are organized retinotopically, the pruning and refinement of connections overtime can be observed as a decrease in the size of tectal neuron receptive fieldsthroughout development [188, 268]. This concept, of structural changes in con-nectivity being observable in the functional properties of neurons, will be revisitedin more detail in Chapter 3. Overall, an image has emerged of the tectum, and thenervous system as a whole, as exhibiting a precocious over-formation of synapsesthat are subsequently pruned during development to produce mature brain connec-tivity [24, 90, 159, 303]. However, this description is certainly a simplificationbecause both additions and retractions of filopodia and synapses occur throughoutdevelopment and into maturity, and mounting evidence suggests that a host of lo-cal signaling processes, initiated by sensory experience, mediate the decisions thataxons and dendrites make as they assemble themselves into functional circuits.221.4.2 Activity-dependent structural development and early learningThe structural development of dendrites is mediated by both innately-expressedand activity-dependent cues [e.g., 40, 113, 117, 136]. Prolific as initial dendriticoutgrowth and synaptogenesis are, even initial contacts must be highly specific,since neurons cannot possibly form even short-lived connections with all poten-tial synaptic partners. Several intrinsic signals first identified for their role inaxon guidance also affect dendritic branching and stratification, including Slit,Semaphorin, and their binding partners [64, 107, 224, 264]. Different cues act ondifferent types of neurons in different ways, contributing to the immense variety ofneuronal morphologies and patterns of connectivity of different neuron types. Thisis perhaps most evident in the fruit fly, Drosophila, where intrinsic, highly vari-able alternative splicing of DSCAM mRNA produces extracellular binding proteinswith specific binding patterns that determine the specificity of synaptic connections[40]. Slit and its receptor Robo guide the branching of one class of dopaminergicneurons, but not others [64].Locally secreted growth factors, such as the neurotrophins (NGF, BDNF, NT-3,NT-4) and associated tyrosine receptor kinase (Trk) receptors also influence pat-terns of dendrite growth, with some factors promoting destabilization [112], whileothers promote growth [e.g., 14], in a cell-identity-specific fashion. Interestingly,in Xenopus, BDNF and TrkB signaling appears to directly affect RGC axons butnot tectal dendrites [116, 166, 236].The ability of activity to guide structural growth patterns is of particular in-terest, because it may allow the structure of the brain to adapt to an organism’sspecific environment as it learns. Indeed, activity influences the growth patterns ofdeveloping dendrites and axons. David Hubel and Torsten Wiesel first describedthe abnormal functional development of cortex in animals deprived of early ex-perience [286]. Since then, numerous studies have demonstrated structural dif-ferences in neurons of animals raised with different types of sensory experience[e.g., 16, 52, 280]. In previously visually deprived tadpoles, visual stimulationinduces dendrite growth [250] and electrophysiologically-measured synapse for-mation [2]. Chronic glutamate receptor blockade decreases dendrite growth inorganotypic cultures of rat cortex [13], and both chronic and acute manipulations23affect Xenopus tectal neurons in vivo [216, 287]. Interestingly, many aspects ofstructural development appear to be unaffected by complete blockade of neuronalfiring across a brain region. For example, chronic blockade of firing activity in theretina with TTX produces no discernible changes in the dendritic morphology ofLGN neurons, and injection of TTX into the tectum does not alter dendrite growthin Xenopus [57, 88, 221, 287].Such global manipulations can be contrasted with studies that manipulated onlysingle neurons. Chronic hyperpolarization of a single neuron in the context of anunaltered circuit results in larger, less refined receptive fields but apparently nor-mal gross dendritic morphology [65], though dendritic motility was not assessed.In contrast, single-cell expression of dominant-negative glutamate receptors [94],which mediate excitatory synaptic transmission, decreased complexity of dendritesand produced more linear branching patterns. These single-cell manipulations andothers suggest that competitive mechanisms consisting of comparisons of activ-ity between neurons, or between dendritic inputs, may mediate activity dependentstructural and functional plasticity [e.g., 23, 53, 128].Single-neuron studies of activity-dependent growth have been further refinedby examining subcellular mechanisms. Several studies have described local den-dritic mechanisms that mediate filopodial formation, growth, stabilization, and re-traction. Here, calcium signalling plays a crucial role.In vitro studies have examined different sources of calcium and their effectson dendritogenesis, demonstrating an important role for local signals associatedwith synaptic transmission. Local calcium release from internal stores, triggeredby synaptic input, is necessary for maintenance of nascent dendritic processes inRGC cultures [157]. Similarly, local calcium transients induce the halting andstabilization of hippocampal neuron dendrites [158]. Synaptic activity, and in-flux of extracellular calcium were shown to halt actin dynamics in dendritic spines[32]. Calcium-dependent intracellular mechanisms, such as CAMKII and smallGTPases, act as downstream effectors of calcium influx, directing cytoskeletal re-modelling [76, 250, 289]. Active synapses are more likely to be maintained [154],and extracellular glutamate can trigger the formation of filopodia and assembly ofsynaptic machinery [140].Earlier studies demonstrating elaboration of dendrites within synapse-forming24regions led to the formulation of the ’synaptotropic hypothesis’: that the forma-tion and stabilization of synapses directs the progressive growth of dendritic andaxonal arbors, promoting elaboration into synaptogenic regions [278]. Accordingto the hypothesis, the formation of synapses drives dendritic branching, which inturn produces more synapses, producing positive feedback (this feedback must ofcourse be balanced by mechanisms that limit the total size of dendrites or synapsenumbers [e.g., 86]).The stabilization of dendritic regions containing synapses can be due to phys-ical adhesions formed by extracellular proteins and/or intracellular signals trig-gered by synaptic contacts and subsequent neurotransmission. Synapses are boundtogether by physical contacts that span the synaptic cleft. The molecules that medi-ate these interactions have specific binding partners and contribute to the specificityof synaptic partners [40, 248], potentially allowing synaptotropic growth to be di-rected into regions of appropriate synaptic partners. Direct two-photon imaging ofdendrites and fluorescently-labelled synaptic proteins indicates that new filopodiaare preferentially added at the sites of existing synapses, while instability and sub-traction of filopodia is preceded by loss of synapses [189]. Consistent with thesefindings, manipulations that stabilize synapses produce more compact but com-plex dendritic arbors, and destabilization of synapses produces more elongate, lessbranched arbors [43, 155].1.4.3 Experience-dependent functional plasticityFrogs and other oviparous vertebrates live independently during early developmentand require a functional visual system to, for example, avoid predators and captureprey. Tectal neurons begin to show visually evoked responses during early phasesof neurogenesis and dendrite growth at Stage 43, soon after the optic nerve firstarborizes within the tectum at Stage 39, on the third day after fertilization [84, 110,232]. Visual responses are highly plastic during early development, and can bemodified in 30 minutes with patterned sensory experience [69], making Xenopusan excellent model system for the study of functional plasticity induced by sensoryactivity.It has long been known that visual experience is extremely important for the25early development of the visual system. During early ’critical periods’ of develop-ment, experience establishes, for example, the appropriate binocular processing ofinputs arriving from the two eyes [286], as reflected in the responses of individualneurons, and their spatial arrangement into functional maps. Sensory experiencedirectly drives the establishment of fine-scale maps in the response properties ofindividual neurons [152], with spatial clusters of similar receptive fields formingin response to experience. Similar changes occur in motor cortex during the learn-ing of motor tasks [135]. The coding significance of these experience-dependentrearrangements across these varied brain regions is still not entirely clear, but theymay reflect optimizations in some combination of neuronal wiring costs, metabolicefficiency, and information processing. Chapter 2 and Chapter 3 introduce this pre-vious work in more detail, describes the formation of similar patterns of receptivefields within the Xenopus tectum, and investigates how rearrangements in neuronalcorrelations and receptive fields during experience-dependent plasticity affect pop-ulation encoding.Changes in neuronal firing are often mediated by changes in synaptic input.Long-term potentiation (LTP) and long-term depression (LTD) denote sustainedstimulus-induced increase or decrease, respectively, of synaptic currents. LTP andLTD occur in virtually every brain region in which they have been investigated,and are thought to be essential for learning. One important mechanism of LTP isspike-timing dependent plasticity (STDP), whereby the precise timing of excitatorysynaptic inputs relative to post-synaptic spiking determines the direction and mag-nitude of plasticity. Excitatory potentials occurring just before a cell fires, whichtherefore predict or drive those firing events, are strengthened, while those arrivingafter the cell fires are weakened. The first demonstration of STDP driving plasticityin vivo was performed in Xenopus retinotectal synapses [295]. Since then, STDPprotocols have been used to induce plasticity in a wide variety of preparations [35].Interestingly, it has been shown that STDP can drive the establishment of mo-tion receptive fields in the tectum [71, 299] via an intriguing mechanism that relieson the topographic spatial arrangement of synaptic inputs. Such studies highlightthe rich interactions between synaptic integration, dendritic morphology, plasticity,neuronal coding, and topographic functional maps.261.5 Scientific motivationsI carried out my doctoral work with two goals in mind. The first was to develop newtechnology and methods that will have long-lasting impacts on the study of learningand plasticity in vivo. The second was to demonstrate the effectiveness of thesemethods by using them to investigate biological questions with broad relevance toneuroscientists.The choice to focus on development, rather than the more mature brain, hasboth methodological and theoretical motivations. Methodologically, the motiva-tions are many. The developing brain is smaller, making it more accessible tooptical methods and more easily held in place. It undergoes changes more rapidlythan the mature brain, providing an exciting opportunity to watch the full courseof learning or plasticity as it occurs. Experiments can be executed more quickly,because animals reach the stage of study sooner.The other great motivation for studying development, however, is a broad strat-egy for understanding how the brain, or any complicated system, works. A lot canbe learned about a machine by watching its parts being put together. A great dealof the challenge in applying artificial neural networks, at present, is in setting so-called metaparameters, dictating the numbers and types of nodes a network startswith, and their possible connections. This initial arrangement of connections, be-fore training begins, greatly impacts how a neural network performs. In addition tothe challenge of identifying learning rules that alter connection strengths betweenneurons, it is still not clear how to first assemble a network that will be capableof learning effectively. The solution to this problem may be to recapitulate the de-velopment of our own brains, which start simply, but rapidly grow in complexityaccording to some as-yet-unknown set of rules. If our ultimate goal is to build abrain, one hope is that we can learn to do that by watching brains build themselves.27Chapter 2Introduction to Manuscript 1:Nonlinear Population Encodingin the TectumAt the time I started my doctoral work, we were beginning to learn a lot aboutfunctional responses of individual neurons in the optic tectum. The receptive fieldsof individual neurons and the information they encode had been mapped acrossdevelopment [e.g., 59, 71, 188]. Methods had been developed to rapidly induceplasticity with natural visual stimulation [69, 296]. An important mechanism ofplasticity, STDP, had been shown to mediate experience-induced plasticity by thelab of Mu-Ming Poo [295]. However, we were lacking an understanding of howthe tectum as a whole encodes information. What challenges does the brain face incoordinating responses, and the plasticity of those responses, to optimally encodethe environment? For example, it is very important that different neurons respondto different stimuli, regardless of how well each neuron encodes its own preferredstimulus. At the limit where all neurons have precisely the same response pattern,the whole circuit conveys no more information than a single neuron does.It has long been known that maximizing the amount of information conveyedby a set of neurons requires minimizing correlations in neuronal responses [15]. Inpractice, however, the brain probably trades off bandwidth for redundancy, and thepossible encoding strategies are limited by the physical structure of brain circuits28and the metabolic cost of neuronal connections [48, 49].To address these questions, we needed to measure correlations in neuronal re-sponses, and assess how patterns of activity change as the brain learns. Electro-physiologists had long been interested in such questions, but were limited in thedensity at which neurons could be sampled in a brain region. Electrophysiologicalstudies in the retinal explants approached the limit of complete neuronal sampling,and provided interesting insights into population encoding of these early visual sig-nals. It was not until the adoption of two-photon calcium imaging that it becamepossible to monitor activity in a large contiguous population of neurons in vivo,making it possible to answer questions regarding encoding by correlated activityin the awake brain. An electrophysiological study had been published on the de-velopment of correlated activity in the tectum, but this study measured differentneurons on separate trials, then predicted correlations assuming these neurons areindependent, a somewhat circular argument [222]. Where studies had used imagingto measure many neurons simultaneously [69, 135, 152], the network was largelytreated as a collection of individual neurons, and how the population encodes stim-uli was not addressed.The following study marks the first investigation of network-level sensory re-sponses in Xenopus, and was, to my knowledge, the first study to track encod-ing of a stimulus in a densely sampled neuronal population across long-lastingexperience-dependent plasticity. It also introduces a number of methods I devel-oped for the processing of calcium imaging data, which continue to be used bymultiple labs. These methods include optimal linear spatial filtering using itera-tive singular value decomposition, optimal linear fitting of baseline fluorescenceusing a Kalman Smoother, and a fast method for decoding population activity us-ing pairwise activity statistics. The most important conclusion of this work is thatthe performance of a brain circuit is not very sensitive to the encoding fidelityof individual neurons, and is rather more sensitive to the distribution of encodingproperties over different neurons in the network. Even when individual neuronsimprove their encoding abilities, the network as a whole can suffer if changes arenot coordinated (as is the case when NMDA receptors are blocked during training).As a result, researchers interested in how well a brain circuit processes informationmust necessarily track activity in many neurons across that circuit.29Chapter 3Manuscript 1: FunctionalClustering Drives EncodingImprovement in the AwakeDeveloping Brain3.1 SummarySensory experience drives dramatic structural and functional plasticity in develop-ing neurons. However, for single-neuron plasticity to optimally improve whole-network encoding of sensory information, changes must be coordinated betweenneurons to ensure a full range of stimuli is efficiently represented. Using two-photon calcium imaging to monitor evoked activity in over 100 neurons simultane-ously, we investigate network-level changes in the developing Xenopus laevis tec-tum during visual training with motion stimuli. Training causes stimulus-specificchanges in neuronal responses and interactions, resulting in improved populationencoding. This plasticity is spatially structured, increasing tuning curve similarityand interactions among nearby neurons, and decreasing interactions among dis-tant neurons. Training does not improve encoding by single clusters of similarlyresponding neurons, but improves encoding across clusters, indicating coordinated30plasticity across the network. NMDA receptor blockade prevents coordinated plas-ticity, reduces clustering, and abolishes whole-network encoding improvement.We conclude that NMDA receptors support experience-dependent network self-organization, allowing efficient population coding of a diverse range of stimuli.3.2 IntroductionThe vertebrate brain exhibits intricate functional organization at many differentspatial scales, from cortical microcolumns dedicated to processing specific recep-tive field properties, to large domains such as somatotopic maps. It is thought thatthis organization of neurons according to shared function optimizes efficiency andeffectiveness of neural processing. During development, the structure [51, 87] andfunction [69, 71, 119, 152] of sensory neural circuits are actively guided by bothendogenous signals and environmental stimuli. However, it is not well understoodhow these changes lead to improved brain function.Here we investigate how plasticity affects developing visual system perfor-mance from the perspective of sensory encoding – the representation of sensorystimuli by activity in populations of brain neurons. Neuronal responses are inher-ently noisy and vary across presentations of the same sensory stimulus, limitinghow much information can be encoded by a single neuron [89]. To optimally en-code environmental stimuli in the presence of noise [273], sensory circuits mustbe organized to balance redundancy, which makes network encoding less sensitiveto neuronal noise, with the ability to encode a diverse range of stimuli. In the ab-sence of noise, a given stimulus feature can be fully conveyed by a small numberof neurons, and to maximize efficiency, other neurons should then encode differ-ent features. If neuronal responses are more variable, more neurons are requiredto reliably convey a given feature. The optimal response pattern for each neuronthus depends on the response properties of other neurons in the network and thereliability of those responses.Encoding is also affected by neuronal interactions. For example, neuronalinteractions may be organized to remove correlations from the network’s input(decorrelation) [15], making the neural code more efficient, and neuronal ensem-bles can synergistically encode information not available from individual neu-31rons [240]. Strategies that coordinate neuronal interactions and optimize encod-ing have been identified in artificial networks under various conditions [273], andencoding schemes have been described and evaluated in mature neural circuits[83, 134, 176, 202]. Further studies have shown that adaptation of neuronal recep-tive fields [245] and correlations [92] can tune encoding in response to changes insensory stimuli in vivo. However, little is known about how encoding schemes ariseduring development or how they are altered during early learning, when dynami-cally growing neural circuits first wire themselves together. Evaluating networkencoding requires simultaneous observation of many neurons, and understandingearly network refinement requires monitoring those networks over the course oflearning and development.The visual system of the X. laevis tadpole has been extensively studied as amodel of neuronal and neural circuit development [43, 51, 66, 69, 71, 188, 211,250]. Transparent albino tadpoles allow minimally invasive in vivo observation ofrapid sensory circuit development, from differentiation [288] to mature neuronsdriving behavioral responses [263]. Studies in the developing brain have describedmechanisms controlling large-scale circuit patterning [258], fine-scale morphogen-esis [43], and rules by which synapses [295], single neurons [71, 188], and smallgroups of neurons [92] refine their response properties with experience. However,it is largely unknown how these developmental changes contribute to network en-coding performance, or how plasticity is coordinated across neurons to producefunctional large networks.Here we use in vivo two-photon calcium imaging [69, 135, 152, 188] to monitornetwork activity and plasticity during early receptive field development in Xenopustadpole optic tectum [232] as we train the brain to respond to a set of visual motionstimuli. Training causes stimulus-specific changes in evoked neuronal responsesand increases stimulus information conveyed by neuronal firing. Decoding of net-work activity using computational models [215] becomes more accurate over thecourse of visual training. Training induces spatial clustering of receptive fields andcorrelations by increasing tuning curve similarity and network interactions amongnearby neurons and decreasing interactions among distant neurons. Blockade ofN-methyl-D-aspartic-acid type glutamate receptors (NMDARs) blocks spatiallygraded plasticity, and prevents decoding improvement with training. By com-32paring decoding in single clusters and groups, we show that increasing networkperformance arises from NMDAR-dependent improvement in encoding of stim-ulus information across clusters, while encoding within single clusters does notimprove with training. We propose that NMDARs support experience-dependentfunctional clustering leading to local redundancy and distant decorrelation, andpromote receptive field diversity by preventing loss of underrepresented receptivefields. These results highlight contributions of network-level organization to theperformance of sensory systems in vivo and identify mechanisms by which visualexperience directs improvement in whole-network function.3.3 Results3.3.1 In vivo monitoring of neuronal firing rates with two-photoncalcium imagingIn vivo two-photon calcium imaging allows simultaneous monitoring of somaticcalcium transients, induced by neuronal firing, in hundreds of neurons in the ver-tebrate brain [69, 135, 152, 188, 190]. We used this method to monitor correlatedvisually evoked responses across the optic tectum, which requires that firing-ratemeasurements are accurate on a single-trial basis and not averaged across trials[215]. Optical readout of calcium transients is hindered by drifting baseline fluo-rescence (F0), bleaching, and saturation, and involves fundamental tradeoffs be-tween imaging area and quality of signal. Moreover, the relationship betweenaction potentials and calcium levels is complicated by the temporal dependenceof calcium concentrations on spiking history and nonlinearities in calcium influx[279]. To overcome these limitations and improve signal quality, we developedtechniques for automated video segmentation to track cell boundaries on the ba-sis of morphology and temporal pixel correlations, spatial filtering to weight thecontributions of pixels within a given cell, and F0 estimation using optimal linearmethods ( Figure 3.10). To extract firing rates from fluorescence data we em-ployed a spike inference algorithm, which takes into account temporal dependenceand nonlinearities in signal [279].To assess the effectiveness of these methods for measuring single-trial evoked33firing rates in the awake brain, we performed in vivo loose seal patch clamp electro-physiological recordings to monitor action potential spiking during simultaneouscalcium imaging and visual stimulation (Figure 3.1e and Figure 3.11). We com-pared firing rates obtained from electrophysiological recordings to two measures ofneuronal firing obtained from fluorescence data: peak ∆F/F0 [69] and firing ratesinferred from spike inference. Though both measures showed significant correla-tions to actual firing, inferred firing rates outperformed peak ∆F/F0 in all neuronsrecorded (Figure 3.12), possibly because burst durations and interspike intervalswere long (Figure 3.1f), resulting in imperfect summation of peak calcium cur-rents. The relationship between inferred firing rates and actual spike counts waslinear (Figure 3.12), showing that in vivo calcium imaging and spike inference isan effective method for monitoring firing rate fluctuations in tectal neurons.We first used rapid two-photon imaging and firing rate inference to character-ize motion receptive fields in untrained tadpoles. Motion stimuli consisted of darkbars moving over a light circular background in each of eight directions (see Meth-ods), with low contrast so as to better detect improvements in neuronal responseswith subsequent training. We found that most motion-responsive tectal neuronsrespond either symmetrically to pairs of opposing directions (orientation selectiv-ity, 59.1%±5.0% of cells; mean ± standard deviation [SD]), and/or specificallyto a narrow band of directions (direction selectivity, 66.3%±11.1%). Neuronsresponding to two opposite directions while strongly favoring one direction canshow both selectivities (36.7%±9.3%). Average responses of individual neuronsto each stimulus direction, called tuning curves, show varying selectivity in a to-pographic organization (Figure 3.2). These results demonstrate the effectivenessof two-photon imaging and spike inference in measuring receptive fields across acontiguous brain network in vivo.3.3.2 Tectal network responses to visual stimuli exhibit noisecorrelations indicating functional interconnectionsBesides the single-neuron properties described above, networks of neurons oftenshow correlations in their firing patterns. Neurons with similar tuning curves showsignal correlations because their firing is driven by the same stimuli [83]. Notably,real neuronal responses also show trial-to-trial deviations from their tuning curves.34Figure 3.1: In vivo imaging of evoked network activity in the unanesthetizeddeveloping brain. (a) Experimental setup. Motion stimuli were presented tothe left eye of awake, immobilized Xenopus tadpoles while imaging the rightoptic tectum. Neurons in the tectum (green circles) extend dendrites to re-ceive visual input from retinal ganglion cells (red) of the contralateral eye.(b) Transmitted light image of a tadpole brain seen through the head. Greenbox, optic tectum. (c) Two-photon image of optical section corresponding togreen box in (b). Tectum is loaded with OGB1-AM, a calcium-sensitive dye.Red box corresponds to the region of tectum monitored in our experiments.(d) Two-photon image of a patched neuron in awake tectum. (e) Simulta-neous recording of somatic fluorescence (∆F/F0, top) and action potentials(green) in response to full field light stimuli of varying intensity, with actual(gray) and inferred (black) firing rates in the 5 s following each stimulus.(f) Expanded voltage trace for electrophysiological recording. Pink shadingmarks time of stimulus. The electrical transients bounding the stimulus pe-riod are clipped. Colored dots mark individual action potentials, which aremagnified in the boxes at bottom.When these trial-to-trial deviations are shared, because of common input or inter-connections, neurons are noise correlated (Figure 3.3a) [83]. Noise correlations arethus correlations in neural firing patterns that are not explained by shared receptivefield properties. Noise correlations can be positive or negative, can differ acrossstimuli, and do not require signal correlations to be present. When trial-to-trialvariability is not shared, neurons are independent.The contribution of neural correlations to network activity patterns is difficult35Figure 3.2: Orientation and direction responses in optic tectum. (a,d) Mapsof direction and orientation selectivity in naive Xenopus tectum obtainedthrough rapid two-photon imaging and firing rate inference. Stimuli weredark bars moving over a light background for 600 ms in eight directions.Black circles mark neurons that responded significantly to stimuli. Coloredarrows mark preferred directions (a) and orientations (d) of neurons showingstimulus specificity. Coronal optical section, rostrum to the left. Scale bar =20 µm. (b,e) Tuning curves of a direction- (b) and an orientation- (e) selectiveneuron highlighted in (a,d). Error bars denote SEM. (c,f) Average temporalresponse of the two neurons to each stimulus direction. Colors match those in(b,e). Gray bar marks time of stimulus presentation. All measures calculatedfrom n = 48 stimulus presentations for each of eight directions (1 h).to determine when observing only individual neurons or small groups [215]. Ef-fects of pairwise interactions on network encoding may only be detectable if manyneurons are taken into account, and even small pairwise interactions strongly im-pact activity patterns when large networks are considered [241]. Thus, when neu-rons are significantly noise correlated, understanding network function requiresobserving activity in large groups of neurons simultaneously [9, 215]. Numerousstudies have investigated the presence of noise correlations in vivo [83, 92, 235],their effects on encoding [8, 134, 195, 273], and the consequences of ignoring them[143]. Conclusions on these topics vary with the brain regions and response prop-erties being studied. It is agreed, however, that the presence and impact of noisecorrelations determines the experimental and theoretical methods we must use tounderstand neural information processing.Examining multineuronal firing patterns elicited by motion stimuli, we find that36Figure 3.3: Tectal noise correlations influence network decoding (a)Recorded responses of two neurons (black and grey) in the same tadpoleto eight consecutive presentations of the same stimulus. Responses vary inamplitude around their means (dotted lines). These neurons were noise corre-lated: variations in amplitude were shared. (b) Distribution of measured pair-wise noise correlations (black dotted e) taken over a 1-h stimulation period,and values expected if neurons were independent (gray). Noise correlationswere more positive (p < 105, t-test) and more variable (p < 108; X2 vari-ance test) than chance. (c) Scatterplot of pairwise linear noise correlationsmeasured in two consecutive 30-min periods. Consecutive noise correlationmeasurements are correlated (r = 0.41, p < 108; linear regression). (d) Distri-bution of decoding errors under independent and noise correlation decodingof actual response patterns (left) and with responses shuffled for each stim-ulus type to remove noise correlations (right). Data from seven tadpoles,277 neurons (b,d), 384 stimulus presentations (c), 192 stimulus presentationseach 30 min. Error bars denote SEM. ∗ : p < 0.05; ∗∗ : p < 0.01.37noise correlations are prominent in the awake developing tectum (Figure 3.3b,c).Noise correlation measurements were correlated over consecutive 30-min periods(Figure 3.3c). Noise correlations varied across stimuli (Figure 3.13), and may thusconvey stimulus information not present in single-neuron responses (Figure 3.14)[179]. Noise correlations between neurons tended to have the same sign as sig-nal correlations (Figure 3.15), indicating that many tectal noise correlations reflectshared errors in similarly responding neurons. These results demonstrate that tec-tal noise correlations can be measured with two-photon calcium imaging and mayhave consequences for information processing in this network.3.3.3 Tectal noise correlations can encode stimulus information, butimpair overall network performanceNoise correlations can both help and hurt network stimulus encoding, dependingon how they vary with stimuli and the response properties of neurons in the network[8, 83, 179, 195, 273]. Because noise correlations are prominent in developing tec-tum and are stimulus dependent, we expected that knowledge of noise correlationsmay be important for downstream neurons to extract all available information fromnetwork activity patterns. However, because we found that tectal noise correlationslargely reflect shared errors, we expected removal of noise correlations from pop-ulation activity would increase the amount of information available in those firingpatterns [89, 273]. To test these predictions, we constructed two model decoders:one that takes into account pairwise noise correlations, and an optimal indepen-dent decoder, which ignores noise correlations. A decoder is a model based on aset of real network responses, which takes a second set of measured activity pat-terns as input and predicts the inducing stimuli [213]. Decoders thus perform thesame task as downstream neurons to recover stimulus information from upstreamnetwork activity. By building decoders, we can ask two distinct questions. Regard-ing encoding: Would population encoding accuracy be altered if noise correlationswere somehow abolished? Regarding decoding: Is knowledge of noise correlationsnecessary to fully decode network activity from a population response? We findthat abolishing noise correlations by shuffling neurons’ responses across trials ofeach stimulus improves accuracy of both decoders (Figure 3.3e). This finding con-firms that encoding would improve overall if responses were uncorrelated, likely38because the noise correlations we observe are largely shared errors among simi-larly responding neurons. Nevertheless, ignoring noise correlations in actual datasignificantly reduced decoding accuracy (Figure 3.3e). This outcome suggests thatsensitivity to noise correlations would help downstream neurons to decode firingrates in this network. However, changes in neural response properties over thesampling period can make noise correlations important for decoding, even in caseswhere they would not be important if responses were stationary [191]. To properlyevaluate the contribution of noise correlations to decoding we must thus determinewhether tectal responses change with repeated stimulus exposure, and manipulatethis contribution by altering neuronal interactions.3.3.4 Visual training induces neural plasticity, improving stimulusencodingDuring development, sensory experience drives dramatic neural plasticity [69, 71,135], but how these changes lead to improved circuit function is not understood.To investigate how stimulus encoding changes in response to visual experience, wepresented tadpoles the eight motion stimuli of different directions repeatedly over 2h. This training improved sensory responses over time, increasing dynamic rangeand response reliability (Figure 3.4a,f,g). Training also shifted neural responseproperties, increasing the proportion of neurons showing combined orientation anddirection selectivity and decreasing the proportion showing only direction selectiv-ity (Figure 3.4e). Encoding was enhanced, evident from increased stimulus mutualinformation conveyed by both individual neurons and neuron pairs (Figure 3.4b,c),and improvement in both independent and noise-correlation–based decoding ofwhole-network activity (Figure 3.5a). To further demonstrate that visual experi-ence modifies network encoding over time, we split the stimulation period into two60-min epochs (early and late), and built decoders for each using firing statisticsfrom either the same or the opposite epoch. Both independent and noise correlationdecoding improved from early to late epochs, and decoding performance decreasedwhen using firing statistics from the opposite epoch (Figure 3.5c), demonstratingthat experience changes how developing brain networks encode stimuli.39Figure 3.4: Effects of visual training on single-neuron response properties.(a) Tuning curve dynamic range, the fraction by which a neuron’s firingchanges in response to different stimuli during early and late epochs. (b,c)Stimulus mutual information conveyed by single neuron (b) and neuron pair(c) firing patterns. Upper asterisks denote difference in the change with treat-ment. Lower asterisks denote significant change across epochs (paired t-test).(d) Evoked firing rates in control (black) and MK-801 treated (gray) tadpolesduring first hour of stimulation. Each point corresponds to a single tadpole;error bars denote standard deviation across neurons within a given tadpole.MK-801 does not acutely affect evoked firing rates (t-test, p = 0.61). (e)Proportion of neurons showing direction (yellow), orientation (blue), both(green), or neither (red) selectivity in control (top) and MK-801(bottom)treated tadpoles, in the first (left) and second (right) hour of stimulation. As-terisks denote significant change across epochs (paired t-test). (f,g) Meannormalized amplitude (f) and response reliability (g) over the course of visualtraining (black). Reliability increased with training (ANCOVA, p < 0.01).Neither measure was affected by MK-801 (gray) (ANCOVA, p > 0.05). Re-liability is the proportion of evoked responses with amplitude larger than themedian spontaneous firing rate. Error bars denote SEM. ∗ : p < 0.05.40Figure 3.5: Training induces NMDAR-dependent improvement of whole-network encoding. (a) Time course of noise-correlationbased (red) and inde-pendent (blue) decoding performance. Light curves, improvement is blockedby MK-801. Bars denote early and late epochs. Decoding improvementis the decrease in decoding error relative to the independent decoder at thefirst timepoint. Both decoders improved from early to late epochs in control,but not MK-801treated tadpoles (paired t-tests). (b) Decoding error of con-trol (left, blue) and MK-801 treated (right, red) tadpoles over first hour ofstimulation. Lighter shades denote decoding using the optimal independentdecoder, darker shades mark noise correlation-based decoding. (c) Improve-ment, relative to the early epoch, of decoders trained on data from early (lefttwo panels) or late (right two panels) epochs, used to decode early or lateneuronal firing patterns. Performance decreased when decoding the epochon which the decoder was not trained (center two panels; ANOVA). Asterisksin rightmost panel denote significant difference from corresponding value inleftmost panel. Error bars denote SEM. ∗ : p < 0.05; ∗∗ : p < NMDAR blockade does not alter basal neuronal or networkresponsesNMDARs act as molecular detectors of correlations between pre- and post- synap-tic firing and are known to mediate several types of functional [69, 153, 295] andstructural [43, 217, 230, 250] plasticity in tectal neurons. To investigate NMDARroles in shaping neuronal correlations and network-level encoding, we tested tad-poles treated with MK-801, a noncompetitive NMDAR antagonist. MK-801 wasinfused directly into the tectum and applied to tadpole bath, conditions we find tocompletely block NMDAR synaptic currents evoked by optic nerve stimulation invivo (Figure 3.16). With calcium imaging, we first investigated the acute effectsof MK-801 on neuronal firing and network performance. NMDAR blockade did41not affect basal neuronal firing rates (Figure 3.4d), or the relative proportions ofdifferent types of motion stimulus selectivities across neurons (Figure 3.4e). MK-801 treatment also did not alter basal network encoding performance (Figure 3.5b)or neuronal reliability (Figure 3.4g). Previous studies have also found that NM-DAR antagonism does not acutely affect tectal motion responses [71], and MK-801does not acutely affect cortical response properties [58], or temporal properties ofevoked tectal firing [211]. Consistent with these studies, we find that NMDARcurrents do not contribute strongly to visually evoked responses in this system.3.3.6 NMDARs mediate experience-driven network plasticityTo investigate NMDAR effects on experience-dependent network plasticity, weperformed the previously described visual training protocol using moving bar stim-uli of eight directions with tadpoles treated with MK-801. We find that distinctcomponents of experience-dependent plasticity are NMDAR dependent and in-dependent. In contrast to untreated tadpoles, training did not shift the propor-tions of different response selectivities in MK801-treated tadpoles (Figure 3.4e).MK-801 reduced improvement in whole-network encoding, dynamic range, andstimulus information of neuron pairs, but not in single-neuron stimulus informa-tion (Figure 3.4a-c and Figure 3.5a). MK-801 also blocked increases in decodingperformance when the stimulation period was split into early to late epochs (Fig-ure 3.5c). In fact, correlation-based decoding with MK-801 worsened from earlyto late epochs when decoded with each epoch’s own training statistics, suggestinga strong role for NMDARs in changes to network interactions and their effects onpopulation encoding.Further aspects of network plasticity observed with training were NMDAR-independent. MK-801 treatment did not affect the time course of neuronal reli-ability or mean response amplitude (Figure 3.4f,g), and a significant portion oftraining-induced increases in mutual information and dynamic range remained inMK-801 treated tadpoles (Figure 3.4ac).423.3.7 Training-induced plasticity and encoding improvement arestimulus specificTo determine whether improvements in network function are specific to the trainingstimuli, we trained tadpoles for 1 h with four of the eight motion stimuli (0, 45, 90,135 deg), followed by probing with the full eight stimuli (0−360 deg), and com-pared network responses to trained versus untrained stimuli. Training improveddecoding of the trained stimuli only for both the correlation-based (Figure 3.6a)and independent (unpublished data) decoders. Relative to naive tadpoles, trainingwith four stimuli increased the proportion of neurons showing combined orienta-tion and direction selectivity and decreased the proportion of responsive neuronsshowing no selectivity (Figure 3.6b). Among direction-selective neurons, directionof selectivity favored the center of the trained directions (Figure 3.6c,d). Dynamicrange was higher in response to trained stimuli, while reliability and evoked firingwere not significantly different between trained and untrained stimuli (unpublisheddata). These results demonstrate that training-induced changes are stimulus depen-dent and favor encoding of the specific visual stimuli experienced.3.3.8 Training induces anatomically structured network plasticityImaging a contiguous population of neurons allows us to relate experience-dependentplasticity to anatomical structure [152]. Similar to visual cortex [190], optic tec-tum has a precise functional architecture [188], where nearby neurons exhibit sim-ilar receptive fields and thus strong signal correlations (Figure 3.2). We also findthat nearby neuron pairs show higher noise correlations and a significant associa-tion between stimulus and noise correlation, consistent with locally shared inputor direct connectivity. We tracked these measures across epochs of visual train-ing among nearby (<25 µm), moderate (25 – 50 µm), and distant (50 – 75 µm)neurons. Tectal somata have diameters of 10 – 15 µm. These measures changedin a distance-specific manner as visual training improved network encoding. Vi-sual training increased signal correlations among nearby but not more distant neu-ron pairs (Figure 3.7a). Visual training also increased nearby noise correlationsand decreased distant ones (Figure 3.7b). Larger signal and noise correlations fornearby neurons indicate increased local redundancy with training, likely because43Figure 3.6: Training-induced changes are stimulus-specific. (a) Decoding error for each direction in tadpoles trained with four ofeight stimuli (0−135 deg), using the correlation decoder. Gray, decoding error of naive control tadpoles. Training-induced decod-ing improvement is specific to the trained stimuli. (b) Proportion of neurons showing direction (yellow), orientation (blue), both(green), or neither (red) selectivity in tadpoles trained with four stimuli. Asterisks denote significant difference from correspond-ing proportion in naive control tadpoles. (c) Angle histogram of preferred directions of direction-selective neurons in tadpolestrained with four stimuli. Points are the proportion of neurons with center directions falling between adjacent stimulus directions.Pink shading indicates the trained directions. Responses strongly favored the center of the trained directions (one-sample t-test,p < 105). Gray dotted line indicates preferred directions in naive control tadpoles. (d) Map of direction selectivity in a tadpole af-ter training with four stimuli. Black circles mark neurons showing significant direction selectivity. Colored arrows mark preferreddirections. Error bars denote SEM. (a - c) n = 3 tadpoles (152 neurons). ∗ : p < 0.05; ∗∗ : p < 0.01.44of strengthening of shared stimulus inputs. The decrease in distant noise correla-tions, however, suggests that encoding strategies thought to improve mature circuitperformance [15, 273], such as network decorrelation, can result from plasticityduring early experience in vivo. These results show that visual training leads toanatomically structured network refinement.NMDAR blockade prevented this refinement and led to degradation of fine-scale functional organization over time. Here, signal correlations were increasedequally for all neuron pairs, regardless of spatial distance, reducing receptive fielddiversity across the tectum (Figure 3.7c and Figure 3.8b). MK-801 also blockedtraining-induced changes in noise correlations (Figure 3.7d), suggesting that de-velopment of efficient network correlation structure is NMDAR-dependent. Theloss of spatial organization we observe with MK-801 over time is consistent withlack of competition between locally represented and distant inputs in the absenceof NMDAR transmission.MK-801induced changes in plasticity were recapitulated by training with thefour-stimulus subset. Tuning curve similarity was greater over untrained stimulithan trained stimuli across moderate and distant, but not nearby, neuron pairs (Fig-ure 3.7e). Networks showed strongly decreased noise correlations to trained stim-uli, while noise correlations to untrained stimuli increased above levels in naivetadpoles. These results show that training with a set of stimuli affects the encodingof unpresented stimuli, and stimuli can compete in determining network connec-tivity (Figure 3.7f) [74, 230, 267].3.3.9 Coordination between neuronal clusters supportsexperience-dependent encoding improvementVisual training induces remarkable spatially divergent plasticity. On one hand,training-induced encoding improvement is associated with lower signal and noisecorrelations among distant neurons. On the other hand, local plasticity opposes thistrend, increasing redundancy between nearby neurons over the course of visualtraining. To determine how these opposing forces contribute to overall networkimprovement, we grouped neurons according to receptive field so as to monitorstimulus decoding within clusters of similarly responding neurons over time. Con-sistent with our measurements of tectal signal correlations, functionally defined45Figure 3.7: Training strengthens clustering of receptive fields and networkcorrelations. (a – d) Tuning curve similarity (a,c) and mean noise correla-tion (b,d) of neuron pairs binned by spatial distance, during early (teal) andlate (purple) epochs, in control (a,b) and MK801-treated (c,d) tadpoles. (e,f)Tuning curve similarity (e) and noise correlations (f) in tadpoles trained withfour stimuli (0−135 deg), binned by distance, in response to trained (orange)and untrained (yellow) stimuli. (f) Shaded area highlights the range of plotsin (b,d). Noise correlations to untrained stimuli were significantly lower thanin naive control animals (p < 105, two-way ANOVA) and those to trainedstimuli were significantly higher than in naive controls (p < 105, two-wayANOVA). Error bars denote SEM. Control, n = 7 tadpoles (277 neurons),MK801, n = 7 tadpoles (255 neurons) (e,f) n = 3 tadpoles (152 neurons).∗ : p < 0.05; ∗∗ : p < 0.01.46groups showed significant spatial clustering (Figure 3.17). Interestingly, decodingsuccess of single clusters did not change with training (Figure 3.8c), suggestingthat interactions between clusters may be more important in supporting overallencoding improvement. To understand how well clusters interact to encode infor-mation, we measured intercluster cooperation, which we defined as the decodingperformance of two clusters taken together minus the maximum decoding perfor-mance of either taken alone. Cooperation is high when clusters encode distinctinformation or encode information synergistically [240], and low when clustersencode the same information. Notably, cooperation increased with visual train-ing in control tadpoles, while training during NMDAR blockade decreased clustercooperation (Figure 3.8d). To further investigate how plasticity in neuronal in-teractions contributes to changes in encoding performance, we again removed thecontribution of noise correlations by shuffling neuronal responses prior to decod-ing (as in Figure 3.3e). Shuffled decoding accuracy did not change from EARLY toLATE epochs, even as nonshuffled decoding accuracy increased in control tadpolesand decreased in MK-801treated tadpoles (Figure 3.18), consistent with a role forneuronal interactions in driving the changes in network performance we observe.These results show that improvements in the brain’s ability to represent visual stim-uli are not due only to improved encoding in single neurons or local groups, but aredriven strongly by changes in the functional organization of the sensory network.3.4 DiscussionThe functional organization of the brain contributes to effective neural processing,and neurons can coordinate or compete to encode distinct stimulus dimensions[83, 96, 176]. We find that developmental plasticity in response to visual experi-ence establishes such organization in the optic tectum (Figure 3.9). This plasticitystrengthens divisions between microarchitectural brain regions specialized to en-code distinct stimuli that the organism experiences. Visual training improves bothindividual neuron and network response properties, but single-neuron changes onlyweakly impact network performance. This weak reliance on single neurons likelyarises because the tectal network is organized in local receptive field clusters thatexhibit high redundancy; information gained from improved fidelity in any indi-47Figure 3.8: NMDAR-dependent coordination between clusters supports network encoding improvement. (a) Preferred directionsof example control (top) and MK801-treated (bottom) tadpoles during early (left) and late (right) epochs. Scale bar = 20 µm. (b)Receptive field diversity across the tectum during early and late epochs, in untreated (black) and MK801-treated (gray) tadpoles.Diversity decreased with training in MK-801 treated tadpoles (paired t-test, p < 0.05). (c) Mean decoding error of independent(blue) and correlation-based (red) decoding of single clusters during early and late epochs, in untreated and MK801-treated (lightershades) tadpoles. (d) Mean decoding cooperation (decoding performance of two clusters taken together minus the maximum de-coding performance of either taken alone) during early and late epochs, in untreated and MK801-treated tadpoles. Cooperationincreased in control tadpoles and decreased in MK-801 treated tadpoles with training (paired t-tests). Number of clusters: un-treated, n = 29; MK801, n = 25 clusters per epoch in seven tadpoles. Error bars denote SEM. Control, n = 7 tadpoles, 277 neurons;MK801, n = 7 tadpoles, 255 neurons. ∗ : p < 0.05; ∗∗ : p < 0.01.48vidual neuron tends to already be available from other nearby neurons. Our resultsshow that the functional organization of the network plays a larger role in the over-all improvement of population encoding with training. This organization consistsof specialization by distinct groups of neurons to convey distinct information, astraining drives distant neurons to become more independent while strengtheninglocal redundancy. This spatially driven plasticity arises from forces acting to in-crease or decrease functional connectivity in the tectum on different spatial scales.Spatial clustering of functional properties is a common feature in the brain[69, 135, 152, 190], which can lead to redundant local encoding. Redundancy isimportant in mitigating effects of variability of individual neuronal responses. Be-cause neuron response fidelity is fundamentally limited by both physics [22] andphysiology [145, 242], redundant encoding by groups can be more practical thandecreasing variability in single neurons. Moreover, response properties in a givenbrain volume are limited by the availability of presynaptic partners, as each neuronmust search its local environment for appropriate connections. In tectum, promi-nent inputs are likely to be shared by nearby neurons because of the localized ar-borization of retinal ganglion cell axons [230], and plasticity that strengthens thoseinputs thus promotes local redundancy. Finally, local similarity can make wiring ofdeveloping networks more economical [49], as neurons responding to a particularstimulus should then receive inputs from a restricted anatomical region. Learning-associated functional clustering and correlation changes similar to those describedhere have been described in mouse motor cortex [135], raising the possibility thatcommon constraints drive functional optimization across network structures andfunctions.Measurement of single-trial firing rates enables monitoring of redundancy andnoise correlations in large populations of tectal neurons. We found that noise corre-lations can be repeatably measured and are altered by training in an experience- andNMDAR-dependent fashion. These results show that two-photon calcium imagingcan be used to investigate shared connections across contiguous brain regions andhow these change in vivo. However, the anatomical substrates underlying tectalnoise correlation plasticity remain unclear, since noise correlations could arise ei-ther from shared retinal inputs or intratectal connections. Plasticity in noise corre-lations may indicate formation and elimination of these connections or alteration49Figure 3.9: Schematic of receptive field and noise correlation plasticity fortrained (red) and untrained (blue) stimuli. Tectal neurons are represented ascircles, circle color marks preferred direction (red, down; blue, up), and dot-ted lines represent noise correlations. Training with down direction increasesand clusters receptive fields oriented toward the trained stimuli and decreaseslong-distance noise correlations (dashed lines). Receptive fields preferringuntrained stimuli (blue) are reduced, and noise correlations to these stimuliare increased on all spatial scales. Note that noise correlations can differacross stimuli and are not necessarily determined by neurons’ preferred di-rections.50of synaptic strengths. We found that accounting for noise correlations improvesdecoding of tectal population activity, but this effect could be due to changes inneural activity patterns over the stimulation period [191]. However, the specificeffects of NMDAR blockade on noise correlation-based decoding with trainingsuggest that noise correlations are indeed important for decoding tectal activity(Figure 3.5c and Figure 3.18). Despite their importance to decoding, we found thatthe presence of noise correlations does not improve network encoding. The reduc-tion of correlations typically enables networks to convey more information [15].Indeed, we found that artificially eliminating noise correlations in network activ-ity data increased decoding performance. Networks whose function is limited bythe number of neurons available for encoding should thus benefit from decreasednoise correlations. Consistent with this prediction, we found that distant networkcorrelations decrease with training in a stimulus-specific manner, as encoding ofthose stimuli improves. Changes on these larger spatial scales, spanning functionalclusters in the tectum, underlie the overall improvement of network encoding withtraining. Our results show that spatial refinement of noise correlations occurs dur-ing experience-dependent plasticity, and changes to such network-level propertiesare important to the development of tectal function with training.We find that visual training with motion stimuli induces extensive plasticityin the tectum, distinct components of which are NMDAR dependent or indepen-dent. Consistent with previous studies [58, 71, 211], we find no effect of NM-DAR blockade on basal motion response properties in tectum. We found thatNMDAR-independent mechanisms mediate training-induced increases in reliabil-ity and partly mediate improvements in dynamic range, single-neuron mutual in-formation, and neuron-pair mutual information. NMDAR blockade does not com-pletely abolish tectal plasticity [69], and NMDAR-independent plasticity has beendescribed in other systems [163]. However, NMDAR blockade has dramatic ef-fects on coordination of plasticity across the network and components of single-neuron plasticity. When NMDARs are blocked, visual training fails to induce spa-tially structured changes in tectal network architecture, and NMDAR-independentplasticity drives neurons toward common receptive fields over time. This progres-sive loss of network organization prevents training from improving whole-networkperformance. Our findings suggest that NMDARs are essential to coordinated51experience-dependent network plasticity by (1) mediating spatial refinement ofnetwork connections, leading to localized redundancy and distant correlation re-duction, and (2) promoting receptive field diversity and preventing loss of under-represented receptive fields even as local similarity increases.Results from training with a restricted stimulus set suggest that competition be-tween synaptic connections underlies network changes in response properties andnoise correlations. Training with a subset of stimuli dramatically increased the pro-portion of responsive neurons with selectivity towards the four stimuli presented,showing that motion-responsive tectal neurons can alter their preferred directionswith training, and that stimuli compete for representation by a limited pool of tec-tal neurons. Furthermore, decreases in noise correlations over the four trainedstimuli were accompanied by increases over the untrained stimuli, showing thatimprovements in stimulus representation can occur at a cost to opposing receptivefields. Training with four stimuli also reduced noise correlations across all spa-tial distances more dramatically than training with the full eight stimuli, showingthat more specific training elicits stronger network plasticity, and suggesting thatreceptive fields compete for efficient representation by the network.A number of competitive mechanisms mediated by NMDARs could supportthe structured plasticity we observe. These mechanisms include removal of axonalprojections from tectal regions dominated by opposing axons [230], spike timing-dependent plasticity [295], shown to intrinsically mediate competition betweensynaptic inputs [255], and NMDAR-dependent metaplasticity [69] that mediatescompetition by altering plasticity thresholds according to a neuron’s overall inputrate. Our results demonstrate a role for NMDAR-mediated plasticity mechanismssuch as these in experience-driven network refinement.For developing neurons to form functional networks, each neuron must pos-sess learning mechanisms that change its response properties to ultimately improvewhole-network performance. Optimal changes depend on both the specific stimuliencountered and the response patterns of other neurons throughout the network[15, 249, 273]. Our findings show that both of these factors guide NMDAR-dependent plasticity induced by structured visual input in the awake, developingbrain.523.5 Methods3.5.1 Animal rearing conditionsFreely swimming albino X. laevis tadpoles were reared in 0.1 Steinberg’s solu-tion (1 Steinberg’s in mM: 10 HEPES, 58 NaCl, 0.67 KCl, 0.34 Ca(NO3)2, 0.83MgSO4, [pH 7.4]) and housed at room temperature on a 12-h light/dark cycle.Experiments were conducted with stage 50 tadpoles in accordance with the Cana-dian Council on Animal Care guidelines and were approved by the Animal CareCommittee of the University of British Columbia Faculty of Medicine.3.5.2 ImagingOregon Green BAPTA-1 AM (Molecular Probes) was pressure injected into theoptic tectum as described previously [69]. 1 h after injection, tadpoles were placedin a bath containing 4 mM pancuronium dibromide for 7 min, then placed in theimaging chamber and immobilized with agar. The imaging chamber was perfusedwith oxygenated 0.1x Steinberg’s solution during imaging. The region imaged wasdetermined by anatomical landmarks and was roughly 200 µm below the surfaceof the tectum. Images were acquired at 5 Hz using a two-photon laser scanningmicroscope adapted from an Olympus FV300 confocal microscope (Olympus) anda Chameleon XR Ti:Sapphire laser (Coherent) tuned to 910 nm. Images wereacquired using a 60X 1.1NA water objective and encompassed a region of roughly50x150 µm.3.5.3 Visual stimulationStimuli were presented on the center of a 6-mm (1,024 x 768 pixels) LCD screen 7mm from the surface of the left eye. The screen was covered by a longpass filter toblock bleed though of stimulus light into detected fluorescence. Stimuli consistedof solid dark bars with a thickness of 0.09 rad moving at 0.6 rad/s. The edges ofthe stimulus region were obscured by a circular Gaussian mask, so that the eightstimuli were identical except for rotation and had identical intensity profiles overtime. The contrast of stimuli was chosen to be at the threshold of the tadpoles’detection ability, to better compare decoding performance across models over the53course of training.Stimulus presentation and timing were controlled in MATLAB using the Psy-chophysics Toolbox extensions [30]. Stimuli were presented repeatedly with inter-stimulus intervals uniform randomly selected from the set (6, 7, 8, 9) s. Movieswere acquired in 4-min periods, with 1-min periods for microscope alignment be-tween movies, during which stimuli were shown but images were not recorded.The order of presentation of stimuli was randomized such that an equal numberof each stimulus was presented in each 4-min period, and the probability that anystimulus followed any other stimulus was roughly equal over stimulus pairs overthe entire experiment.Tadpoles were presented with one of two stimulus paradigms, 4STIM or 8STIM.Starting 1 h after dye loading, the 4STIM group was presented with a set of fourstimuli corresponding to one half of the stimulus space (0 to 135deg) for 1 h, fol-lowed by 1 h of the full stimulus space. The 8STIM group was presented with thefull stimulus space for 2 h. MK801-treated tadpoles received tectal and ventricularmicroinjections of 20 µm MK-801 after dye loading.3.5.4 Two-photon guided patch recording and calcium imagingFor simultaneous imaging and electrophyisiological recording, loading and imag-ing of calcium indicators were performed as described above. Tadpoles’ headswere mounted in a clear acrylic chamber and held in place by mesh, with tails freeto allow respiration. Patch pipettes (tip resistance 7 MOhm), filled with tadpoleextracellular solution (115 mM NaCl, 4 mM KCl, 3 mM CaCl2, 3 mM MgCl2, 5mM HEPES, 10 mM glucose,10 mM glycine; [pH 7.2], adjusted with NaOH; os-molality 255 mOsm) were inserted through the ventricle, approaching the tectumfrom the medial side. Two-photon imaging was used to guide the pipette tip toresponsive neurons and gentle suction was applied to achieve loose seals (80200MOhm) at which point action potentials could be clearly discerned. We obtainedloose patch recordings at command voltages, which resulted in no net current flowto detect endogenous activity with minimal effect on neuronal firing properties[198]. Imaging and recording were performed while stimulating the contralateraleye with brief flashes from a red LED. Electrical recordings were acquired using an54Axon Instruments Axopatch 200B amplifier, digitized at 10 kHz using a Digidata1322A board, and recorded using pClamp 9 software.3.5.5 Fluorescence data processingFluorescence data stacks were xy aligned using Turboreg (ImageJ, NIH) [271]. Ex-periments that showed vertical drift after alignment were discarded (approximatelyone in four cases). Custom-written software was used to identify and track regionsof interest (ROIs) for each cell over the course of each experiment. Initial ROIswere formed on the basis of morphological characteristics and temporal correlationand excluded cell edges, ensuring no overlapping signal from neighbouring cells.ROIs were then expanded, and these regions were refined and fluorescence sig-nal was denoised using iterated singular value decomposition (SVD), where onlypixels with common weighting indicating a positive correlation with cell calciumconcentration were retained in successive SVD iterations. Pixels in the expandedregion were only retained if they predicted signal in the initial ROI, and if theyshowed less correlation to overlapping ROIs than the maximum correlation of anypixel in the initial ROI. Raw fluorescence for each cell was the reconstructed time-varying mean pixel intensity based on SVD weightings. The fluorescence timeseries for each cell was then calculated as (F−F0)/F0. The time-varying baselinefluorescence, F0(t), was fit for each cell using a Kalman smoother implementingthe Rauch-Tung-Striebel algorithm. The model used for the Kalman smootherconsisted of a signal with no velocity and Gaussian noise of constant amplitude tomodel the slowly drifting baseline. The observation of F0 at each timepoint was theminimum of the smoothed fluorescence trace in a 10 s window around the time-point, and the covariance was the variance of the raw fluorescence trace within thatwindow, to reflect the confidence that the baseline was observed in that window.At this point, cells were excluded from the dataset: (1) If fewer than 80% ofpixels from the original morphological ROI had common weighting in the SVDdecomposition over 80% of the duration of the experiment; this implied that thesingular value did not adequately track the calcium concentration of the cell, whichshould always be positively correlated to fluorescence intensity. (2) If the estimatedsignal-to-noise ratio for the calcium trace in the cell was less than 1.55Spiking parameters for each cell, including the maximum likelihood spiketrain, were fit using nonlinear state space methods [279], with initial parameterestimates for spike amplitude, calcium channel time constant, and saturation de-termined from 10-kHz two-photon imaging line scan data acquired under the sameconditions, and fit to each cell using expectation-maximization. After fitting, spikerate time series for each cell were temporally aligned to each other on the basis of xand y position, to account for the amount of time required to acquire a video frame.Because this model can only place one spike per time bin, it is effective when in-terspike intervals are consistently longer than the bin width used for inferring spiketimings. Over 92% of interspike intervals in electrophysiological recordings duringvisual stimulation were greater than the 50-ms bin width used for spike inference,and less than 0.1% of time bins contained two spikes, with no bins containing three.3.5.6 Single-neuron propertiesTemporal response curves for each stimulus type were generated by averaging neu-rons’ firing rate in the temporal vicinity of each stimulus over all stimulus presen-tations of that type. Each neuron’s evoked response to each stimulus presentationwas the neuron’s mean firing rate between an onset and offset latency after the stim-ulus, which were chosen to maximize the variance of the neuron’s activity acrossstimulus types. Most tadpoles showed potentiation of evoked responses over time(seven of nine 8STIM; seven of nine 8STIM+MK801; three of five 4STIM). Tad-poles showing significant decrease of response amplitude from the first to secondhour of training were not included in analyses.Evoked responses for each neuron were normalized to their mean over each 4-min imaging period to ensure that any changes in overall measured activity wouldnot affect subsequent analyses. Tuning curves were calculated as the mean evokedactivity in response to each stimulus over all imaging periods within an epochof interest. Dynamic range of neuronal tuning curves was defined as the meanabsolute deviation of the normalized tuning curve from its mean, 1. Dynamic rangeis thus the average fraction by which firing rate is altered in response to differentstimuli. To compare trained and untrained stimuli in Figure 3.6, dynamic rangewas calculated in the same way over each set of four stimuli.56Each neuron’s baseline firing rate during each 4-min movie was defined as themedian of its spiking rate binned at 200-ms intervals. The 5-s period following eachstimulus presentation was excluded from baseline estimation. Neuron reliability inresponse to each stimulus type was defined as the fraction of stimulus presentationsto which the neuron responded with a firing rate greater than baseline.Orientation and direction selectivity were measured in the manner of Zhang[294]. The centers of the resulting orientation and direction curves, as plotted inFigure 3.2, were determined by fitting a cos2 or angular Gaussian function, respec-tively. Neurons were considered significantly selective if the amplitude of these fitswas significantly different from 0, with variability in initial measurements takeninto account. The preferred overall directions plotted in Figure 3.8 were deter-mined by the 2D vector sum of neuron-tuning curve values to each direction. Thedirection of the resulting vector was the preferred orientation.Single-neuron mutual information is the mutual information between a singleneuron’s responses and the stimuli, corrected for bias because of limited samplesize [275]. For the calculation of mutual information and decoding, evoked activ-ities were discretized into five bins for each neuron, with each bin containing anequal number of samples.3.5.7 Network propertiesNeuron-pair mutual information is the mutual information between a bivariate neu-ronal activity distribution and the stimuli. p-Values for bivariate mutual informa-tion were estimated by generating random samples with the same number of obser-vations from the independent distribution having the same single-neuron marginalprobabilities.Noise correlation was measured as the correlation between the responses of apair of neurons to a single stimulus type. With the exception of results presentedin Figure 3.3b,c we use mutual information between the two neuron’s responses asour measure of correlation, so as not to limit our investigation to linear correlation.In Figure 3.3b,c linear correlations were used to illustrate that these correlationsare positive and the nature, not merely the degree, of the correlation is stable.Tuning curve similarity was defined as the Pearson correlation between tuning57curve values across stimuli. To better detect shifts in similarity and because sim-ilarity differed across imaging regions, initial similarity was normalized throughmean subtraction in each tadpole.Receptive field diversity is a measure of how well the tuning curves of observedneurons cover the full space of possible receptive fields. We defined this as thevariance, across neurons, of tuning curve amplitudes to a given stimulus, summedover all stimuli.Cooperation in cluster decoding was defined as the decoding performance oftwo groups taken together minus the maximum decoding performance of eithertaken alone. Decoding performance is the negative of the mean classification er-ror of the decoder, in degrees. To better display shifts in decoding success withtraining and under different decoding conditions, decoding performance in Fig-ure 3.5a,c was normalized to initial decoding success of the independent model ineach tadpole.3.5.8 Decoding algorithmsBecause we do not know the methods that downstream neurons use to decode net-work information, we build optimal decoders which calculate stimulus probabil-ities as accurately as possible given an underlying model so as to measure theoverall encoding capability of the network. Decoding of network responses con-sists of assigning a probability P(S | R) that each stimulus (S) was presented, giventhe network response vector R = [r1,r2, ...rn], where each ri is the activity of neuroni. The most common approach to this task is to calculate the inverse distribution,P(R | S), and use Bayes’ rule to obtain the desired result:P(S | R) = P(R|S)∗P(S)P(R) (Bayes’ Rule)Maximum a posteriori (MAP) decoding consists of identifying the peak ofthis distribution, useful for categorical classification. These probability distribu-tions are hard to estimate from biological data because the number of neurons,the dimensionality of R, is high compared to the number of samples available.A simplifying assumption that is often made is to assume that the firing ratesof all neurons are conditionally independent given the stimulus S. In this case,58P(R | S) = ∏i P(ri | S) . This model requires fewer observations to fit because itrequires estimation only of the one-dimensional distribution of ri for each stimu-lus. To perform categorical classification that is sensitive to pairwise interactionsbetween neurons, we used a simple model that relies on the pairwise conditionalprobability distributions P(ri,r j | S), which are more easily estimated than the fulldistribution but can capture more complexity than the independent model:P(R | S) = ∏i, jP(ri,r j|S)Z∏i P(ri|S)N(1)whereZ = ∑k∏i, j P(ri,r j|Sk)∏i P(ri|Sk)NWhere the denominator in (1) is a correction for the overrepresentation ofsingle-neuron probabilities in the product of pairwise tables. The optimal valueof the parameter N depends on the size of the network and its correlation structure.In practice, we selected N a priori on the basis of a linear regression of the opti-mal N against sample size in separate test data. A separate regression for N wasused for cluster decoding presented in Figure 3.8. A prior probability was addedto both models to assure that undersampling would not result in zero probabilitybeing assigned to a stimulus-response pair. Parameter settings, i.e., number of binsand prior probability, were chosen to maximize absolute decoding success underthe independent model, but results were similar under a wide range of parametersettings. Stimulus probabilities generated by both models were adjusted such thatlong-run probabilities of all stimuli given the training data were equal.Assuming sufficiently large samples, this model performs identically to theindependent model when neuronal firing is actually independent. For small devi-ations from independence, consisting of increased probability of a single networkpattern, it categorizes stimuli more accurately than the independent model. This de-coder outperformed the independent model on virtually all real data we collected,and in artificial datasets of size 3 to 150 neurons having small pairwise correlationsand varying sample sizes (unpublished data). Notably, this model does not makeany assumptions about the nature of the bivariate relationships within the network,59unlike parametric models such as copulas [19], and allows for graded activity, un-like the Ising model [272].In all cases, statistics for decoding were calculated from a training set separatefrom the test set to be decoded, using a leave-one-out strategy, in which each shortsegment (eight stimuli) of activity was decoded using statistics calculated on thebasis of all other stimulus presentations in the epoch of interest. For Figure 3.5c,decoders were trained either on the same epoch being decoded, using a leave-one-out strategy, or all stimulus presentations in the opposite epoch.Decoding error was defined as the absolute difference between the MAP esti-mate and the actual stimulus presented, measured in degrees. Decoding improve-ment is change in decoding error, with positive values representing a decrease inerror. Decoding improvement in Figure 3.5 was measured relative to performanceof the independent decoder at the first timepoint.3.5.9 ClusteringClusters (Figure 3.8c,d) were initially formed using the normalized cuts graph clus-tering algorithm [247] over neuron-pair tuning curve similarity. This was followedby gradient descent to generate groups of uniform size (nine neurons) having maxi-mum within-group similarity. Groups that did not reach a threshold value of within-group similarity were not included in decoding. The group size was made uniformto better compare decoding performance across groups. The number of neuronsper group was selected to maximize the difference between the minimum pairwisewithin-group similarity and the maximum pairwise across-group similarity over alldatasets.The median distance between pairs of neurons within these functionally de-fined clusters (Figure 3.17) was used to measure spatial clustering. These medianswere compared to the bootstrapped distribution of randomly generated clusters us-ing the same neuron positions for each tadpole.3.5.10 StatisticsWhere statistics are not described in the methods and figure captions, unpairedt-tests were used to compare mean values across tadpoles.603.6 Supplementary figuresThe following pages contain supplementary figures referenced in the text.61Figure 3.10: Methods for fluorescence data processing. (A) Initial ROIs were identified automatically on the basis of morpholog-ical properties and pixel-to-pixel correlations, with cells automatically tracked from video to video. Cells not highlighted driftedout of the imaging plane in one or more videos over the course of the experiment. (B) Expansion of green bounded region in (A).Morphological ROIs are conservative and do not overlap. (C) ROIs are then expanded for spatial filtering using iterated singularvalue decomposition. (D) Pixel weights indicating the relative contribution of pixels to fluorescence signal reconstruction fortheir respective cells. Brighter pixels indicate higher weighting. (E) Time-varying baseline fluorescence (F0) was fit using aKalman filter smoother taking into account the amplitude of spiking to estimate the accuracy of baseline estimates.62Figure 3.11: Correlation of optical and electrophysiological firing rate measurements. Left, simultaneous recording of somaticfluorescence (∆F/F0, top) and action potentials (green) in response to full field light stimuli of varying intensity, with actual(gray) and inferred (black) firing rates in the 5 s following each stimulus, for three different cells. Right, expanded voltage tracesfor the regions marked in red at left. Pink shading marks time of stimulus. The electrical transients bounding the stimulus periodare clipped. Colored dots mark individual action potentials, which are magnified in the boxes at bottom.63Figure 3.12: Optical measures of firing rate are correlated with electrophysi-ological measurements. Left, scatterplots of number of spikes evoked by vi-sual stimuli versus inferred firing rate (b) measured optically. Evoked spikesrefers to total number of spikes evoked in the 5-s period following stimu-lus onset. Each point represents a single stimulus presentation, and sym-bol colors correspond to distinct neurons. All optical recording parameters(duration, frame rate, optical setup) and fitting method for spike inferencewere identical to experiments performed with optical methods alone. Right,correlation between visually evoked firing rates obtained from cell-attachedrecording and (left) inferred firing rates or (right) peak ∆F/F0. Firing rateinference outperformed peak ∆F/F0 (paired t-test, p = 0.001). n = 5 visuallyresponsive neurons.64Figure 3.13: Noise correlations differ across stimuli. Distribution of mag-nitude of Fisher’s z (normalized to expected SD) for all pairwise compar-isons of noise correlation coefficients in neuron pairs. Dotted line representsthe null distribution (normal with unit variance). Observed noise correla-tions between neuron pairs vary across stimuli 14% more than expected bychance if they were actually equal (p < 1012; Chi-square variance test).65Figure 3.14: Noise correlation encoding. Responses of two example neuronsto two stimulus types. Arrows denote the two stimulus directions plotted.As their single-neuron firing distributions (top and right) indicate, neitherneuron taken alone significantly discriminates the two stimuli. However,because noise correlations differ for the stimuli, the joint firing distribu-tion (center) does discriminate them: when presented with a left movingstimulus (blue), neuron 2 is strongly active only when neuron 1 is inactive(negatively correlated); when presented with a right moving stimulus (red),neuron 2 is strongly active only when neuron 1 is strongly active (positivelycorrelated). As discussed in the text, such encoding is not prominent in thetectum.66Figure 3.15: Receptive field similarity and noise correlation are associated.(a) Scatterplot of signal correlation versus mean linear (Pearson’s) noisecorrelation between tectal neuron pairs. Black points fall outside two SDsof mean of the null distribution. (b) Quantification of (a). Mean signalcorrelation binned for extreme (>two SDs from the mean) and moderatenoise correlations. Values are mean ± standard error of the mean (SEM).67Figure 3.16: The noncompetitive NMDA receptor antagonist MK-801 blocksevoked NMDA receptor currents in Xenopus tectal neurons in vivo. Wholecell patch clamp recordings were performed at a holding potential of +55mV while stimulating axonal inputs at the optic chiasm in the presenceof CNQX (10 µm) to block AMPA receptor currents. Addition of 20 MMK-801 caused a progressive blockade of evoked synaptic NMDA receptormediated currents. Colors denote recording trials before (black), and thefirst, tenth, and 29th stimulation trials after MK-801 application, with a 10-s interstimulus interval. Complete blockade of NMDA receptor-mediatedcurrents were observed in a total of five neurons recorded from five tad-poles.68Figure 3.17: Neuron receptive fields are spatially clustered. Median neuron-neuron distance within groups generated by the clustering algorithm, whichis based only on tuning curves. This is the median distance between pairs ofneurons belonging to the same group, averaged across all groups in a giventadpole. Values are the mean± SEM over n = 7 tadpoles (29 clusters). Dot-ted line is the mean value of this measure across 1,000 randomly selectedclusters in each tadpole using the same neuron positions that were includedin the real clusters. Neurons with similar receptive fields are closer to eachother than expected by chance (two-sample I-test). ∗p < 0.05; ∗∗ p < 0.01.69Figure 3.18: Performance of shuffled decoders does not change with training.Performance of decoders trained and tested on shuffled (blue) or unshuffled(orange) data during early (left) and late (right) epochs in control (top) andMK-801 treated (bottom) tadpoles. To generate shuffled data, responses toeach stimulus type were shuffled for each neuron, a procedure that removesnoise correlations but maintains neuronal tuning curves. Asterisks denotesignificant difference relative to the same decoder in the early epoch (pairedt-test). ∗∗ p < 0.01.70Chapter 4Introduction to Manuscript 2:Brighter Two-PhotonFluorophoresOur studies of population activity reported in the previous chapter left many ques-tions to be answered regarding the computation of motion receptive fields by tectalneurons and the nature of their functional connectivity. More generally, I was eagerto understand the complement of functional inputs received by a single neuron, andhow those inputs or their integration change when the neuron undergoes plasticity.To this end, I started to construct a random access microscope to image synap-tic calcium transients across entire individual tectal neurons. An important studyby Bollmann and Engert [25] had already monitored subcellular dendritic calciumtransients in tectal neurons, by filling single neurons with the calcium-sensitive or-ganic dye Oregon Green BAPTA (OGB). These transients were NMDA-receptor-dependent and showed stimulus tuning, suggesting a synaptic origin, but were mea-sured across entire branches and could not be attributed to individual synapses.To be able to identify individual synapses in tectal neurons loaded with OGB, Ineeded a dye that could be loaded into neurons at the same time as OGB, and labelsynapses belonging to the loaded neuron. The only previously available method,single-cell overexpression of fluorescently labelled synaptic proteins, was unreli-able and rarely produced sufficient contrast to definitively identify synapses, par-71ticularly when a second fluorophore was present in the neuron. Moreover, overex-pression of synaptic proteins could alter neuronal physiology, and would requiretwo separate manipulations- one to introduce DNA encoding the synaptic label intoa single target neuron, and a second to introduce the OGB.To avoid these complications, I developed a fluorescently labelled peptide thatbinds to the synaptic protein PSD-95, and can be introduced into a neuron at thesame time as OGB. One challenge to this strategy was that contrast with intracel-lular labels is difficult to attain, because washing away the label after applicationwould require dialyzing the cell. Therefore, it’s important to introduce the labelat the correct concentration (close to the Kd of the peptide-target interaction) tomaximize the difference in signal between the unbound cytosolic label and thebound target. To improve signal-to-noise, it is very important to have bright singlemolecules of dye bound to each peptide.The need for a bright organic dye for this project prompted me to investigate aclass of dyes that had previously been suggested as very bright but had not yet beensystematically evaluated for in vivo labeling: squaraines. The following chapterpresents the results of a screen of a large number of squaraine derivatives to identifybright, photostable, and nontoxic two-photon fluorophores, and their application tothe production of a synaptic label.72Chapter 5Manuscript 2: Ultra-Bright and-Stable Red and Near-InfraredSquaraine Fluorophores for InVivo Two-Photon Imaging5.1 SummaryFluorescent dyes that are bright, stable, small, and biocompatible are needed forhigh-sensitivity two-photon imaging, but the combination of these traits has beenelusive. We identified a class of squaraine derivatives with large two-photon ac-tion cross-sections (up to 10,000 GM) at near-infrared wavelengths critical for invivo imaging. We demonstrate the biocompatibility and stability of a red-emittingsquaraine-rotaxane (SeTau-647) by imaging dye-filled neurons in vivo over 5 days,and utility for sensitive subcellular imaging by synthesizing a specific peptide-conjugate label for the synaptic protein PSD-95.5.2 IntroductionSince its inception [61], two-photon excitation microscopy (TPM) has been a pow-erful tool in cell and systems biology. In TPM, fluorophores are excited by the73simultaneous absorption of two photons, each with lower energy than that requiredfor single-photon excitation. TPM imaging has high resolution because two-photonabsorption is proportional to the squared incident light intensity, confining excita-tion to a small volume near the focal point. TPM is particularly suited for in vivoimaging [62, 205, 274], since tissues are transparent to the red-shifted excitationwavelengths used, and photodamage is reduced due to the limited excitation vol-ume [259].The need for better, biocompatible fluorophores with large two-photon cross-sections for high sensitivity in vivo imaging has previously been identified [12,142, 226]. The strongest limitations on imaging depth [103] and signal rate [46] inTPM have been photodamage and/or background fluorescence due to high excita-tion powers required to adequately excite fluorophores. Brighter dyes would allowimaging deeper into tissues, at higher rates, with lower excitation power. An idealfluorophore for sensitive in vivo two-photon imaging should show the followingcharacteristics: 1) Strong two-photon absorption in the near-infrared (NIR) win-dow of wavelengths, between 750 and 950 nm, where absorption and scattering bytissues is minimized 2) High fluorescence quantum yield 3) High photostability,to ensure that each molecule can be excited many times before bleaching 4) Highchemical stability within cells 5) No toxicity or other side effects to cells.The brightest organic dyes and proteins commonly used for in vivo two-photonimaging, such as Rhodamine B and eGFP, have peak two-photon action cross-sections of 200 GM in the NIR window. Quantum dots and other nanocrystalsshow stronger absorption, with action cross-sections as large as 47,000 GM, andcan be excited repeatedly without photobleaching [142]. However, quantum dotsare considerably larger than organic dyes, and require appropriate coatings, whichcan be bulky, to reduce cytotoxicity and aggregation [12, 171]. Larger fluorophoresmay be problematic for studies tracking the motion of tagged biomolecules, suchas fluorescence correlation spectroscopy [70] and single-molecule imaging [56],due to potential effects on protein diffusion, localization, and interactions. Therehas consequently been interest in developing smaller probes with brightness andstability similar to existing quantum dots [300].745.3 ResultsSquaraine dyes are a class of red/NIR fluorophores produced by a condensationreaction of electron-rich molecules with squaric acid, a small dibasic acid with aunique square carbon backbone (Figure 5.1a). The donor-acceptor-donor structureof squaraines is conducive to strong two-photon absorption [4], and squaraineshave been identified with two-photon action cross-sections as large as 33,000GM[50], but with molecular weights orders of magnitude smaller than quantum dots.The remarkable properties of squaraines have received attention for applicationsin photoconductivity, solar cells, and nonlinear optics [260, 285]. Squaraine-basedfluorescent sensors have been developed for a variety of analytes including calcium[3], pH [210], protein and DNA, and squaraine-based labels exhibit an increase influorescence intensity and lifetime upon binding to biomolecules [270]. The pho-tostability of squaraine dyes is comparable to those of conventional cyanine dyes[270], but can be substantially increased by the synthesis of a squaraine-rotaxane[7], an interlocked structure wherein a macrocycle encases the electrophilic squaryliumcore, preventing its exposure to nucleophilic attack in solution (Figure 5.1a).Despite their unique optical properties, squaraines and squaraine-rotaxaneshave yet to be applied to biological TPM, and little is known about their pho-tochemical stability, toxicity, or two-photon fluorescence properties in living tis-sue. To address these questions, we screened a library of commercially availablesquaraine derivatives for strong excitation in the NIR window. We identified sev-eral dyes with spectra well suited for in vivo TPM, with action cross-sections inexcess of 10,000GM (Figure 5.1b). One promising candidate is the squaraine-rotaxane SeTau-647 (K9-4142), having a large two-photon cross-section acrossthe NIR window, exceeding 6,000 GM at 740 nm, with a second broad excitationmaximum of 3,000GM at 920 nm (Figure 5.2a). Such a wide absorption spectrumfacilitates multiplex imaging with other fluorophores using a single excitation laser.Many fluorescent dyes show strongest two-photon absorption at twice the wave-length of their one-photon absorption peak. This is predicted for dyes that lacksymmetry for which the absorption induces significant polarization [63], but thispattern can be violated in strongly absorbing dyes [4], and squaraines [50] showtwo-photon absorption strongly blue-shifted relative to twice their one-photon ex-75Figure 5.1: Squaraine derivatives with large two-photon action cross-sectionsin the NIR window. a) Structure of squaraine-rotaxanes. Squaraines con-tain a characteristic squarylium core flanked by nucleophilic motifs, formingan electron Donor-Acceptor-Donor structure. The macrocyclic ’cage’ ster-ically shields the more reactive squarylium core, increasing its stability. b)Two-photon action cross-sections of squaraine derivatives. Dye names witha prefix of K8 denote squaraines, K9 prefixes denote squaraine-rotaxanes.Cross-sections were obtained by ratiometric imaging and calculated usingemission spectra measured from BSA-conjugate (K8–1342), IgG-conjugate(K8–1384), or free dye (all others).citation peak. This shift allows imaging at lower excitation wavelengths, improvingresolution and facilitating co-excitation with spectrally distinct fluorophores. Forexample, the 920 nm excitation peak of SeTau-647 overlaps peaks in two-photonexcitation of eGFP and Alexa Fluor 488, but their emission spectra are well sepa-rated.To be suitable for in vivo imaging, a dye should be stable and unreactive in theintracellular environment, and resistant to photobleaching. We determined the sta-76Figure 5.2: Brightness and photostability of the SeTau-647 measured in vitro and in vivo. a) Fluorescence emission spectrumof SeTau-647 (shaded curve) and two-photon action cross-section, compared to published cross-sections of a bright organic dye(Rhodamine B, red) and fluorescent protein (eGFP, green). For SeTau-647, data in closed circles were obtained by ratiometricfluorescence measurement, and data in open circles were obtained by Z-scan. b) Simultaneous two-photon photobleaching ofSeTau-647 and Alexa 488 at 920 nm in water and fitted mono-exponential decay curves. The abscissa is the product of illuminationtime with the two-photon action cross-section of each dye, proportional to number of photons emitted per dye molecule. c-e)Photobleaching of SeTau-647 and Alexa 594 dextran conjugates in neuronal dendrites in vivo. c) Neurons were electroporatedwith 1 mM Alexa 594- (top) or SeTau-647- (bottom) dextran, and distal dendritic segments were imaged at the minimum laserintensity which allowed clear visualization. The 1st (left) and 60th (right) consecutive images are shown. d)Photobleaching inexample dendritic segments loaded with Alexa 594 (red) or SeTau-647 (blue). Lines are fitted monoexponential decay curves.Intensities are normalized to the maximum and minimum of the fit curve. e) Distribution of fitted photobleaching rate (τ) forneurons loaded with Alexa 594 or SeTau-647. Each data point is the median fit bleaching rate over all dendritic segments imagedin one neuron. Horizontal bars denote the mean of each distribution.77bility of SeTau-647 under two-photon excitation by simultaneously photobleachingSeTau-647 and Alexa Fluor 488 and measuring the time constant of decay in flu-orescence for each dye. Despite showing a much higher cross-section, and thusundergoing a larger number of fluorescence transitions per molecule, SeTau-647exhibits a longer photobleaching time constant than Alexa Fluor 488 under thesame illumination conditions. We calculate that, on average, a SeTau-647 moleculeemits 110 times as many photons as an Alexa Fluor 488 molecule prior to bleaching(Figure 5.2b). Next, to determine photostability in the intracellular environment,we produced a dextran conjugate of SeTau-647 and introduced the dye into sin-gle neurons in the brain of Xenopus laevis tadpoles by single-cell electroporation[93]. To compare the performance of SeTau-647 to a commonly used fluorophoreunder typical experimental conditions, we electroporated individual neurons usingthe same concentration of dextran-conjugated SeTau-647 or Alexa Fluor 594 andperformed in vivo imaging at the minimum laser intensity that allowed clear visu-alization of distal dendritic segments (Figure 5.2c). SeTau-647 labelled neuronswere clearly imaged using laser powers 4 to 8 times lower than those labelled withAlexa Fluor 594 (10 mW vs. 4080 mW at the objective). Under these conditions,bleaching rates for Alexa Fluor 594 were over 10 times faster compared to SeTau-647 (Figure 5.2ce).To determine the long term intracellular stability of SeTau-647, we imageddextran-conjugated SeTau-647 in neurons in vivo over a period of 5 days (Fig-ure 5.3a). All neurons imaged on the day of electroporation were present whenimaged 5 days later (7/7, 100%). Neurons were imaged at the same laser inten-sity on both days and remained extremely bright 5 days after electroporation, withno signs of toxicity. Labelled neurons continued to grow and elaborate dendriticarbors (Figure 5.4), suggesting that SeTau-647 does not interfere with cellular func-tions, and is stable over long periods in the intracellular environment.A major benefit of a small, bright fluorophore is the ability to label discretestructures with low amounts of dye. To demonstrate such an application, we devel-oped a fluorescent probe for PSD-95, a postsynaptic protein used as an indicatorof excitatory synapses. While many excitatory synapses form at dendritic spines,synapses can also form without overt morphological correlates, and many neuronsare non-spiny. A label that acutely identifies synapses would complement existing78Figure 5.3: Cellular and subcellular labelling with SeTau-647 in vivo. a) Z-projection images acquired 3 hours (left) and 5 days(right) after single-cell electroporation with SeTau–647–dextran. Images were acquired at an excitation power of 10 mW at theback aperture of the objective (<8 mW at the focal plane). b) Targeted electroporation of a squaraine-tagged PDZ binding peptidelabels postsynaptic density protein PSD-95. (top) Z-projection of a neuron filled with OGB-1 (green) and the squaraine-taggedpeptide (red). (bottom) Expanded views of dendritic regions numbered above show punctate labelling of SeTau-647 (arrows). c)Anti-PSD-95 puncta colocalize with SeTau puncta in labelled tectal neuron dendrites. (left) Anti-PSD-95 immunostaining (green)of brain section containing dendrites of neurons labelled with SeTau PDZ-binding peptide (red) and Cascade Blue dextran (blue)as a space filler. (right) Expanded views of dendritic regions numbered at left. Scale bars: 10 µm.79techniques relying on fusion protein overexpression, and facilitate imaging stud-ies of synaptic physiology [46]. To acutely label endogenous PSD-95 in individualcells in vivo, we synthesized a peptide with high affinity for PDZ domain 3 of PSD-95 [244], fused to SeTau-647. When labelling endogenous proteins, target sites arelimited and label concentration must be much lower than the peak concentration ofthe target. Optimal signal-to-noise ratios are achieved at a peptide concentrationnear the Kd of the peptide-target interaction. Imaging of small amounts of peptidebound at limited target sites benefits greatly from the higher gain provided by abrighter fluorophore.To introduce labelled peptide into cells, we used two-photon guided single-cell electroporation [125] of tectal neurons, which allows targeting of individualneurons and observation of label diffusion after electroporation. Bright fluores-cent puncta were distinguishable in neuronal dendrites within 5 minutes of elec-troporation, consistent with synaptic labelling (Figure 5.3b,c). Each neuron wasco-loaded with a green dye (Oregon Green BAPTA-1) to distinguish puncta fromvolume effects. Puncta could be clearly imaged at low excitation laser powers(<20 mW). Anti-PSD-95 immunostaining confirms that puncta coincide with en-dogenous PSD-95 (Figure 5.3c). Nearly all (69/74; 93%) SeTau-647 puncta in thesynaptic neuropil of the tectum overlapped with anti-PSD95 puncta, suggestingspecificity of this label to PSD-95containing compartments of the cell. A largeproportion (69/97; 71%) of anti-PSD-95 puncta within labelled cells in the neu-ropil overlapped with SeTau-647 puncta. Some anti-PSD-95 puncta not showingSeTau-647 labelling may belong to other neurons, but overlap with labelled cellsdue to resolution limits. These results suggest that the PDZ-binding peptide actsas a specific and possibly sensitive marker for excitatory synapses in labelled neu-rons. Co-loading of a synaptic marker peptide and a fluorescent calcium indicatoras described here may prove useful for identifying and tracking excitatory synapticactivity [46].5.4 DiscussionThe squaraine dyes offer unprecedented fluorescence properties ideal for in vivotwo-photon imaging. Squaraines have two-photon cross-sections and photosta-80bility approaching those of quantum dots, at molecular weights similar to otherorganic dyes, allowing labelling of subcellular structures, such as synapses, at lowconcentrations and with reduced functional impact. Critically, we demonstrate thatsquaraine derivatives are non-toxic, allowing long-term neuronal imaging. Thedemonstrated brightness and stability of these dyes promise to extend the limits offluorophore concentration, imaging rate, illumination depth, and imaging durationfor in vivo two-photon microscopy.5.5 Materials and methods5.5.1 Fluorescent labelsAll squaraine and squaraine-rotaxane labels including SeTau 647 (K9-4142) arecommercially available from SETA BioMedicals, Urbana IL. Dextran- and peptide-conjugates of SeTau-647 were synthesized from the dye NHS-ester as describedbelow.5.5.2 Animal rearing conditionsFreely-swimming albino Xenopus laevis tadpoles were reared in 0.1x Steinbergssolution (1x Steinbergs in mM: 10 HEPES, 58 NaCl, 0.67 KCl, 0.34Ca(NO3)2,0.83 MgSO4, pH 7.4) and housed at room temperature on a 12 hr light/dark cycle.Experiments were conducted with Stage 48 tadpoles in accordance with the Cana-dian Council on Animal Care guidelines, and were approved by the Animal CareCommittee of the University of British Columbia Faculty of Medicine.5.5.3 Two-photon action cross-sectionsTwo-photon cross-section measurements and in vivo imaging were performed ona custom-built two-photon microscope adapted from an Olympus FV300 confo-cal microscope (Olympus, Center Valley, PA) and a Chameleon XR Ti:Sapphirelaser (Coherent, Santa Clara, CA). Two-photon cross-sections were determined bytwo-channel ratiometric fluorescence intensity measurement with Alexa Fluor 488hydrazide (Invitrogen, A10436) and the open aperture Z-scan method, at wave-lengths between 760 and 980 nm in 20 nm increments. For ratiometric intensity81measurements (Figure 5.1b, Figure 5.2a), the laser was scanned over a 100x100µm region at a depth of 5 µm in a well containing the two dyes at 5 µM con-centration. Z-scan transmission measurements were made in a 1 mm path lengthof 10 µM SeTau 647, using a silicon detector (10D, UDT Sensors, Hawthorne,CA). Fluorescence emission for ratiometric measurements was collected throughthe following optics: 1.1 NA water-immersion objective (Olympus LUMFLN 60xW), 700 nm dichroic (700dcxr), 700 nm shortpass filter (ET700sp-2p), 585 nmlongpass dichroic (ET585/40 m; Ch1 transmitted, Ch2 reflected), 610 nm longpassfilter (E610LPv2, Ch1 only), 500 to 550 nm bandpass filter (HQ525/50M, Ch2only). All filters were purchased from Chroma Technology, VT. We performed si-multaneous current measurements on the two detectors channels while imaging themixture of dyes. We subtracted from these currents the current measured with wa-ter alone at the same excitation wavelength. The subtracted currents were convertedto brightnesses for each channel by compensating for the known parameters of themicroscope/sample system: the emission spectra of the two dyes, the transmissionspectrum of the objective, dichroics, and emission filters, and the gain and sensitiv-ity spectra of the detectors. We performed linear unmixing to convert the channelintensities to brightnesses for each dye, using empirically measured bleedthroughof each dye on our microscope system. Bleedthrough was lower than 1% in allcases. Two-photon action cross-sections were then obtained by multiplying the ra-tio of brightnesses of the two dyes by the known action cross-section of Alexa Fluor488 hydrazide. To confirm accuracy of cross-section measurements, we measuredthe cross-section of Alexa Fluor 594 using the same methods, obtaining valueswithin 15% of those previously reported. Below 800nm, single-photon fluores-cence and scattering become non-negligible in some bulk dyes, interfering with ra-tiometric measurement of cross-sections. At these wavelengths, we used the Z-scanmethod to measure and fit two-photon absorption, and single-photon fluorescencequantum yield to infer action cross-sections (Figure 5.2a), which agreed well withthose measured by the ratiometric method. Previously published cross-sectionswere obtained from the Developmental Resource for Optical Imaging Electronics(DRBIO, Labelled dextran synthesis and single-cell electroporationSeTau-647-dextran was synthesized via coupling of the dye-NHS-ester with equimo-lar amounts of a 10 kDa amino dextran (Invitrogen, D-1860) in pH 8.5 boratebuffer, and subsequent purification using a 6 kDa MWCO desalting column (Bio-Rad, 732 – 6228). For in vivo photobleaching and multi-day imaging experiments,individual neurons within the optic tectum of the intact tadpole brain were filledwith SeTau-647 or Alexa Fluor 594 dextran (D-22913) using single-cell electropo-ration [27]. Tadpoles were briefly anaesthetized with 0.02% 3-aminobenzoic acidethyl ester (MS222, Sigma). A sharp glass pipette (0.6 µm tip diameter) filled withdye dissolved in ultrapure water was inserted into the cell body layer of the tectum,and electroporation was performed using an Axoporator 800A (Molecular Devices,Sunnyvale, CA; stimulus parameters: pulse intensity = 1 µA; pulse duration = 300µsec; pulse frequency = 900 Hz; train duration 14 msec).5.5.5 In vivo imaging and photobleachingIn vitro photobleaching was performed at a depth of 5 µm in a well (2 mm di-ameter) containing 0.5 mM SeTau-647 and Alexa Fluor 488 hydrazide, using thesame detection setup described above. Intensity traces were obtained while scan-ning the excitation laser across a square region (˜100x100 µm) at 500 ms/frame.Resulting traces were fit with an exponential decay curve and normalized to themaximum intensity and minimum of the fit. For in vivo photobleaching experi-ments (Figure 5.2ce) tadpoles were reversibly paralyzed with 4 mM pancuroniumdibromide prior to imaging, and placed in an imaging chamber perfused with oxy-genated Steinbergs solution throughout imaging. Neurons were electroporated us-ing 1 mM-equivalent concentrations of Alexa 594- or SeTau-647-dextran. Thesewere concentrations of dextrans with absorption at 600 nm equal to that of 1 mMfree dye. Alexa 594 and SeTau-647 were imaged at 800 nm and 920 nm, respec-tively. Distal regions of loaded neuronal dendrites were imaged at the minimumlaser intensity, which allowed dendrite boundaries to be clearly visible in a singleframe (256x256 pixels 34x34 µm, 500 ms/frame). Both dyes were imaged with thefilter set for Ch1 described above. This filter set, combined with the quantum effi-ciency of the detector (Hamamatsu R4632), results in detection of 2% and 0.8% of83Alexa 594 and SeTau-647 emission, respectively. For 5-day imaging experiments(Figure 5.3a, Figure 5.4), neurons were imaged on the day of electroporation (Day0) and 5 days later (Day 5) using the same laser intensity (<8 mW at the sam-ple). Image stacks were registered by automated rigid-body transformation andmaximum intensity projections are shown. Neurons were drawn in 3D using ourcustom written software Dynamo for analysis of dendritic arbor morphology andgrowth (Figure 5.4).5.5.6 Targeted electroporation and PSD-95 labellingTo label endogenous PSD-95, we used the peptide sequence RVRLQTSV, whichhas high affinity for PDZ domain 3 of PSD-95, as identified in a phage displayscreen [244]. This sequence was modified by addition of a triglycine linker and anarginine residue (RGGGRVRLQTSV), and N-terminal coupled with SeTau-647.Coupling was performed during synthesis before resin cleavage to improve separa-tion of product from the similar-sized reagents. The labelled peptide was purifiedby high performance liquid chromatography on a C8 column after cleavage.For two-photon guided single-cell electroporation, the tectum was loaded withCascade Blue Dextran (Invitrogen, D1976) to visualize the tectum and identify ma-ture cells for electroporation. This dye shows virtually no two-photon absorptionabove 900 nm, and does not interfere with subsequent cellular imaging. Targetedelectroporation was performed at 830 nm excitation. The electroporation pipette,containing 1.8 mM Oregon Green Bapta-1 (Invitrogen, O6806) and 0.8 mM SeTau-labeled peptide was inserted into the tectum and placed against a single mature tec-tal cell. Each cell was electroporated with a single 10 ms square pulse of between1825V. Morphological imaging was performed at 920 nm excitation.5.5.7 ImmunohistochemistrySeveral neurons per brain were co-loaded with Cascade Blue Dextran and theSeTau-647 peptide label using single-cell electroporation. 30 minutes after load-ing, brains were extracted, fixed in BT fixative (4% paraformaldehyde, 4% sucrose0.12 mM CaCl in 0.1 M PBS, pH 7.4), washed. Brains were mounted in OCTmedium and cryostat sectioned at 10 um. PSD-95 was detected with a monoclonal84primary antibody (clone 7E3-1B8, Millipore) and Alexa-488-coupled secondary(Invitrogen, A11001). Sections were imaged with a confocal microscope (Olym-pus FV1000). SeTau-647 was excited at 633 nm and the emission was detectedat 650750 nm. Alexa Fluor 488 was excited and imaged separately from CascadeBlue and SeTau647 to ensure no bleedthrough between channels. Puncta werecounted from seven imaged sections containing labelled cell dendrites across 3tadpoles. Alexa-488 puncta were considered to be within labelled cells if the entirepunctum overlapped with dendritic staining.85Figure 5.4: SeTau-647 dextran labelled neurons continue to grow and elaborate branches. a) Maximum intensity projection imagesof a neuron loaded with SeTau 647, on the day of electroporation (left) and 5 days later (right). b) Tracings of 3D image stacksin a. Greyscale intensity indicates Z-position of traced processes. Black circles indicate neuron somata. c) Total dendritic branchlength (TDBL), including filopodia, of 5 neurons traced as in b, on the day of electroporation with SeTau 647, and 5 days later.Scale bars: 20 µm.86Chapter 6Summary of Manuscript 3:Non-negative Deconvolution6.1 IntroductionCalcium transients in neurons, such as those induced by action potentials, have acharacteristic shape consisting of a rapid rise followed by a slow exponential decay.This slow decay can be either beneficial or problematic, depending on the situation.Longer decay times make it easier to detect signals when imaging slowly. On theother hand, if action potentials occur much faster than the decay time, they canblend together to form a plateau, and their timing can be difficult to distinguish.A common method to recover the onset times of transients has been to deconvolvethe known profile of a single action potential from the recorded timecourse. Thistype of unconstrained linear deconvolution is efficiently implemented with highlyoptimized algorithms, such as the discrete Fourier transform. If all transients arethe same, and sum linearly, and there is no noise, this method produces the onsettimes of action potentials.In reality, however, there is often a great deal of noise, and not all transientsare the same. Noise in the measured trace can produce large positive deflectionsfollowed by equally large negative ones in the deconvolved signal. One obviousconstraint to apply, then, is to require that the deconvolved signal be non-negative,since the count of action potentials (the underlying signal being recovered) is non-87negative.Intriguingly, a perfectly analogous problem arises when processing data fromphotomultiplier tubes (PMTs), the most commonly used detector in two-photon mi-croscopes. Photons striking the PMT cathode generate a current. When terminatedinto even a relatively low resistance (e.g., 50 Ohm), the very high capacitance ofthe PMT stretches these currents into the same fast rise and slow exponential de-cay observed with calcium transients. These transients limit the speed at whichphotons can be counted, and require PMTs to be terminated into very low resis-tance to avoid bleedthrough between adjacent pixels, which in turn increases therate at which analog sampling of PMT currents must be conducted and reduces thesignal-to-noise ratio of the sampled signals.In the course of building our random access microscope, the similarity of thesetwo problems became apparent to me. Previous methods for non-negative decon-volution had been fast enough for real-time processing of most population calciumimaging traces, which tended to be recorded at no more that 30Hz. However, Iplanned to record from multiple PMTs at 5MHz, hundreds of thousands of timesfaster. To be able to apply non-negative deconvolution to improve PMT signalprocessing, I sought to develop a much faster algorithm to implement the deconvo-lution.The resulting algorithm and its applications for both PMT signal processingand deconvolution of calcium transients were published in the Journal of Biopho-tonics [204], but are not reproduced here for the sake of brevity. The Abstract andIntroduction of that paper are included below as a summary.6.2 SummaryLaser scanning microscopy (LSM) is a common technique for high resolution fluo-rescent imaging. Here we describe a fast algorithm for non-negative deconvolutionand apply it to readout of LSM detector photocurrents. By broadening photonimpulses and deconvolving sampled photocurrent, effective quantum efficiency ofthe imaging system is increased. Using simulation and imaging with a custom-builttwo-photon microscope, we demonstrate improved fidelity of images acquired atshort dwell times over a wide range of photon rates. Images formed show increased88correlation-to-sample equivalent to a 25% increase in photon rate, lower noise, andreduced bleed-through compared to conventional image generation.Fluorescence microscopy has become a leading tool for investigating cellularand systems biology, allowing imaging of labeled structures and molecules withsubcellular resolution. Genetically encoded fluorescent proteins [38] have revolu-tionized our understanding of fundamental biological processes on scales rangingfrom protein interactions [139], to cellular structure [43], to entire tissues [109].Quantitative imaging of fluorescent sensors has allowed continuous, spatially re-solved monitoring of cellular signaling, complementing or replacing more invasivetechniques and enabling studies in awake animals [205]. Laser scanning confocaland multi-photon microscopy (LSM) allow high resolution fluorescence imagingby collecting light from a single focal plane, and have become dominant techniquesin this field [197]. In LSM, a focused laser beam is scanned in two or three dimen-sions through a sample, and fluorescence emitted by the sample is collected andtypically focused onto a single photodetector, which reads out photons for eachpixel in sequence. Because pixels are acquired in sequence, little time is availableto sample each pixel, particularly when the imaging area is large, the sample rateis high, or the image must be of high fidelity. Such conditions are common in bio-logical experiments, and most applications thus require that photons be detected bythe imaging system at a high rate, so that pixel brightness can be accurately mea-sured in a short amount of time. This need for high photon rates has motivated thedevelopment of bright fluorescent dyes [173] and highly efficient photo-detectorsused in current imaging applications.Efficient signal processing and image generation are also important to LSM,and the manner in which photodetector signals are conditioned and processed candramatically impact the sensitivity and dynamic range of the imaging system. Sig-nal processing is particularly important in quantitative imaging, where pixel bright-ness must be measured with high precision, and systematic errors can bias ex-perimental results. Here we introduce a signal processing method, non-negativedeconvolution (NND), which improves LSM performance over a wide dynamicrange. We compare NND to conventional deconvolution and to two existing meth-ods, photon counting and photocurrent binning. As LSM equipment has becomereadily available, many investigators have built custom microscopes designed to89address specific research questions or as lower-cost alternatives to commercial sys-tems [91, 161, 208]. The flexibility of such custom systems allows researchers torapidly implement new signal processing methods such as the one described here.With such researchers in mind, an optimized form of the NND algorithm is avail-able upon request.90Chapter 7Introduction to Manuscript 4:Comprehensive DendriticImaging7.1 Relating neuronal structure and functionThe exquisite morphological diversity of neurons having distinct information-processingroles hints that the structure of neurons is important for understanding their con-tribution to circuit function, and vice versa. A relationship between neuronal mor-phology and function is also evident from correlations between structural abnor-malities and circuit dysfunction in brain disorders [138]. For example, Rett Syn-drome [17], Autism Spectrum Disorders [27], Schizophrenia [172], and Epilepsy[37] are associated with aberrant neuronal morphologies, which likely result fromerrors in morphological growth during the critical period of early brain develop-ment. How structural differences arise and are maintained, and how they impactsubsequent activity, plasticity, and circuit specification, are thus important ques-tions in understanding these disorders. While the shape and size of neurons clearlyimpacts possible function, it is also accepted that activity during early critical peri-ods contributes to directing morphological growth, since altering activity dramati-cally affects growing neuron morphology [94, 287]91A great advantage of the Xenopus model system is the ability to induce ro-bust plasticity through controlled visual stimulation protocols delivered duringtwo-photon imaging, allowing monitoring of both structural growth and functionalplasticity as they occur in real time. Multiple lines of evidence indicate a temporalcorrelation between functional and structural plasticity in the Xenopus tectum. Forexample, training paradigms that tend to strengthen functional responses at Stage50 were shown to also stabilize tectal dendritic filopodia in younger neurons [44].To date, however, studies have limited themselves to monitoring changes in eitherstructure or function, but not both in the same neurons. It is particularly importantto measure structure and function simultaneously because of the heterogeneity oftraining-induced functional changes in neurons. Training can induce potentiation,depression, or no change in neurons in the same brain, depending on their activ-ity patterns [69]. To associate structural and functional changes, both need to bemeasured at the same time.7.2 Summary of resultsThe following chapter describes a series of studies that can roughly be divided intotwo categories. The first experiments aimed to establish the relationship betweenneuronal firing patterns and growth behaviour. These involved monitoring somaticfiring throughout plasticity-inducing training, while also measuring complete pat-terns of growth in the same neuron. Because activity patterns vary across neurons,this required full reconstruction of large numbers of neurons – 45 neurons across atleast 5 timepoints each for just for this first series of experiments – to make reason-able measurements of growth for each activity pattern and manipulation. I limitedthis study to a single type of neuron in a specific region of the tectum, to furtherreduce variability. The main result of these experiments is that neurons with differ-ent firing profiles throughout training show very different patterns of growth, butneurons with the same firing profile are quite similar. Moreover, we manipulatedcytoskeletal dynamics, synaptic transmission, and neuronal firing, and found thatactivity drives structural changes and not vice versa, at least over the time scalesmeasured here. Interestingly, important details of the relationship between somaticactivity and structural changes described in our results, while not shown defini-92tively by simultaneous structural and functional imaging, were recently predictedby Kishore and Fetcho [129]The second set of experiments represent the first application of our randomaccess microscope to synaptic imaging, and the first case of activity being simulta-neously monitored across entire 3D neurons in vivo. Here we sought to understandthe dendritic activity patterns that occur in neurons during sensory experience, andwhether local learning rules can explain the changes we saw in the first series ofexperiments. We identified a simple set of local learning rules, based on the rela-tive amplitude of sensory-evoked calcium transients at the base and tip of filopodia,that determine filopodial stability or motility in vivo and could fully account for ourprevious observations. These results vividly show that fine-scale activity directsthe growth of these neurons, and that filopodial motions that previously appearedto be random are in fact, when animals are awake, being precisely guided by theinformation being transmitted by the developing neural network.7.3 Dynamo software for Dynamic MorphometricsDynamic Morphometrics refers to the rapid 4D (3 spatial dimensions over time)quantification of complete neuronal growth patterns. To accurately reconstruct andanalyze neuronal structures over time, I created an Open Source toolbox for Dy-namic Morphometrics called Dynamo. This software records the full 4D structureof the neuron, and facilitates tracing of series of image stacks by automating partsof the drawing process. It provides rudimentary automatic tracing (AutoDraw),but more importantly, automatically registers drawings to successive image stacksto compensate for movement and growth. Once the first timepoint is traced, lit-tle human input is required to produce tracing of all subsequent timepoints. Thisfunctionality solves an important problem in random access microscopy, namelythe identification and tracking of random access imaging sites. Dynamo thereforeserves two purposes: to analyze morphological changes in structural imaging ex-periments after acquisition, and to track the structure of moving neurons onlineduring imaging. Dynamo automates analysis of neuronal structure and motilityby tracking additions, subtractions, and changes in length of dendritic filopodia,transitions of filopodia into stable branches, and how all of these are arranged in93space. It also provides functions for Monte Carlo simulation to detect clusteringof processes, and being open source, can easily be extended to include additionalanalyses.7.4 Random access microscope designWe used a relatively simple design for our random access microscope, using asingle pair of large-aperture TeO2 crystals operating in transverse mode to deflectthe excitation laser beam in two dimensions, as has been previously implemented[121]. We compensated for spatial dispersion with a single custom-cut prism thataccurately cancels dispersion in the center of the deflectors’ field of view. Thisdesign is simple but does not fully compensate for spatial dispersion at the edgesof the field of view, which can be a limitation for scanning large fields of view butdoes not strongly limit imaging performance over the fields we image, ˜100x100microns. We also employed a simple approach to compensate for temporal disper-sion. I initially constructed a prism-based adjustable prechirper, which was usedin initial experiments described in Manuscript 3. We ultimately purchased a laserwith integrated dispersion compensation based on chirped mirrors (Chameleon Vi-sion II, Coherent), which removed the need for the earlier prechirper. Both of thesemethods compensate only for group velocity dispersion, and not higher-order tem-poral dispersion, though this is sufficient to produce strong two-photon excitationat the sample with the laser power available. We also use a relatively simple ap-proach to axial scanning, by mounting the objective on a piezo focusing stage. Thismaintains excellent beam quality along the entire imaging volume, but is slightlyslower than remote focusing methods [26], and is limited to a roughly 100 Hzvolume acquisition rate.More sophisticated designs have been proposed that more precisely compen-sate for both spatial and temporal dispersion, and other aberrations, introduced byacousto-optic deflectors, and enable random access scanning in all three dimen-sions [227]. While some of these approaches may one day be incorporated intoour microscope, they were not necessary to achieve the resolutions and fields ofview required for imaging tectal neurons. Notably, having a single conventionallyscanned axis is not a major limitation in our application, because imaging sites94fill a contiguous region of the Z-axis. To increase the amount of time availablein imaging planes containing more target sites, we can alter the trajectory of theobjective, but every plane within the range of the neuron must be visited to com-prehensively track activity. Several refinements in the design of our microscopeare worth noting here. I used a photodiode to measure laser power in real time,allowing compensation for fluctuations in excitation power, which includes noisein the laser intensity as well as fluctuations in deflector diffraction efficiency. Thisphotodiode also allows us to calibrate excitation across the field of view to produceuniform illumination. The microscope features an extremely short collection lightpath, of 11 cm from the back aperture of the objective to the detector cathode, facil-itating collection of scattered light. One major advance that made synaptic imagingfeasible in vivo was the ability to prevent virtually all sample movement. Awaketadpoles can be paralyzed with curare, which blocks the cholinergic neuromuscularjunction, and has been used extensively for in vivo imaging and electrophysiologystudies. To further stabilize tadpoles, I designed head clamps that apply pressureto the head from all sides, keeping it fixed in place, while leaving the tail free tomove for oxygen exchange, or for behavioural assessment of swimming motionsin unparalyzed tadpoles. We fabricated these head clamps from laser-cut acrylic,with molded silicone-based inserts to conform to the head sizes of various stagesof tadpoles. The other major advance facilitating these imaging experiments hasbeen software development. The ability to tailor the operation of the microscopeto the specifics of each experiment is invaluable, because numerous details of theexperiment affect relatively low-level operation of the microscope. The integra-tion of Dynamo into the imaging software is an excellent example of this, alreadydiscussed. As another example, the appropriate excitation laser intensity differsacross sites, with the neuronal soma being much brighter than distal dendrites un-der resting conditions. Notably, however, the appropriate excitation intensity alsovaries across trials; dendrites become much brighter when probed with stimuli towhich they respond. To ensure adequate excitation without saturating the detec-tors (and losing part of the recording), it is important to predict the brightness ofeach sample point for each trial, on-line, and adjust excitation accordingly. Suchadjustments, and others like it, would be difficult or impossible with a commercialmicroscope.95Chapter 8Manuscript 4: Sensory ActivityInstructs PatternedDendritogenesis in the AwakeBrain8.1 SummarySensory activity is necessary for normal development of many neural circuits, but itremains unknown to what extent experiences shape detailed neuronal structure andconnectivity in the developing brain. Here, we designed an ultrafast random-accesstwo-photon microscope to simultaneously record activity at closely-spaced pointsthroughout the entire 3D dendritic arbor of developing neurons in awake animals.We characterized complete spatiotemporal patterns of synaptic input, firing out-put, and structural changes throughout plasticity-inducing visual training, findingthat sensory experience induces spatial clustering of dendritic filopodia respondingto the same stimuli. Experience-driven strengthening of sensory-evoked neuronalfiring causes filopodial pruning, while non-responding neurons extend spatially-clustered filopodia. We identified several activity-dependent rules governing whenand where filopodia are added, removed, or stabilized, depending on the ampli-96tude of stimulus-evoked calcium transients at the base and tip of each filopodium.These rules shape the spatial organization of dendrites and their synapses accordingto specific experiences.8.2 IntroductionDuring learning, activity passing through neural networks changes those networksto improve future processing [80, 169, 199, 205, 223]. These changes involvean intricate interplay between neuronal structure and function that is most dra-matically realized during early brain development, when neurons first assemblethemselves into functional circuits. Although sensory experience is essential forlearning, it remains unclear how specific experiences guide the development of op-timized information-processing circuits. Here, we investigate how sensory-evokedactivity shapes the detailed structure and function of neurons in the developingbrain.Activity is necessary for many aspects of functional [118, 131, 182] and fine-scale structural [5, 80, 157, 158, 169, 250] brain development. A central questionregarding activity-dependent development has been whether activity is instructiveor permissive [55]. Where it is instructive, patterns of activity determine circuitstructure, allowing the brain to adapt to the specific environment it encounters.Where it is permissive, the presence of activity is required, but its detailed pat-tern is not important. For a cue to be instructive, different patterns in the cuemust correspond to the distinct outcomes studied. Patterns of firing across neuronscan, for example, instruct which neurons integrate into an emerging response cir-cuit [291]. However, determining whether activity instructs fine-scale patterns ofdendrite growth requires monitoring activity at the corresponding spatial scale ofindividual dendritic processes.To address this question, we developed methods to track an individual neuron’sdendrite growth, somatic firing and complete 3D dendritic arbor activity simulta-neously in real time. We term this approach ’comprehensive imaging’. Imaging ofcalcium-sensitive reporters is widely used to measure somatic firing and excitatorysynaptic input in dendrites [45, 124, 205]. A major obstacle to large-scale in vivosynaptic imaging has been the need to rapidly image 3D volumes at high resolution.97To simultaneously track activity of all dendritic processes, shafts, and the soma ofa labelled neuron, we employed random access multi-photon (RAMP) microscopy[121, 218]. RAMP microscopy enables imaging of arbitrary sets of points withoutpassing through the intervening space, using acousto-optic deflectors to steer thelaser focus. We performed comprehensive imaging in the awake brain of devel-oping Xenopus tadpoles, while conducting plasticity-inducing sensory training, toinvestigate how experience can shape detailed neuronal structure and function.The visual system of the Xenopus laevis tadpole provides in vivo access toa conserved vertebrate brain circuit undergoing rapid development [25, 69, 151,183, 184, 205, 250]. At 2-3 weeks of age (Stages 49-50) the retinotectal sys-tem processes visual input, but continues to undergo rapid maturation of receptivefield properties [84, 205]. Transparent albino tadpoles allow imaging through theskin without surgery or anaesthesia, permitting observation of ongoing activity-dependent growth and information processing in the awake brain. Developing den-drites exhibit numerous filopodia that act as sites of synaptic contact [151, 189] andundergo rapid activity-dependent structural changes [25, 183, 250]. At these earlystages of development, backpropagating action potentials produce only modestdendritic calcium signals [25, 269], allowing recording of localized dendritic cal-cium transients without voltage-clamp, which can interfere with plasticity events[183]. Visual training induces robust structural and functional plasticity in tectalneurons within 30 minutes [44, 69], providing a unique opportunity to observe thecomplete process of experience-induced neural plasticity as it occurs in real time.8.3 Results8.3.1 Dual in vivo imaging of sensory-driven growth and somaticfiringRecent studies have associated sensory-evoked neural firing, plasticity, and learn-ing with dendritic structural changes during development [44, 223] and in the ma-ture brain [80, 128, 199]. To directly correlate these measures within the samecell, we imaged both somatic activity and dendrite growth in individual neuronswhile training the brain to better discriminate a sensory stimulus. We used targeted98single-cell electroporation [93, 130] (TSCE) to selectively label neurons accordingto their evoked firing properties (Figure 8.1a,b, Figure 8.8). We filled the tectumwith the cell-permeable calcium-sensitive dye Oregon Green BAPTA 1-AM. Weselected a target neuron using two-photon imaging of visually-evoked somatic cal-cium transients, which reflect bursts of action potentials in tectal neurons [205],and approached it with a pipette containing Alexa Fluor 594-dextran, which wasdriven into the cell with a brief train of voltage pulses (Fig. Figure 8.1b). Welabelled only type 13b pyramidal neurons [147] in a small target region of rostraltectum responding to a 50 ms full-field OFF stimulus presented to the contralateraleye via a red LED [69]. The sensory-evoked responses of these cells and relativelystable dendritic morphology reflect neurons in a late stage of development, wellintegrated into visual circuits.We induced long-lasting plasticity using a Spaced Training (ST) protocol thatprimarily drives long-term potentiation (LTP) of somatic responses [69]. Visualstimulation consisted of 30 minutes of baseline probing to assess OFF-evokedresponses, followed by 30 minutes of ST, composed of trains of high frequencystimuli spaced by periods of invariant light, followed by two 30-minute probingsessions to assess plasticity. In initial experiments, we performed 2D imagingof evoked somatic calcium transients during probing, and collected full 3D vol-ume images of neuronal dendrites between probing sessions to monitor structuralchanges (Figure 8.1e). We fully reconstructed each neuron’s entire dendritic arborat every 3D imaging time point using our custom-written Dynamo software, allow-ing comprehensive morphological analysis of all dendritic processes across time, amethod we term Dynamic Morphometrics (Figure 8.1c,d).8.3.2 Experience-induced dendrite growth patterns correlate withevoked somatic activity and plasticityWe used TSCE to label individual neurons exhibiting OFF-evoked somatic re-sponses and neurons not showing OFF-evoked responses adjacent to respondingneurons. We grouped all neurons into four ’activity profiles’ according to theirinitial somatic response to the stimulus and the plasticity of that response with ST(Figure 8.1e). Neurons were classified as having an initial response that increasedin amplitude with training (LTP, 7 neurons), an initial response that did not change99Figure 8.1: Targeted Single-Cell Electroporation and Dynamic Morphometrics of neurons with identified visual responses. a) An awaketadpole is shown visual stimuli during calcium imaging of the tectum to identify responding neurons for targeted electroporation. b) Neuronsloaded with OGB immediately before (left) and after (right) electroporation of a single responsive neuron with space-filling Alexa 594 dextrandye. c) Z-projection (left) of a neuron image stack and corresponding 3D reconstruction (right) produced by Dynamo software. d) (top)Overlay of two substacks of the same neuron, acquired 30 minutes apart. The overlay is white where structures are stable, cyan where theyextend, and magenta where they retract. Arrowheads mark sites of addition (green), subtraction (red), growth (cyan) and retraction (magenta).(bottom) Corresponding motility plot generated by Dynamo. Circles mark sites of motility, and their area represents the change in length ofthe filopodium. Dynamo allows clear 3D visualization of motility not easily detected in 2D overlays. e) Dynamo motility plots and meanevoked responses (∆F/F0, inset) for neurons of the four most common activity profiles.100(No Change, 5 neurons), no initial response but a response after training (BecameResponsive, 5 neurons), or no response before or after training (No Response, 4neurons). The four groups had consistent dendritic arbor sizes, indicating similarmaturational state (Figure 8.9). These different activity profiles likely reflect dis-tinct upstream inputs and receptive fields, as neurons not responsive to OFF stimulioften responded to other visual stimuli upon testing (30/73; 41%).Neuronal growth patterns and activity profiles were strongly associated (Fig-ure 8.1e,Figure 8.2a). Activity profiles explain most of the variation in growthpatterns of these neurons (64%, MANOVA F statistic, p < 0.0003). LTP neu-rons showed decreased dendritic filopodial addition rates during training (58% ofbaseline), remaining low over the next hour, and increased subtraction rates duringtraining (176% of baseline), returning to baseline over the next hour. No Changeneurons showed no significant fluctuations in additions, subtractions, or motility,but maintained constant levels of dynamic growth. Became Responsive neuronsshowed a strong increase in filopodial additions during training (220% of baseline)that remained high over the following hour. Subtraction rates increased strongly(243% of baseline), but only after these neurons became responsive. Interestingly,No Response neurons showed strong modulation of growth patterns in responseto ST. In these neurons, addition rates increased during training (171% of base-line) and remained high over the following hour. Subtraction rates greatly de-creased (40% of baseline), returning to baseline over the next hour. Throughoutall groups, extensions (elongation of existing processes) and retractions (decreasedlength without process loss) showed modulation patterns similar to additions andsubtractions, respectively (Figure 8.10). These results show that each neuronalactivity profile corresponds to a distinct pattern of dendritic growth and pruning.Across groups, training-induced potentiation of somatic responses was followedby filopodial pruning, while absence of somatic responses was accompanied bygrowth during training.8.3.3 Experience-induced functional plasticity is upstream ofstructural plasticityThese strong correlations between growth patterns and functional plasticity couldarise in three ways. Structural changes could cause functional plasticity, e.g. by en-101Figure 8.2: Neuronal activity determines experience-induced structural plasticity a) Mean filopodial addition (top) and subtraction (bottom)rates during each Epoch, for Control neurons, neurons following APV injection, neurons expressing inhibitory DREADD hM4Di followingCNO injection, and neurons before (Baseline) and after actin blocking cocktail (ABC) injection. Rates are expressed as percent of totalfilopodia at Epoch 1, minus mean rate during Epoch 1. Control: N=7 (LTP); 5 (No Change); 5 (Became Responsive); 4 (No Response). APV:N=4 (Responding); 4 (No Response). hM4Di: N=4 (Responding); 4 (No Response). ABC: N=5. ∗ : p < 0.05 ∗∗ : p < 0.01, relative to sametreatment at first Epoch. ANOVA followed by Tukey LSD. b) (left) Mean responses in a neuron expressing GCaMP6m alone (Control, top) orGCaMP6m+hM4Di before and after CNO administration. Shading denotes SEM (n=16 trials per epoch). (right) Average response amplitudesbefore and after CNO administration in initially-responding control and hM4Di neurons. (Control, N=5; hM4Di, N=4 neurons). ∗∗ : p < 0.01,paired t-test. c) Proportion of neurons showing each possible activity profile in tadpoles receiving Control or ABC tectal injections. Actinblockade does not reduce the proportion of neurons undergoing plasticity with training. Control: N=5 tadpoles, 423 neurons. ABC: N=5tadpoles, 399 neurons. ns: not significant, χ2 test. d) (top) Overlays at 30 minute intervals of a dendritic branch. After 30 min, the tectumwas infused with ABC, followed by visual training. Overlays are white where dendrites are stable, cyan where they extend, and magentawhere they retract. Arrowheads mark sites of extension (cyan) and retraction (magenta). b) Dynamo motility plot and mean evoked responses(∆F/F0, inset) for the same neuron. Circles represent filopodial additions (green), subtractions (red), extension (cyan) or retraction (magenta).The area of each circle represents change in length of the filopodium. This neuron underwent LTP in response to training.102abling new connections, removing old ones, or altering biophysical properties. Al-ternatively, functional plasticity could cause structural changes, e.g. by triggeringcytoskeletal remodeling [193, 250]. Finally, the two effects could be independent,but share upstream regulators. To test whether cytoskeletal remodelling is neces-sary for long-lasting functional plasticity in these neurons, we used a mixture ofJasplakinolide (4mM) and Latrunculin A (5mM), inhibitors of actin polymerizationand depolymerization, respectively, to inhibit actin dynamics throughout stimula-tion [81, 192, 220]. Tectal injection of this actin blocking cocktail nearly abolishedfilopodial additions, subtractions, and motility for over 2 hours (Figure 8.2a,d). De-spite this dramatic morphological stabilization, and in contrast to studies of moremature spine plasticity [81, 192, 220], we found no effect of cytoskeletal freezingon plasticity of somatic firing responses with ST (Figure 8.2c).To determine whether functional plasticity is needed for training-induced struc-tural changes, we used D-APV (50 µM), an NMDA receptor (NMDAR) blocker.NMDARs conduct localized synaptic calcium currents, affecting dendritic growth[165, 187, 250] and plasticity [160]. Tectal APV injection reduced the propor-tion of Became Responsive and LTP neurons compared to vehicle-injected controls(Figure 8.11). Neurons treated with APV showed no change in motility throughoutthe stimulation protocol, regardless of somatic responsiveness (Figure 8.2a), indi-cating that NMDARs are important for experience-induced structural plasticity.NMDAR blockade interferes with firing plasticity and synaptic signalling, butdoes not significantly alter neuronal firing responses. To investigate effects of de-polarization and firing on structural plasticity, we expressed hM4Di, a designerreceptor exclusively activated by a designer drug [6] (DREADD). Tectal injectionof the exogenous ligand CNO (5 µM) largely silenced evoked firing of hM4Di-expressing neurons (Figure 8.2b). We grouped silenced neurons according to theirOFF responses prior to CNO administration as either Responding or Not Respond-ing. These groups showed similar levels of additions and subtractions, which werenot modulated by training (Figure 8.2a).These results show that NMDAR activation and neuronal firing are necessaryfor training-induced modulation of dendritic structure, but structural changes arenot necessary for functional plasticity. We conclude that neuronal activity actsupstream of experience-induced structural changes.1038.3.4 Hotspots of structural plasticity are mediated by intracellularsignalsLocal clustering of spines has been observed along neurons’ dendritic arbors [68,169]. Similarly, we found that filopodia are clustered along tectal dendrites (Fig-ure 8.12). Such clustering must arise due to biases in filopodial additions or sub-tractions, and could result from localized signaling. To assess clustering of training-induced motility, we compared observed nearest-neighbour distances between sim-ilar growth behaviors to distributions expected if they occurred randomly. Training-induced additions and extensions were more clustered than expected by chance,while subtractions and retractions were not (Figure 8.3a). We next investigatedwhether this spatially clustered structural plasticity is mediated by intracellular sig-naling or extracellular cues such as diffusible molecules. We reasoned that intracel-lular signals would follow intracellular distances between filopodia, while extracel-lular factors would more closely follow straight-line extracellular distances (Fig-ure 8.3b). We sorted pairs of filopodia into two groups: those with intracellular dis-tance similar to their extracellular distance (ID≈ ED), and those with intracellulardistance much longer than their extracellular distance (ID ED; Figure 8.3c). Ex-tracellular distance only predicts clustering for pairs with ID≈ ED, while intracel-lular distance predicts clustering regardless of extracellular distance (Figure 8.3d).These results demonstrate that local signals directing clustered structural changesin these neurons are intracellular, consistent with mechanisms described in morereduced preparations [238].8.3.5 Tectal neuron filopodia show localized evoked glutamatergiccalcium transientsThe clustering of experience-induced dendritic outgrowth by local cues suggeststhat natural sensory activity could instruct the specific patterning of a growingneuron’s dendrites. To investigate this possibility, we used GCaMP6m, a genet-ically encoded calcium indicator, to image dendritic calcium transients [45]. Indeveloping dendrites, filopodia act as sites of synaptic contact similar to spines[151, 189]. Dendritic filopodia showed robust, highly localized calcium transientswith reliable tuning toward OFF, ON, or both stimuli (Figure 8.4b,c,Figure 8.5b).104Figure 8.3: Filopodial motility is clustered along dendrites by intracellular distance. a) Clustering index (Pro-portion of nearest neighbour distances less than 5 µm, relative to chance) for filopodia showing each motilitytype. Filopodial additions and extensions occur significantly closer together than expected by chance. Sub-tractions and retractions are not significantly clustered. N=1504 additions, 1102 subtractions, 1264 extensions,1175 retractions in 21 neurons. ∗∗ : p < 0.01 vs. 0, t-test. b) Distances between filopodia can be measuredeither extracellularly (i.e. straight-line distance, blue), or intracellularly along the dendritic arbor (black). c,d,e)To determine whether motility clustering is mediated by intracellular (ID) or extracellular (ED) distance, wecompared pairs of filopodia having similar ID and ED (ID ≈ ED) to pairs having much longer ID than ED(ID ED). f) Probability that both filopodia in the pair extend, binned by the extracellular distance. ShortEDs predict similar motility for ID ≈ ED pairs, but not for pairs with larger ID. Pairs with larger IDs showless similar motility even when ED is short. g) Probability that both filopodia in the pair extend, binned by theintracellular distance between them. ID predicts similarity in motility regardless of ED. Dashed lines are fitlogistic regressions. ∗∗ : p < 0.01, Likelihood Ratio test. n=2500642 pairs in 21 neurons.In neurons that showed evoked somatic firing, filopodial transients were followedby dendritic shaft transients with slower onset, delayed peak, and lower ampli-tude (Figure 8.4c). These shaft signals were strongly attenuated in filopodia, in-dicating that tectal neuron filopodia can isolate calcium signalling bidirectionally.Consistent with previous findings [25], localized filopodial transients were greatlyreduced upon tectal injection of APV (Figure 8.4b,c). These results indicate thatlocalized filopodial calcium transients reflect glutamatergic synaptic transmissionin tectal neurons.105Figure 8.4: Tectal neuron filopodia show highly localized sensory-evoked NMDA-receptor dependent calciumtransients. a) Raster-scanned two-photon image of a dendritic segment of a GCaMP6m-expressing neuron.Regions of interest are shown for traces in (c). Average of 140 frames. b) Individual frames from serialscanning movies of segment in (a), showing localization of stimulus-induced calcium transients c) Visuallyevoked fluorescence transients from filopodia (F1,F2) and adjacent dendritic shafts (D1,D2). Transients wereevoked by alternating full-field ON and OFF illumination (bottom). After baseline probing, the tectum wasinfused with the NMDAR antagonist APV. Vertical lines denote times of movie frames shown in (b). Filopodiashow robust, localized, NMDAR-dependent evoked transients that are greatly attenuated in adjacent dendriticshaft.1068.3.6 Comprehensive imagingTo date, studies of synaptic activity in vivo have mostly been limited to samplinga small number of synapses at a time. Where large numbers of synapses werestudied, these were imaged sequentially [46, 123, 124, 277] or in more optically-accessible cultures [131]. However, thorough measurement of spatiotemporal in-put patterns, and how these signals are integrated to convey information or driveplasticity, requires more comprehensive imaging techniques. To this end, we con-structed a RAMP microscope capable of imaging over 1500 dendritic sites at 20Hzacross a 100x100x120 µm volume at depths over 200 µm in the intact, awakebrain (Figure 8.13). This microscope can densely sample the entire dendritic ar-bor of these tectal neurons, which have 200-600 synapses, estimated by electronmicroscopy [151]. Our software suite, ARAMiS, incorporates Dynamo to iden-tify and track all filopodial tips, all branchpoints, and dendritic shaft sites at 2 µmincrements, for comprehensive functional and structural imaging (Figure 8.5a).For comprehensive imaging experiments, we used the same timing of probingand training epochs described previously, while imaging individual GCaMP6m-labelled neurons (Figure 8.14). We used a modified probing protocol incorporatingOFF and ON stimuli in alternating sessions, allowing measurement of the tuningpreferences of filopodia between these stimuli. Probing epochs consisted of eight40-second sessions of four trials each. Between each session, a full 3D volumeimage was rapidly acquired for morphological tracking.All neurons showed numerous sites of evoked dendritic calcium transients,even in the absence of evoked somatic firing (Figure 8.5b). Amplitudes of evokedsomatic responses were correlated with the sum amplitude of dendritic responsesacross neurons ( R2 = 0.69, p < 0.001 ; N=14 neurons) and across trials within neu-rons (R2 = 0.24, p < 0.001; N=14 neurons). As observed with 2D imaging, den-dritic responses often originate within filopodia, since 46% (635/1389) of filopodiashowed consistently larger evoked calcium transients at their tip than at their baseto ON or OFF stimuli (N=14 neurons).107Figure 8.5: a) Maximum intensity projection (left) of a GCaMP6-labelled neuron and corresponding wire diagram (right) with random accessimaging sites overlaid. ARAMiS selects sampling points at every branch tip (red), branch point (cyan), the soma, and along each branch shaft(yellow). Scalebar: 15µm b) Dendrogram (left), mean ON and OFF evoked calcium transients (center), and mean spontaneous ∆F/F0 (right)recorded at each site on a neuron. Dots (bottom) denote time of stimulus. n=16 responses per stimulus (ON,OFF), 3200 frames (Spontaneous).c)Simultaneous ∆F/F0 signals at the base and tip of a filopodium. Imaging sites shown at right. A mixing coefficient (red) is fit to estimatesignals spreading from the base to the tip. Scalebar: 2 µm. d) (top) Overlaid single-trial ∆F/F0 traces from the base (blue) and tip (red) ofthe same filopodium. (bottom) Diffusion-compensated signals (navy blue) superimposed over uncompensated tip signals (red). e) Diffusion-compensated mean responses and spontaneous activity of neuron shown in (b). Black bars denote filopodium tips that were not measured. f)Clustering of ON and OFF tuning of evoked tip calcium transients and diffusion-compensated responses in Control tadpoles, tadpoles rearedin complete darkness and tadpoles reared in MK-801. Diffusion-compensated inputs are clustered by input tuning in Control, but not Dark- orMK-801- reared tadpoles. Control: N=13 tadpoles. Dark Rear: N=4 tadpoles. MK-801: N=4 tadpoles. ∗ : p < 0.05, ∗∗ : p < 0.01 vs. 0, t-test.1088.3.7 Sensory-evoked dendritic activity is spatially structuredDendritic signal processing can be strongly influenced by the distribution of inputsalong a neuron’s arbor. Sublinear, linear, and supralinear summation are importantmechanisms of computation, facilitating integration of inputs with spatial organi-zation ranging from distributed to clustered. The clustering of inputs on dendriticbranches can allow branches to act as independent computational units, enablingrich nonlinear integration [207]. Ordering of inputs along dendrites can give riseto emergent receptive field properties [25, 71]. Comprehensive imaging enablesassessment of such clustering, as well as direct measurements of coincident ac-tivity relevant to studies of nonlinear integration [146, 246]. Coincident activitydistributed across dendrites [183, 293] or within a local dendritic region [100, 158]can cue the recruitment of structural and functional plasticity.We examined the organization of ON- and OFF-evoked calcium transientsacross tectal dendrites. Pairs of filopodia with shared ON or OFF preferenceshowed shorter nearest-neighbour distances than expected by chance (Figure 8.5f),indicating that sensory stimulation evokes locally patterned calcium transients.This organization could be due to clustered innervation by similarly-tuned in-puts, or to signals spreading within dendrites. Spreading signals correlate cal-cium transients at neighbouring dendritic sites, causing information to be sharedor confounded between them. The extent of spread depends on buffering, extru-sion, physical compartmentalization, and active or passive dendritic conductances.To measure the extent of spread we calculated noise correlations [83], which re-move the contribution of each site’s tuning to the correlation between two sites,leaving behind correlations due to spreading signals or common input. Noise cor-relations in evoked activity increased at intracellular distances less than 25 µm(Figure 8.6a,c).To compensate for passive signal spread, we treated the calcium signal in eachfilopodium as the sum of active input to the filopodium and diffusing signal en-tering via its base. We quantified the influence of signals at the base of eachfilopodium on those at the tip (Figure 8.5c), and subtracted the scaled base signalfrom the tip signal as a measure of diffusion-compensated activity (Figure 8.5c,dFigure 8.15). Compensation dramatically reduced spatially-structured noise corre-109lations (Figure 8.6b,c), indicating that noise correlations in uncompensated activitylargely represent the scale of spreading calcium signals in tectal neurons.Filopodial activity continued to show strong clustering of ON and OFF pref-erence after diffusion compensation (Figure 8.5f), indicating that tuning of activefilopodial inputs is clustered along dendrites. Neurons of tadpoles reared in dark-ness or raised in 10 µM MK-801 (a noncompetitive NMDAR blocker) showed noclustering of diffusion-compensated responses, indicating that compensation ad-equately removes effects of spreading signals and that establishment of clusteredinputs is activity-dependent.Together, these results show that stimulus-evoked activity is spatially clusteredalong tectal neuron dendrites, and that establishment of this organization requiressensory experience and NMDAR transmission.8.3.8 Local activity cues instruct structural plasticity in dendriticfilopodiaWe next sought to determine whether local patterns of sensory activity instructstructural changes within developing dendrites in vivo. Synaptic transmission[165, 168, 183, 187] and associated calcium signalling [150, 158] act as cues tosynaptic plasticity and structural changes in spines [140, 158, 162]. Calcium sig-nals are sufficient to induce filopodial growth or stabilization in vitro [144, 156–158]. Several types of rules are proposed to govern how calcium transients instructstructural or functional changes. For example, plasticity at a dendritic site mightdepend only on the amplitude of calcium signals at that site [164, 266]. Otherproposed rules are heterosynaptic, involving a comparison of signals across sites[23, 78, 99, 100], for example through the correlation of activity at a synapse to itsneighbours [131] or to somatic spiking [21, 246].Using comprehensive imaging, we evaluated several possible rules regulatingexperience-induced structural changes. We compared motility of filopodia duringST with evoked calcium transients at their tip and base, as well as comparisons be-tween simultaneously measured activity at the tip, adjacent dendritic shaft, neigh-bouring filopodia, and somatic firing. We similarly analyzed spontaneous activitycollected during intervals between stimulus presentations. Spontaneous activitylevels in filopodia were not correlated to evoked responses, suggesting that these110Figure 8.6: Noise Correlations reflect spreading dendritic signals. a) Dendrogram and noise correlations in evoked calcium transient ampli-tudes between all measured sites on a neuron during Epoch 1. Noise correlations are sparse, and largest between nearby sites, indicative ofspreading signals. b) Noise correlations between all measured filopodium tips for the same neuron after diffusion compensation. Compensationabolishes local structure of correlations. c) Mean noise correlations in evoked responses between pairs of filopodium tips in all neurons mea-sured, for calcium transients (dark) and diffusion-compensated signals (light blue), binned by intracellular distance. N=13 neurons, 267404pairs. Residual noise correlations in compensated activity are not the result of systematic over- or under- compensation, as altering diffusioncompensation only increased noise correlations (Figure 8.15).111two types of activity represent distinct sets of inputs (Figure 8.16).Calcium transients measured at filopodium tips did not strongly predict training-induced motility (Figure 8.7a). Correlations of evoked calcium signals or diffusion-compensated activity to those in nearby filopodia or the soma were also not strongpredictors of motility (Figure 8.17). However, nearly all filopodia showing struc-tural changes during training had lower tip than base calcium transients evokedby the training stimulus (78/92, 85%) (Figure 8.7b,d). Filopodia with larger tipthan base responses were nearly all stable (350/364, 96%). This relationship wasspecific to the training (OFF) stimulus, and did not hold for responses to the un-trained stimulus or spontaneous activity (Figure 8.7b). This stimulus-specificityimplies that unstable filopodia do not only represent processes lacking synapses[189], but that training-induced motility is sensitive to the tuning of filopodial in-puts. Training-induced subtractions and retractions occurred in regions of largerdendritic shaft responses than additions, which most often occurred on brancheswith moderate evoked response amplitudes (Figure 8.7c, Figure 8.18).These results demonstrate that patterns of evoked calcium transients predict themotility of dendritic filopodia, and suggest that a comparison between the base andtip of a filopodium may mediate this interaction.8.4 DiscussionIs neuronal activity instructive or permissive to early brain circuit development?We addressed this question on two spatial scales. At the circuit level, we inves-tigated whether patterns of somatic firing direct neuron-to-neuron variations indendrite growth. We found that neurons with different somatic activity profilesexhibit specific, distinct growth patterns, and that blocking either NMDARs orfiring abolishes these differences. Notably, potentiation of evoked somatic firingdrives filopodial pruning, while neurons not showing somatic responses grow dur-ing training. These results are consistent with recent studies associating naturally-induced LTP with decreases in synapse number [28], and learning with loss ofdendritic spines [141, 237]. In the tectum, stable dendritic branches show fewer,larger and more mature synapses [151]. Similarly, afferent retinal axons stabilizewhen their activity is correlated to post-synaptic tectal neurons, and grow when112Figure 8.7: Sensory-evoked activity determines structural changes. a) Mean pretraining OFF-evoked calcium transients at base and tip forstable, subtracted, extended, or retracted filopodia. Base and tip evoked calcium transients alone were not associated with training-inducedmotility. b) Difference in transient amplitudes at tip and base for evoked ON and OFF or spontaneous activity for the four motility classes.Motile filopodia showed lower tip than base evoked calcium transients specific to the training (OFF) stimulus. ∗∗ : p < 0.01 comparedto stable, ANOVA followed by Tukey LSD. c) Cumulative distribution of OFF evoked calcium transients for dendritic shafts at sites ofadditions, subtractions, and stable filopodia and all dendritic shaft sites. ∗∗ : p < 0.01, Kolmogorov-Smirnov test d) Base and Tip OFF-evokedcalcium transients prior to training, and motility during training, for stable, subtracted, extended, or retracted filopodia. (a-d) N = 12 neurons,1089 stable filopodia, 41 subtractions, 37 additions, 27 extensions, 24 retractions, 1972 shaft sites. e-g) Examples of motile processes andcorresponding off-evoked activity at nearby sites. Arrowheads denote sites of subtraction (red), extension (cyan) and retraction (magenta).Upper panels, imaging sites with corresponding mean evoked ∆F/F0 traces at right. Black dot denotes time of OFF stimulus. Timestampson morphological images denote time relative to onset of Spaced Training. Yellow arrowheads denote processes extending beyond the imageframe. Scalebar, 2µm.113it is not [184]. We hypothesize that this pattern of activity-dependent structuralplasticity permits growing neurons well-integrated into functional circuits to refineresponse properties, while affording as-yet unintegrated neurons opportunities toform new relevant connections.We next investigated instructive cues to dendritogenesis at a subcellular scale.We showed that experience-induced additions and extensions of filopodia are spa-tially clustered according to local intracellular cues. Using comprehensive imag-ing, we identified rules that instruct the formation, maintenance, and eliminationof dendritic filopodia in response to focal and global evoked activity. Our resultsdelineate a model where filopodia with lower evoked activity at their tip than theadjacent dendritic shaft are destabilized, subtractions occur in regions of highershaft activity, and additions occur in regions of more moderate shaft activity.These rules can explain the patterns of motility observed in neurons of differentactivity profiles. Neurons undergoing potentiation will have increased evoked den-dritic shaft activity driven by somatic firing and synaptic input. Higher shaft activ-ity, in turn, increases filopodial subtraction rates. Neurons without firing responsesto training nevertheless show responsive subthreshold inputs, leading to moderateshaft activity levels that can recruit clustered filopodial additions. In neurons thatbecome responsive, these additions are followed by a transition to suprathresholdinputs driving filopodial subtraction. These rules may also explain why filopodialsubtractions are not clustered (Figure 8.3a), because regions of less active filopodiaare unlikely to drive high levels of shaft activity.Previous studies have shown that both spontaneous and task-associated activ-ity cause clustering of synaptic inputs [80, 131, 169]. In contrast, serial map-ping of spine tuning in vivo has shown no clustering in several neuron populations[46, 123, 277], but see [45]. The current study presents a possible explanation ofthese disparate results. Tectal neurons show clustered input tuning that is activity-dependent, and the rules governing those inputs are stimulus-specific. If inputsto cortical neurons operate by similar rules, clustering could be specific to stimulithat the neuron processes preferentially, or that the animal previously encountered.Indeed, neurons using nonlinear integration to perform a computation would beexpected to cluster inputs specific to that computation, and not other inputs. Suchpatterns may not be detectable in maps of tuning to a set of pre-selected stimuli.114We envision that a limited set of learning rules will eventually account for thefull variety of experience-dependent growth patterns observed in developing neu-rons. The same learning rules, subject to different patterns of synaptic input drivenby varied experience, will produce different neuronal morphologies and patterns ofconnectivity.8.5 Methods8.5.1 Animal rearing conditionsFreely-swimming albino Xenopus laevis tadpoles were reared in 0.1x Steinbergssolution (1x Steinbergs in mM: 10 HEPES, 58 NaCl, 0.67 KCl, 0.34 Ca(NO3)2,0.83 MgSO4, pH 7.4) and housed at room temperature on a 12 hr light/dark cy-cle. For MK-801 rearing experiments, tadpoles were reared in Steinberg’s solutioncontaining 10 µM MK-801 (Tocris) from Stage 30 to Stage 50, when imagingwas performed. Rearing solution was changed every 12 hours. For dark-rearingexperiments, tadpoles were placed in darkness, and handled only under infraredillumination, from Stage 30 to Stage 50, when imaging was performed. Experi-ments were conducted in accordance with the Canadian Council on Animal Careguidelines, and were approved by the Animal Care Committee of the University ofBritish Columbia Faculty of Medicine.8.5.2 Imaging conditionsFor all experiments, tadpoles were placed in a bath containing 4mM pancuroniumdibromide (a reversible paralytic) for 5 minutes immediately before imaging, thenplaced in a custom-fabricated chamber that immobilizes the tadpole head for awakeimaging and visual stimulation. The tadpole tail was perfused with oxygenated0.1x Steinbergs solution during imaging.8.5.3 Calcium indicator loading and tectal infusionsFor population calcium imaging and targeted single-cell electroporation experi-ments, Oregon Green BAPTA-1 AM (OGB; Molecular Probes, Eugene, OR) waspressure injected into the left optic tectum as described previously, 30-60 min be-115fore imaging. Where actin blocking cocktail was used, 4mM Jasplakinolide (Rand D Systems) and 5 mM Latrunculin A (R and D Systems) were included in thedye-loading solution. Where APV was used, D-APV (50 µM) in Ringer’s solutionwas injected prior to experiment onset (or included in the dye-loading solution inOGB experiments), followed by a second injection of D-APV (50 µM) in Ringer’ssolution prior to training onset. Control experiments in all cases consisted of thesame injection protocol with drug omitted.8.5.4 Targeted neuronal silencingIn neuronal silencing experiments, we co-electroporated GCaMP6m (from Ad-dgene plasmid 40754, inserted into a modified pN1 vector, 3µg/µl) at Stage 35-36and the DREADD HM4Di (3µg/µl) by ventricular electroporation68 at Stage 35-36. Control tadpoles were electroporated with GCaMP6m (3µg/µl) alone. Tad-poles were screened for expression of GCaMP6m in isolated neurons at stage 50.For these experiments, we acquired an additional probing epoch of activity data(’Baseline’) prior to our stimulation protocol, to assess response amplitudes priorto CNO administration. Following the Baseline epoch, tadpoles were anaesthetizedand 5 µM CNO was infused into the tectum. The normal stimulation protocol (seebelow) was initiated following recovery from anaesthesia. During this imaging, theimaging chamber was perfused with Steinberg’s solution containing 0.5 µM CNO.8.5.5 Target neuronsFor all experiments, we targeted type 13b pyramidal neurons [147] in the dorsolat-eral tectum of Stage 50 tadpoles. In this target region at this developmental stage,these neurons show robust visually evoked responses and are relatively stable mor-phologically, with motility rates that can be accurately captured by imaging at 30-minute intervals. Nevertheless, these neurons also robustly show both structuraland functional plasticity with training.8.5.6 Selection of stimuliTo select the stimuli used in these experiments, we earlier tested stimuli in neuronsrandomly labelled by electroporation of G-GECO1.2 [297]. Receptive fields of116neurons in the target region were characterized by presenting spatiotemporal pat-terns of illumination on an array of 8 LEDs arranged in the naso-temporal plane.Full OFF flashes (all 8 LEDs) were most successful in eliciting detectable somaticcalcium responses (8/73 randomly labelled neurons), followed by Full ON flashes(7/73 neurons). 34/73 neurons responded to one or more of the 10 stimuli tested(Localized flash, Full flash, Simulated Motion, Looming, and Receding patternswith ON and OFF stimuli)8.5.7 Conventional two-photon imagingExperiments not involving random-access imaging were performed with a scan-mirror based two-photon microscope adapted from an Olympus FV300 confocalmicroscope (Olympus) and a Chameleon XR Ti:Sapphire laser (Coherent). Pop-ulation calcium imaging (Figure 8.2c,Figure 8.11) was performed at 5Hz, as de-scribed previously [205]. Single cells and their neighbours within a 28x28 µmfield of view were imaged at 6.6Hz (Figure 8.1e,Figure 8.2abd,Figure 8.10).8.5.8 Visual stimulation protocolVisual stimuli were presented via a 635 nm LED (LTL–709E, Lite-On Inc.) placedagainst the eye contralateral to the imaged tectum. Where OFF stimuli were pre-sented, the LED (on at start of trial) was turned off for 50 ms. Where ON stimuliwere presented, the LED (off at start of trial) was turned on for 50 ms. In all cases,Spaced Training (ST) consisted of three 5-minute bursts of high-frequency (0.3Hz,50 ms) OFF stimuli spaced by 5-minute periods of invariant light.For conventional two-photon imaging experiments, visual stimulation consistedof three probing epochs of 30 minutes each (Figure 8.1,Figure 8.2), during whichwe presented one OFF stimulus per minute while recording somatic activity of thetarget neuron and its neighbours. Probing was halted during morphological imag-ing, for a 5 minute period between each epoch. ST was performed between the firstand second probing epoch, for a total stimulation period of 2 hours.Similarly, visual stimulation for random access imaging consisted of three 30-minute probing epochs, with OFF ST presented following Epoch 1, for a totalstimulation period of 2 hours. Each probing epoch consisted of at least 8 prob-117ing sessions. During each probing session, 4 stimuli were shown, consisting ofeither ON or OFF flashes presented with pseudorandom inter-stimulus intervalsranging from 8-12 seconds. ON probing trials and OFF probing trials were al-ternated throughout the experiment. During Morphological imaging, registration,and acquisition planning (roughly 2 minutes between each probing session; seeARAMiS, below), the LED was turned on and off at pseudorandom intervals of 5to 10 seconds, to ensure roughly equal amounts of light and darkness over time.8.5.9 Processing of calcium imaging dataFor conventional calcium imaging experiments, image registration, ROI selection,spatial filtering, and baseline fitting were performed as described previously to ob-tain ∆F/F0 fluorescence traces [205]. Initial data processing for random accessimaging is described separately below. To measure response amplitudes at eachdendritic or somatic site, ∆F/F0 traces were filtered using non-negative deconvo-lution [204] (NND) and the mean of the unfiltered trace over 0.5 seconds priorto the stimulus was subtracted from the peak of the NND filtered trace within 2seconds after the stimulus. A decay time constant of 0.75s was used for filtering.This constant is shorter than the ’off’ time constant of GCamP6m and was selectedto avoid oversmoothing, while reducing high-frequency noise. P-values for eachresponse were calculated as the probability of observing a peak value of the mea-sured size in a filtered trace of Gaussian noise with the same variance as the ∆F/F0trace over 2.5 seconds prior to the stimulus. Mean responses were calculated asthe mean of the unfiltered ∆F/F0 trace over all stimuli of a given type (ON orOFF) in an Epoch. Neurons were classified as responding in a given epoch if thepeak of their mean response differed significantly (p < 0.05) from noise, P-valuescalculates as above.For all experiments, neurons responsive during Epoch 1 were classified as un-dergoing LTP if their response amplitude increased (p < 0.05, t-test); LTD, if theirresponse amplitude decreased (p < 0.05); or No Change, if response amplitudesdid not change, over both Epoch 2 and Epoch 3. Neurons not initially responsivewere characterized as Becoming Responsive, if they were responsive in Epoch 2,or No Response, if they were not responsive for any epoch. Neurons not falling118into any of these categories were labelled Other.8.5.10 Targeted electroporationNeurons were labelled by targeted single-cell electroporation (TSCE) after bo-lus loading of OGB. Because the tectum consists of many cell types [147], andrelatively few cells respond measurably to any given probing stimulus (8/73 forOFF;11%), TSCE is a useful technique as it labels a functionally and morpho-logically homogeneous neuronal population. Using two-photon imaging of OGBfluorescence responses evoked by the stimulus, we select a neuron on the basis ofits response and approach it with a pipette containing Alexa Fluor 594-dextran,which is driven into the cell using a brief train of voltage pulses (Figure 8.1b).Typical electroporation parameters were 18ms train duration, 16V, 200 µs/pulse,and 200 pulses/s. We targeted either responding neurons in the target region ornon-responding neurons adjacent to responding neurons in the target region.Electroporation caused an immediate increase in somatic OGB fluorescenceand transient loss of neuronal responsiveness that returned to pre-electroporationlevels within 20 minutes. Neurons that did not recover response amplitudes within20 minutes invariably showed morphological signs of damage, such as dendriticblebbing (data not shown). Physical damage, if present, is usually evident withinminutes of electroporation.To determine the effects of electroporation on neurons, we compared stimulus-driven growth, activity and plasticity in neurons immediately after TSCE labelling,neurons labelled one day prior, and non-electroporated neurons (activity imagingonly). We measured ST-induced functional plasticity in both the electroporated tar-get neuron and its nearest neighbours loaded with OGB. Electroporated and neigh-bouring non-electroporated cells showed no differences in plasticity outcomes withST (Figure 8.8). We compared comprehensive morphometric measures in neuronslabelled 30 minutes before the experiment and neurons labelled one day earlier by’shadow TSCE’ [130]. Neurons electroporated 30 minutes before the experiment,neurons electroporated one day earlier, and neurons labelled by ventricular electro-poration of GCaMP6m at Stage 35-36 (and subsequently screened to find isolatedexpressing type 13b neurons) showed indistinguishable patterns of motility, filopo-119dial additions, and subtractions (Figure 8.8). Altogether, tectal neurons appear toexhibit normal evoked responses and undergo normal experience-induced plastic-ity and structural growth after electroporation.8.5.11 Morphological imaging and dynamic morphometricsAll morphological imaging, for both conventional and random-access microscopes,was performed by 3D raster scanning with a Z resolution of 1.5 µm. 3D Rasterscans can be performed quickly (<1 min) by the random access microscope be-cause ARAMiS selects a minimal region to scan that encompasses the full XY ex-tent of the neuron, allowing for drift or motion since the last morphological imagestack, and because random access imaging minimizes dead time at the end of eachscan line. Neuron tracing and Dynamic Morphometrics were performed using ourfreely available Dynamo software ( Neuron imagestacks were traced at 30-minute intervals. Image stacks were denoised with a mod-ified version of CANDLE [54] using the following settings: smoothing parameter,0.8; patch radius, 1; search radius, 3. Filopodia were defined as any unbranchedprocess shorter than 5 µm. Processes longer than 5 µm were classified as branches.The axon and axonal processes were excluded from analysis. For clustering analy-ses, interstitial filopodia were defined as those originating more than 5 µm from abranch tip. For analyses of activity of motile filopodia (Figure 8.7), filopodia werecharacterized as extending or retracting during training if their measured lengthchanged by 0.5 µm or more during the training epoch, and stable otherwise. Allanalyses were performed in Matlab.8.5.12 Spatial clusteringTo measure spatial clustering of various properties, we computed the nearest neigh-bour distance (NND) for that property, i.e. the distance to the nearest filopodiumwith the same type of motility. For example, the NND for a filopodium added at agiven timepoint is the distance to the base of the nearest other filopodium added atthat timepoint. We excluded filopodia at the tips of branches (’terminal filopodia’)from our clustering analyses; terminal filopodia show different patterns of motil-ity than interstitial filopodia [115], and are inherently clustered by virtue of being120at branch tips. To measure clustering, we compared the proportion of NNDs lessthan 5 µm to the proportion expected under the null (non-clustered) distribution,obtained by Monte Carlo sampling. For properties of existing filopodia (such assubtractions, growth, retraction, or response amplitude) the null distribution wascalculated by randomly reassigning observed values among all existing filopodiaand measuring the NND on this randomized set. The null distribution for additionswas calculated differently, by uniform randomly reassigning the observed numberof addition sites along the length of the dendritic arbor. These reassignments wereperformed 10000 times for each neuron to generate the null distribution for eachquantity. The term ’clustering index’ represents the observed proportion of NNDsless than 5 µm minus the expected proportion. Error bars denote standard error ofthe proportion, from observed data and Monte Carlo sampling, propagated throughthe above calculations.8.5.13 2D synaptic imaging2D imaging of localized dendritic calcium transients (Figure 8.4, supplementalmovie 1) was performed on a conventional raster-scanning two-photon microscope.We imaged dendritic segments of mature tectal neurons in stage 50 tadpoles at 930nm excitation. Short segments with filopodia visible in the focal plane were imagedat 10Hz while visual stimulation switched between on and off at a mean intervalof 8s. After this initial imaging period, tadpoles were removed from the imagingchamber and the tectum injected with 50 µM D-APV in frog Ringer’s solution,or Ringer’s solution alone, and a second round of imaging was performed. Sup-plementary movies and corresponding individual frames of 2D synaptic imagingwere produced by fitting ∆F/F0 for each pixel in the frame, after rigid body imageregistration.8.5.14 Random access microscope designTo image localized dendritic activity, we constructed a random access microscope.Lateral (X-Y) scanning was performed with a pair of crossed large-aperture (9mm)acousto-optic deflectors (AODs; Isomet OAD1121-XY). Prior to scanning, the ex-citation beam was collimated and expanded through coupled achromatic doublets,121and its polarization rotated to optimize diffraction by the AODs. The diameterof the beam entering the AODs was controlled by an aperture (A1). Combinedwith interchangeable achromatic doublets acting as zoom lenses, adjustment ofthis aperture allows trade-offs between AOD access time, scan range, and focalspot size. In dendritic imaging experiments A1 was set to 6mm, producing an8mm spot fully filling the back aperture of the objective (LUMFLN 60XW, 1.1NA,Olympus), a lateral scan range of 110x110 µm, and an access time (time taken forthe acoustic wave in the AOD to fully cross the excitation beam) of approximately10 µs. We used a Ti:Sapphire excitation laser with integrated tunable dispersioncompensation (Chameleon Vision II, Coherent) to compensate for temporal dis-persion of ultrafast pulses introduced by the AODs. All random-access imagingof GCaMP6m-expressing neurons was performed at 910 nm excitation. Angulardispersion introduced by the AODs was compensated using a prism in the opti-cal path immediately after the deflectors [148] oriented at 45 degrees to the twodeflection axes. The apex angle of the prism was selected to best compensate forangular dispersion in the center of the field of view. A small portion of the exci-tation beam is reflected by a coverslip to a photodiode (UDT Sensors, PIN-10D)prior to entering the objective, allowing us to calibrate the intensity of the diffractedbeam across the field of view for spatially uniform excitation, and to compensatefor laser intensity fluctuations during imaging. Axial scanning of the excitationbeam was produced by a piezo objective mount (PiezoJena, MIPOS 100PL). Dur-ing random access dendritic imaging, the objective was scanned at high frequency(20-100Hz) under closed loop control, through the entire neuron volume. Oscil-lation of the objective was initiated 10 seconds prior to initiation of imaging toensure stability, and the motion profile was kept constant by software PID controlduring imaging. Radio frequency signals driving the AODs were produced by twocustomized programmable Direct Digital Synthesis units that generate sequencesof high-resolution chirped signals from internal memory stores. Emitted fluores-cence was detected with H7422-40 PMT modules (Hamamatsu Photonics, Japan).PMT signals were amplified with SR570 transimpedance amplifiers (Stanford Re-search Systems, Sunnyvale, CA). Hardware was controlled by, and signals wereacquired on, two PCI-6110 DAQ boards (National Instruments, Austin, TX), at arate of 5MHz. PMT signals were deconvolved with non-negative deconvolution122(NND) to recover pixel intensities [204].8.5.15 ARAMiS - a random access microscopy suiteAll random access imaging was performed using our freely available imaging suite,ARAMiS. We recorded comprehensive dendritic activity from single neurons inthe target region of the tectum labeled with GCaMP6m by ventricular electropo-ration. After collecting an initial image stack, Dynamo was used to trace the neu-ron. ARAMiS automatically selects imaging sites at every dendritic branch andfilopodium tip, every branchpoint, and at 2 µm intervals along the entire dendriticarbor. Activity at these sites was recorded simultaneously at 20Hz in 40-secondprobing sessions (see Visual Stimulation, above). We obtained a full morphologi-cal image stack between each session, to compensate for drift or growth, and trackmorphological plasticity.For each 40-second probing session, ARAMiS optimizes the range and profileof the objective Z-axis scan to match the time spent in any given imaging plane tothe number of points to be imaged in that plane. Each target point is sampled upto, but no more than, 5 times during each oscillation, including both its rising andfalling phase.Morphological image stacks obtained between probing sessions were used forautomated registration. Registration consisted of initial global registration fol-lowed by registration of local regions (approximately 5 µm in each direction) sur-rounding each branch tip and branch point. Intermediate points were inferred usingregistered branch points as landmarks. All registrations performed were rigid-bodytranslations based on image stack cross-correlations to the previous morphologicalimage stack, typically acquired roughly 2 minutes earlier. Where the error betweenregistered images exceeded a threshold, corresponding points were selected man-ually. Newly added filopodia were manually traced in Dynamo at the start of eachnew Epoch (every 30 minutes) for inclusion in comprehensive imaging.During fast scanning, each site was imaged as a short line scan 2.7 µm inlength, allowing detection of drift and sample motion. Recordings that showeddrift or motion artifacts were discarded from the time of motion onwards. In ourparalyzed tadpole preparation, only a small fraction (<10%) of recordings showed123detectable drift over a single imaging cycle, which lasts roughly 2 minutes.8.5.16 Processing of random access imaging dataFollowing recording, correct registration of imaging sites to the neuron’s structurein each session was confirmed by manual examination. Each line scan was regis-tered across frames of the fast scan to detect drift or sample motion, and each moviewas further manually inspected for drift. Traces that showed drift of more than 2pixels were discarded from the time of drift onwards. Drift was rare, with lessthan 10% of all traces showing more than 2 pixels of drift over 40 seconds. Rawfluorescence traces for each imaging site were generated by summing intensity ofthe center 7 pixels of each aligned 11-pixel line scan. ∆F/F0, response amplitudes,and plasticity metrics were obtained from these raw fluorescence traces in the samemanner as for conventional calcium imaging data.8.5.17 Compensation of spreading calcium signalsTo compensate for calcium signals spreading into filopodium tips from the den-dritic shaft, so as to isolate signals originating within filopodia, we adapted a tech-nique used to compensate for backpropagating action potentials in mouse corticalneuron spines15. The joint distribution of base and tip ∆F/F0 amplitudes typicallyshows two arms (Figure 8.5c) or a fan shape, but always has a clear lower limit onthe ratio of base to tip fluorescence. We attribute this lower limit to some proportionof calcium signals at the base that enter the tip by diffusion or similar processes. Wefit this lower limit (the mixing coefficient) for each filopodium, for each probingsession, using robust regression (function robustfit in the Matlab statistics toolbox)for data frames where base ∆F/F0 exceeded tip ∆F/F0. Compensated tip signalswere calculated as (tip ∆F/F0) (mixing coeffient)*(base ∆F/F0) and NND filteredwith a 0.75s time constant.8.5.18 Spontaneous activity’Spontaneous’ activity was recorded during periods between stimulus presenta-tions. All activity recorded from 5 seconds following each stimulus presentationup to the frame before the following stimulus was included in analysis of sponta-124neous activity. Mean spontaneous activity (Figure 8.5b,e,Figure 8.16,Figure 8.17)is the mean ∆F/F0 value for a given site over this time.8.5.19 Measurement of correlationsCorrelations were calculated as the Pearson correlation between traces for each pairof sites on the neuron. To obtain noise correlations, correlations in evoked activitywere calculated separately over OFF and ON trails then averaged, thus removingthe effect of stimulus identity on the correlation. Prior to calculation of noise cor-relations, global variations in evoked responses (e.g., due to run-down of responsesto repeated stimulation) were removed by subtracting the reconstructed signal pro-duced from the largest singular value of the response matrix. For spatial analysis ofdiffusion compensated noise correlations, correlations between filopodia sharing ameasurement point at their base (i.e. filopodia with a distance of 0) were not in-cluded in analysis, to avoid including spurious correlations due to shared variationin the base data used for compensation.Correlations in calcium signals and diffusion-compensated inputs (Figure 8.17)are total correlations, not noise correlations; they include variations due to theidentity of the stimulus, and reflect the overall similarity in calcium fluctuations atthe two sites being correlated. In these figures, ’Correlation to nearby filopodia’refers to the mean correlation to all filopodia pairs at intracellular distances lessthan 10 µm.8.5.20 Spatial clustering of local dendritic responsesWe used diffusion compensated signals to identify the tuning preferences of indi-vidual filopodia (see Diffusion Compensation). To assess clustering of ON/OFFtuning of filopodial responses, we calculated the clustering index by measuringNNDs separately for each stimulus, over filopodia with mean diffusion-compensatedresponses to that stimulus exceeding a threshold. The threshold for selection wasthe 80th percentile of evoked response amplitudes for each neuron, bounded aboveand below at 0.2 and 0.6 ∆F/F0.1258.5.21 Image processingFor large-scale maximum intensity projections (Figure 8.1c, Figure 8.5a), regionsmore than 2 microns away from neuronal processes (as determined by Dynamo)were dimmed by a factor of 10, to make it possible to see small processes through-out the entire depth of the image. The image in Figure 4c was square-root trans-formed to show dimmer processes.8.5.22 StatisticsStatistical tests are reported alongside p-values. Matlab code for Monte Carlo sim-ulations and statistical tests are available by contacting the authors.8.6 Supplementary figuresThe following pages contain supplementary figures referenced in the text.126Figure 8.8: Neurons electroporated by TSCE show normal patterns of activity and growth. a) Distribution of plasticity outcomesto OFF spaced training for OFF responsive neurons electroporated with TSCE and OFF responsive unelectroporated neighboursin OGB-filled tecta. Electroporated neurons show the same distribution of plasticity outcomes as neighbours. b) Plasticity ofneighbour cells compared to plasticity of the target cell. c) Total dendritic branch length (TDBL), addition rate, and subtractionrate, measured 20 minutes, or 1 day after TSCE. Neurons show show normal patterns of growth measured soon after labelling.127Figure 8.9: Neurons of different activity profiles show similar dendritic arbor sizes. Total dendritic branch length for neuronsshowing each of the 4 activity profiles in response to OFF Spaced Training. Adjacent black markers and errorbars denote corre-sponding means and standard errors for each group. ns: p > 0.05, ANOVA.128Figure 8.10: Motility and stability of filopodia in neurons of different evoked activity profiles during OFF Spaced Training. a)Net Motility expressed as average distance change per filopodium (extensions-retractions). Neurons undergoing LTP show netretraction of filopodia during spaced training, while initially nonresponsive neurons show increased net extension relative tobaseline. Positive numbers denote more extension than retraction, negative numbers vice versa. Data points are mean +/- SEM,with neurons as replicates. ∗ : p < 0.05, ∗∗ : p < 0.01 relative to same group at Epoch 1, ANOVA followed by Tukey LSD. b)Survival of filopodia emerging during each epoch, grouped according to the neuron’s activity profile. Points are the proportionof filopodia born in the specified epoch remaining after each subsequent epoch. Insets show comparisons of survivorship afterEpoch 3; ∗ : p < 0.05, ∗∗ : p < 0.01, χ2 test over all groups, followed by pairwise 2x2 comparisons.129Figure 8.11: Distribution of activity profiles for neurons in OGB-loaded tecta of tadpoles infused with APV, or vehicle alone.APV/vehicle infusions were performed during initial loading of OGB and again immediately prior to Spaced Training. ∗ : p <0.05, ∗∗ : p < 0.01, chi2 testFigure 8.12: Filopodia are clustered along tectal neuron dendrites. 8.5% more filopodia have nearest neighbours within an intra-cellular distance of 5 µm than would be expected if filopodia occurred randomly along dendritic branches. Filopodia near thetips of branches were excluded from analysis. N = 1752 interstitial filopodia in 22 neurons.130Figure 8.13: Schematic of RAMP microscope. Lateral (X-Y) scanning was performed with a pair of crossed large-aperture (9mm)acousto-optic deflectors (AODs; Isomet OAD1121-XY). Prior to scanning, the excitation beam was collimated and expandedthrough coupled achromatic doublets (L1 L2), and its polarization rotated (/2) to optimize diffraction by the AODs. The diameterof the beam entering the AODs was controlled by an aperture (A1). Combined with interchangeable zoom lenses (L3, L4),adjustment of this aperture allows trade-offs between AOD access time, scan range, and focal spot size. In dendritic imagingexperiments A1 was set to 6mm, producing an 8mm spot fully fulling the back aperture of the objective (OBJ; LUMFLN 60XW,1.1NA, Olympus), a lateral scan range of 110x110µm, and an access time of approximately 10 µs. We used a Ti:Sapphireexcitation laser with integrated tunable dispersion compensation (Chameleon Vision II, Coherent) to compensate for temporaldispersion of ultrafast pulses introduced by the AODs. All random-access imaging of GCaMP6m-expressing neurons was per-formed at 910 nm excitation. Angular dispersion introduced by the AODs was compensated using a prism in the optical pathimmediately after the deflectors oriented at 45 degrees to the two deflection axes. The apex angle of the prism was selected tobest compensate for angular dispersion in the center of the field of view. A small portion of the excitation beam is reflected bya coverslip (CS) to a photodiode (PD; UDT Sensors, PIN-10D) prior to entering the objective, allowing calibration of diffractedbeam intensity across the field of view for spatially uniform excitation. Axial scanning of the excitation beam was produced by apiezo objective mount (PiezoJena, MIPOS 100PL). During random access dendritic imaging, the objective was scanned at highfrequency (20-100Hz) under closed loop control, through the entire neuron volume. Oscillation of the objective was initiated10 seconds prior to initiation of imaging to ensure stability, and the motion profile was kept constant by software PID controlduring imaging. Radio frequency signals driving the AODs were produced by two customized programmable Direct DigitalSynthesis units that generate sequences of high-resolution chirped signals from internal memory stores. Emitted fluorescencewas detected with H7422-40 PMT modules (Hamamatsu Photonics). PMT signals were amplified with SR570 transimpedanceamplifiers (SRS). Hardware was controlled by, and signals were acquired on, two PCI-6110 DAQ boards, at a rate of 5MHz.131Figure 8.14: During RAMP experiments, an initial image stack was acquired and traced with Dynamo. After initial tracing,comprehensive imaging of activity was alternated with rapid volume acquisition, which was used to adjust for sample drift andcompensate for neuron growth. Each cycle of random access and morphological imaging takes roughly 2 minutes.132Figure 8.15: Effect of biasing fitted mixing coefficients on measured noise correlations in diffusion compensated responses. Eitherincreasing or decreasing mixing coefficients decreases independence of resulting diffusion-compensated signals, indicating thatour fitting is not systematically biased. Errorbars denote corresponding means and standard errors. N = 14 neurons, 267404filopodium pairs133Figure 8.16: Spontaneous activity and evoked activity show different patterns across filopodia. a) Noise correlations for sponta-neous activity in raw calcium signals (top) and diffusion compensated signals (bottom). Spontaneous activity continues to showspatially structured correlations after diffusion compensation. Compare Figure 6b. b) Mean noise correlations in all neuronsmeasured for spontaneous (red tones) and evoked (blue tones) activity, between pairs of filopodium tips, for calcium transients(black) and diffusion-compensated signals (blue), binned by intracellular distance. Spontaneous activity is more highly correlatedthan evoked activity. ∗∗ : p < 0.01, ANCOVA followed by Tukey LSD. N=13 neurons, 267404 pairs. c) Relationship betweenevoked response amplitudes and spontaneous activity levels in dendritic filopodia. Response amplitudes are not correlated tospontaneous activity levels, suggesting that stimulus-evoked activity and spontaneous activity represent largely distinct circuitsin the tectum. N = 12 neurons, 778 filopodia,134Figure 8.17: Correlations in activity are not strongly associated with training-induced motility in tectal neuron filopodia. Motile orretracted filopodia did not differ from stable filopodia in correlations to neighbours (< 10 µm), or to the soma, in either evokedor spontaneous activity. N = 12 neurons, 1089 stable filopodia, 41 subtractions, 27 extensions, 24 retractionsFigure 8.18: Cumulative distribution of OFF evoked calcium transients for dendritic shafts at sites of filopodial retractions (ma-genta) and stable filopodia (light gray). ∗ : p < 0.05, Kolmogorov-Smirnov test. N = 12 neurons, 1089 stable filopodia, 24retractions135Chapter 9ConclusionsHere I have presented several methods that facilitate optical monitoring of infor-mation processing in the awake brain, and their application to problems in develop-mental neuroscience. Throughout these studies, I have emphasized the importanceof measuring patterns of activity across many sites. It is an inconvenient but unde-niable fact that information is conveyed by patterns, which are difficult to measurewith conventional tools. We are only recently making headway into this problem,which has required the development of new microscopes, contrast agents, label-ing methods, analysis techniques, and mathematical models. This thesis describescontributions to each of these fields.9.1 Learning mechanismsWhat new conclusions can be drawn from my work regarding the developmentof functional neural circuits? Despite being conducted on very different spatialscales, the studies presented in Chapter 3 and Chapter 8 provide converging linesof evidence for a novel learning mechanism.Training drives the specialization of neurons that are already responsive to thestimuli being trained. This is evidenced by changes in their tuning curves, whichbecome sharper, by their functional network connectivity, which becomes morespecific, and by the loss of filopodia not activated by the training stimulus. Theseresults, at both the single-cell and circuit levels, fit well with our established un-136derstanding of developmental learning, which associates neuronal maturation withsynaptic pruning.At the same time, however, neurons that do not respond to a stimulus grow out,forming new filopodia (Chapter 8) and becoming more functionally interconnected(Chapter 3). This growth likely serves multiple purposes. It may broaden the spaceof inputs that these neurons receive, allowing them to incorporate into the activatedcircuits. Moreover, it may combat the overspecialization of the network as a whole,by producing synapses responsive to less active stimuli, offsetting their loss amongthe more active neurons. This synapse formation may maintain the representationof a diverse range of receptive fields within the network even as some neurons be-come highly specialized. As demonstrated in Chapter 3, NMDAR-mediated mech-anisms that prevent overspecialization are essential for improvements in populationencoding. Though far less studied, these processes which drive growth and con-nectivity of less-active elements in the developing neural circuit are probably asimportant for learning as pruning in active neurons.One of the challenges that has limited our ability to train large artificial neu-ral networks has been the so-called ’variance trap’. Most artificial networks, andour brain, rely on stochastic activation to explore the high-dimensional space ofpossible activity patterns. The variance trap occurs when artificial networks spe-cialize too rapidly, with some nodes being driven so strongly that they are eitheralways active, or always inactive, and alternative patterns over those nodes arenever sampled. This overspecialization constrains the space of activity patternsthat the network explores, and prevents the circuit from learning because of a lackof error signals. This problem is not solved by slowing specialization, because thatprevents learning equally quickly. The variance trap is considered one of the mostsignificant problems preventing the effective training of large neural networks, anda general solution to this problem has yet to be described.One interpretation of the learning mechanisms described above – respondingneurons becoming more refined while non-responding neurons grow out – is thatthey allow neural circuits to specialize while maintaining enough diversity to sam-ple a wide space of potential activity patterns, avoiding a variance trap. Testingthis hypothesis will require modeling of synaptic and population activity with andwithout these processes at play. This theory also predicts that new synapses formed137during training would be tuned to underrepresented upstream features pruned frommore active neurons, which has yet to be shown. The emerging field of compre-hensive imaging would certainly benefit from further experimental manipulations.Nevertheless, it is exciting that such questions regarding learning rules and infor-mation processing can now be asked and answered in biological systems with thetools I have developed. The ability to measure the quantities relevant to artificialneural networks - synaptic input driving neuronal firing, functional and structuralconnectivity – made possible by these tools, enables a richer correspondence be-tween neuroscience and artificial intelligence.Another important advance has been the identification of structural learningrules driven by sensory activity patterns in awake animals. Previous studies haveelucidated mechanisms governing dendrite growth and pruning, in far greater depththan I have here, in the context of spontaneous activity in brain slices or other invitro preparations. However, to understand how these mechanisms and others coulddrive learning and information processing requires studies in awake animals. In theintact brain, patterns of activity represent information about the environment. Thespace of possible activity patterns that pass through a neural circuit in the brainis highly structured, and it is this structure that the brain adapts to when learning.Such patterns are vastly altered in vitro, and while many of the same plasticitymechanisms likely apply in vivo, they may combine to produce emergent phenom-ena, and other mechanisms may be recruited. By tracking patterns of activity inresponse to multiple sensory stimuli, I have shown how information, not simplyactivity, drives structural and functional plasticity.The ability to comprehensively image information processing in a single neu-ron represents a wholly new paradigm in the study of circuit function. The synapticinputs to a neuron represent, in one way, a complete set: all the presynaptic neu-rons within the circuit that contribute to the computation that the target neuronperforms. We cannot observe neuronal somata with a similar completeness. Evenrecent efforts, yet to be published, to image neuronal activity at cellular resolu-tion across the entire brain surface with large-field-of-view microscopes, will leavemost of the neurons relevant to performing a given computation unobserved. Fromthe perspective of reductionism, with the goal of carving nature at its joints, com-prehensive imaging represents a promising horizon. The ability to fully observe138individual neurons and their connections may greatly inform our understanding ofthe circuits they form.9.2 Strengths, weaknesses, and future directionsThe data I collected have provided new insights into how the brain coordinatesfunctional plasticity and connectivity across groups of neurons and how experi-ence shapes the detailed structure and function of individual neurons. One strengthof these studies has been the richness of the data collected and the detailed level ofanalysis conducted. For example, these studies include the most detailed data everpresented on the structure and activity patterns of individual neurons in vivo. Manyphenomena can be observed in the supplemental videos associated with Chapter 8,such as backpropagating action potentials, spreading activity, and dendritic inte-gration, which have yet to be analyzed. Because we measured the activity of entireneurons, we have produced a rich database of structural and functional data thatcan be retroactively analyzed to form new hypotheses. I hope that both the meth-ods and the data already collected represent a lasting contribution to the study ofexperience-dependent developmental plasticity.These studies also have significant limitations that I hope will be addressed inthe future. Perhaps the most important limitation deals with the relationship be-tween optically measured calcium transients and changes in membrane potential,which ultimately underlie all long-distance communication between neurons. Ifappropriate sensors were available, measuring voltage could in some ways providericher information than monitoring calcium by enabling direct observation of ex-citatory and inhibitory potentials. In general, voltage fluctuations are more rapidthan calcium currents, and the size of synaptic calcium transients is not necessarilyassociated to the amplitude of post-synaptic potentials [253]. Alternatively, mea-surement of extracellular glutamate [167] or other neurotransmitters may be a moreappropriate measure where presynaptic activity levels are of interest.Either of these modalities would strain, though possibly not exceed, the imag-ing speeds achievable with our current microscope. Our microscope is capable ofimaging an entire tectal neuron at 100Hz, or 10 ms/volume. Synaptic glutamatetransients measured with iGluSnRF [167] last roughly 50 ms. The speed of mea-139sured voltage transients is limited by conformational changes in the sensor, andmay or may not exceed our imaging rates, but no sensor has yet been producedwith signals that even approach those required for in vivo imaging of subcellularpotentials.Throughout these studies, we have used intensiometric calcium indicators ratherthan ratiometric ones. This is certainly a limitation for the development of mech-anistic models of activity-dependent plasticity. We have shown that the relativeamplitudes of calcium transients are important in determining structural changesin dendrites, but more quantitative measurements would facilitate both modelingefforts and the identification of subcellular mechanisms mediating these changes.Ratiometric indicators would allow us to monitor concentrations of calcium at eachsynapse, whereas intensiometric measurements are only coarsely related to the am-plitude of currents, and give little indication of actual concentrations.There are many things that could have been studied with our microscope thatare not presented in this thesis. Most conspicuously, the random access imagingstudies focus on how activity drives structural changes, while many neuroscientistsmay be more interested in functional changes and dendritic integration relevantto learning in more mature neurons. A large amount of functional data collectedafter training (Epochs 2 and 3) is relevant to these issues, but remains unreported,and will certainly be analyzed in the future. Current efforts in the lab focus onsome of these questions, including building biophysical models of tectal neuronsand simulating the underlying conductances that produce calcium fluctuations, aswell as monitoring neurons responding to richer stimuli to investigate dendriticcomputation and stimulus encoding.The approach to synaptic imaging described here would not be suitable forstudies in adult or even adolescent mice, for reasons of sample motion outlined inthe introduction. Nevertheless, I am confident that it will very soon be possibleto perform comprehensive imaging of entire cortical neurons in awake, behavingmice. The most promising approaches to this problem have yet to be presented inthe scientific literature, but are active areas of research for several laboratories.140Bibliography[1] G. B. Airy. On the diffraction of an object-glass with circular aperture.Transactions of the Cambridge Philosophical Society, 5:283, 1835. →pages 13[2] C. D. Aizenman and H. T. Cline. Enhanced visual activity in vivo formsnascent synapses in the developing retinotectal projection. Journal ofneurophysiology, 97(4):2949–2957, 2007. → pages 23[3] E. U. Akkaya and S. Turkyilmaz. A squaraine-based near IR fluorescentchemosensor for calcium. Tetrahedron letters, 38(25):4513–4516, 1997.→ pages 75[4] M. Albota, D. Beljonne, J. L. Brdas, J. E. Ehrlich, J. Y. Fu, A. A. Heikal,S. E. Hess, T. Kogej, M. D. Levin, S. R. Marder, D. McCord-Maughon,J. W. Perry, H. Rckel, M. Rumi, G. Subramaniam, W. W. Webb, X. L. Wu,and C. Xu. Design of organic molecules with large two-photon absorptioncross sections. Science (New York, N.Y.), 281(5383):1653–1656, Sept.1998. ISSN 0036-8075. → pages 75[5] A. Antonini, M. P. Stryker, and others. Rapid remodeling of axonal arborsin the visual cortex. Science, 260(5115):1819–1821, 1993. → pages 97[6] B. N. Armbruster, X. Li, M. H. Pausch, S. Herlitze, and B. L. Roth.Evolving the lock to fit the key to create a family of g protein-coupledreceptors potently activated by an inert ligand. Proceedings of the NationalAcademy of Sciences, 104(12):5163–5168, 2007. → pages 103[7] E. Arunkumar, C. C. Forbes, B. C. Noll, and B. D. Smith.Squaraine-derived rotaxanes: sterically protected fluorescent near-IR dyes.Journal of the American Chemical Society, 127(10):3288–3289, Mar. 2005.ISSN 0002-7863. doi:10.1021/ja042404n. → pages 75141[8] B. B. Averbeck and D. Lee. Neural noise and movement-related codes inthe macaque supplementary motor area. The Journal of Neuroscience: TheOfficial Journal of the Society for Neuroscience, 23(20):7630–7641, Aug.2003. ISSN 1529-2401. → pages 36, 38[9] B. B. Averbeck, P. E. Latham, and A. Pouget. Neural correlations,population coding and computation. Nature Reviews. Neuroscience, 7(5):358–366, May 2006. ISSN 1471-003X. doi:10.1038/nrn1888. → pages 36[10] D. Axelrod. Total internal reflection fluorescence microscopy in cellbiology. Traffic, 2(11):764–774, 2001. → pages 13[11] M. W. Bagnall, L. E. McElvain, M. Faulstich, and S. du Lac.Frequency-independent synaptic transmission supports a linear vestibularbehavior. Neuron, 60(2):343–352, 2008. → pages 7[12] R. Bakalova, Z. Zhelev, D. Kokuryo, L. Spasov, I. Aoki, and T. Saga.Chemical nature and structure of organic coating of quantum dots is crucialfor their application in imaging diagnostics. International Journal ofNanomedicine, 6:1719–1732, 2011. ISSN 1178-2013.doi:10.2147/IJN.S17995. → pages 74[13] R. E. Baker, P. Wolters, and J. van Pelt. Chronic blockade ofglutamate-mediated bioelectric activity in long-term organotypicneocortical explants differentially effects pyramidal/non-pyramidaldendritic morphology. Developmental brain research, 104(1):31–39, 1997.→ pages 23[14] R. E. Baker, P. A. Dijkhuizen, J. Van Pelt, and J. Verhaagen. Growth ofpyramidal, but not non-pyramidal, dendrites in long-term organotypicexplants of neonatal rat neocortex chronically exposed to neurotrophin-3.European Journal of Neuroscience, 10(3):1037–1044, 1998. → pages 23[15] H. Barlow and P. Foldiak. Adaptation and decorrelation in the cortex.1989. → pages 3, 28, 31, 45, 51, 52[16] M. F. Bear. NMDA-receptor-dependent synaptic plasticity in the visualcortex. Progress in brain research, 108:205–218, 1996. → pages 23[17] P. V. Belichenko, E. E. Wright, N. P. Belichenko, E. Masliah, H. H. Li,W. C. Mobley, and U. Francke. Widespread changes in dendritic andaxonal morphology in mecp2-mutant mouse models of rett syndrome:Evidence for disruption of neuronal networks. Journal of ComparativeNeurology, 514(3):240–258, 2009. → pages 91142[18] M. R. Bennett. History of the Synapse. CRC Press, Sept. 2003. ISBN9780203302545. → pages 12[19] P. Berkes, F. Wood, and J. W. Pillow. Characterizing neural dependencieswith copula models. In Advances in neural information processing systems,pages 129–136, 2009. → pages 60[20] J. Bewersdorf, R. Pick, and S. W. Hell. Multifocal multiphotonmicroscopy. Optics letters, 23(9):655–657, 1998. → pages 17[21] G.-q. Bi and M.-m. Poo. Synaptic modifications in cultured hippocampalneurons: dependence on spike timing, synaptic strength, and postsynapticcell type. The Journal of neuroscience, 18(24):10464–10472, 1998. →pages 110[22] W. Bialek and S. Setayeshgar. Physical limits to biochemical signaling.Proceedings of the National Academy of Sciences of the United States ofAmerica, 102(29):10040–10045, July 2005. ISSN 0027-8424.doi:10.1073/pnas.0504321102. → pages 49[23] E. L. Bienenstock, L. N. Cooper, and P. W. Munro. Theory for thedevelopment of neuron selectivity: orientation specificity and binocularinteraction in visual cortex. The Journal of Neuroscience, 2(1):32–48, Jan.1982. ISSN 0270-6474, 1529-2401. → pages 24, 110[24] D. L. Bishop, T. Misgeld, M. K. Walsh, W.-B. Gan, and J. W. Lichtman.Axon branch removal at developing synapses by axosome shedding.Neuron, 44(4):651–661, 2004. → pages 22[25] J. H. Bollmann and F. Engert. Subcellular topography of visually drivendendritic activity in the vertebrate visual system. Neuron, 61(6):895–905,2009. → pages 10, 71, 98, 105, 109[26] E. J. Botcherby, R. Jukaitis, M. J. Booth, and T. Wilson. An opticaltechnique for remote focusing in microscopy. Optics Communications, 281(4):880–887, 2008. → pages 94[27] T. Bourgeron. A synaptic trek to autism. Current opinion in neurobiology,19(2):231–234, 2009. → pages 91[28] J. N. Bourne and K. M. Harris. Coordination of size and number ofexcitatory and inhibitory synapses results in a balanced structural plasticityalong mature hippocampal CA1 dendrites during LTP. Hippocampus, 21(4):354–373, 2011. → pages 112143[29] S. Bradbury. The Evolution of the Microscope. Elsevier, May 2014. ISBN9781483164328. → pages 12[30] D. H. Brainard. The psychophysics toolbox. Spatial Vision, 10(4):433–436, 1997. ISSN 0169-1015. → pages 54[31] K. L. Briggman and W. Denk. Towards neural circuit reconstruction withvolume electron microscopy techniques. Current opinion in neurobiology,16(5):562–570, 2006. → pages 12[32] I. Brunig, S. Kaech, H. Brinkhaus, T. G. Oertner, and A. Matus. Influx ofextracellular calcium regulates actin-dependent morphological plasticity indendritic spines. Neuropharmacology, 47(5):669–676, 2004. → pages 24[33] G. Buzsaki. Large-scale recording of neuronal ensembles. Natureneuroscience, 7(5):446–451, 2004. → pages 9[34] P. J. Campagnola and C.-Y. Dong. Second harmonic generationmicroscopy: principles and applications to disease diagnosis. Laser &Photonics Reviews, 5(1):13–26, 2011. → pages 16[35] N. Caporale and Y. Dan. Spike timing-dependent plasticity: a hebbianlearning rule. Annu. Rev. Neurosci., 31:25–46, 2008. → pages 26[36] P. Caroni, F. Donato, and D. Muller. Structural plasticity upon learning:regulation and functions. Nature Reviews Neuroscience, 13(7):478–490,2012. → pages 2[37] F. Cesca, P. Baldelli, F. Valtorta, and F. Benfenati. The synapsins: keyactors of synapse function and plasticity. Progress in neurobiology, 91(4):313–348, 2010. → pages 91[38] M. Chalfie, Y. Tu, G. Euskirchen, W. W. Ward, and D. C. Prasher. Greenfluorescent protein as a marker for gene expression. Science, 263(5148):802–805, 1994. → pages 89[39] B.-C. Chen, W. R. Legant, K. Wang, L. Shao, D. E. Milkie, M. W.Davidson, C. Janetopoulos, X. S. Wu, J. A. Hammer, Z. Liu, and others.Lattice light-sheet microscopy: Imaging molecules to embryos at highspatiotemporal resolution. Science, 346(6208):1257998, 2014. → pages 13[40] B. E. Chen, M. Kondo, A. Garnier, F. L. Watson, R. Pettmann-Holgado,D. R. Lamar, and D. Schmucker. The molecular diversity of dscam isfunctionally required for neuronal wiring specificity in drosophila. Cell,144125(3):607–620, May 2006. ISSN 0092-8674.doi:10.1016/j.cell.2006.03.034. → pages 23, 25[41] B. L. Chen, D. H. Hall, and D. B. Chklovskii. Wiring optimization canrelate neuronal structure and function. Proceedings of the NationalAcademy of Sciences of the United States of America, 103(12):4723–4728,2006. → pages 3[42] J. L. Chen, O. A. Pfffli, F. F. Voigt, D. J. Margolis, and F. Helmchen. Onlinecorrection of licking-induced brain motion during two-photon imagingwith a tunable lens. The Journal of Physiology, 591(19):4689–4698, 2013.ISSN 1469-7793. doi:10.1113/jphysiol.2013.259804. → pages 19[43] S. X. Chen, P. K. Tari, K. She, and K. Haas. Neurexin-neuroligin celladhesion complexes contribute to synaptotropic dendritogenesis via growthstabilization mechanisms in vivo. Neuron, 67(6):967–983, 2010. → pages22, 25, 32, 41, 89[44] S. X. Chen, A. Cherry, P. K. Tari, K. Podgorski, Y. K. K. Kwong, andK. Haas. The transcription factor MEF2 directs developmental visuallydriven functional and structural metaplasticity. Cell, 151(1):41–55, 2012.→ pages 92, 98[45] T.-W. Chen, T. J. Wardill, Y. Sun, S. R. Pulver, S. L. Renninger, A. Baohan,E. R. Schreiter, R. A. Kerr, M. B. Orger, V. Jayaraman, and others.Ultrasensitive fluorescent proteins for imaging neuronal activity. Nature,499(7458):295–300, 2013. → pages 9, 11, 97, 104, 114[46] X. Chen, U. Leischner, N. L. Rochefort, I. Nelken, and A. Konnerth.Functional mapping of single spines in cortical neurons in vivo. Nature,475(7357):501–505, 2011. → pages 74, 80, 107, 114[47] H. J. Cheng, M. Nakamoto, A. D. Bergemann, and J. G. Flanagan.Complementary gradients in expression and binding of ELF-1 and mek4 indevelopment of the topographic retinotectal projection map. Cell, 82(3):371–381, Aug. 1995. ISSN 0092-8674. → pages 21[48] D. B. Chklovskii and A. A. Koulakov. Maps in the brain: what can welearn from them? Annu. Rev. Neurosci., 27:369–392, 2004. → pages 29[49] D. B. Chklovskii, T. Schikorski, and C. F. Stevens. Wiring optimization incortical circuits. Neuron, 34(3):341–347, 2002. → pages 29, 49145[50] S.-J. Chung, S. Zheng, T. Odani, L. Beverina, J. Fu, L. A. Padilha,A. Biesso, J. M. Hales, X. Zhan, K. Schmidt, A. Ye, E. Zojer, S. Barlow,D. J. Hagan, E. W. Van Stryland, Y. Yi, Z. Shuai, G. A. Pagani, J.-L. Brdas,J. W. Perry, and S. R. Marder. Extended squaraine dyes with largetwo-photon absorption cross-sections. Journal of the American ChemicalSociety, 128(45):14444–14445, Nov. 2006. ISSN 0002-7863.doi:10.1021/ja065556m. → pages 75[51] H. Cline and K. Haas. The regulation of dendritic arbor development andplasticity by glutamatergic synaptic input: a review of the synaptotrophichypothesis. The Journal of physiology, 586(6):1509–1517, 2008. → pages31, 32[52] P. D. Coleman and A. H. Riesen. Evironmental effects on cortical dendriticfields. i. rearing in the dark. Journal of Anatomy, 102(Pt 3):363, 1968. →pages 23[53] M. Constantine-Paton and P. Ferrari-Eastman. Pre-and postsynapticcorrelates of interocular competition and segregation in the frog. Journal ofComparative Neurology, 255(2):178–195, 1987. → pages 24[54] P. Coupe, M. Munz, J. V. Manjn, E. S. Ruthazer, and D. Louis Collins. ACANDLE for a deeper in vivo insight. Medical image analysis, 16(4):849–864, 2012. → pages 120[55] M. C. Crair. Neuronal activity during development: permissive orinstructive? Current opinion in neurobiology, 9(1):88–93, 1999. → pages97[56] M. Dahan, S. Lvi, C. Luccardini, P. Rostaing, B. Riveau, and A. Triller.Diffusion dynamics of glycine receptors revealed by single-quantum dottracking. Science (New York, N.Y.), 302(5644):442–445, Oct. 2003. ISSN1095-9203. doi:10.1126/science.1088525. → pages 74[57] M. B. Dalva, A. Ghosh, and C. J. Shatz. Independent control of dendriticand axonal form in the developing lateral geniculate nucleus. The Journalof neuroscience, 14(6):3588–3602, 1994. → pages 24[58] N. W. Daw, B. Gordon, K. D. Fox, H. J. Flavin, J. D. Kirsch, C. J. Beaver,Q. Ji, S. N. Reid, and D. Czepita. Injection of MK-801 affects oculardominance shifts more than visual activity. Journal of Neurophysiology, 81(1):204–215, Jan. 1999. ISSN 0022-3077. → pages 42, 51146[59] K. E. Deeg, I. B. Sears, and C. D. Aizenman. Development of multisensoryconvergence in the xenopus optic tectum. Journal of neurophysiology, 102(6):3392–3404, 2009. → pages 21, 28[60] W. Denk and K. Svoboda. Photon upmanship: why multiphoton imaging ismore than a gimmick. Neuron, 18(3):351–357, Mar. 1997. ISSN0896-6273. → pages 15[61] W. Denk, J. H. Strickler, and W. W. Webb. Two-photon laser scanningfluorescence microscopy. Science (New York, N.Y.), 248(4951):73–76, Apr.1990. ISSN 0036-8075. → pages 15, 73[62] W. Denk, K. R. Delaney, A. Gelperin, D. Kleinfeld, B. W. Strowbridge,D. W. Tank, and R. Yuste. Anatomical and functional imaging of neuronsusing 2-photon laser scanning microscopy. Journal of NeuroscienceMethods, 54(2):151–162, Oct. 1994. ISSN 0165-0270. → pages 74[63] B. Dick and G. Hohlneicher. Importance of initial and final states asintermediate states in two-photon spectroscopy of polar molecules. TheJournal of Chemical Physics, 76(12):5755–5760, 1982. → pages 75[64] S. Dimitrova, A. Reissaus, and G. Tavosanis. Slit and robo regulatedendrite branching and elongation of space-filling neurons in drosophila.Developmental biology, 324(1):18–30, 2008. → pages 23[65] W. Dong and C. D. Aizenman. A competition-based mechanism mediatesdevelopmental refinement of tectal neuron receptive fields. The Journal ofNeuroscience: The Official Journal of the Society for Neuroscience, 32(47):16872–16879, Nov. 2012. ISSN 1529-2401.doi:10.1523/JNEUROSCI.2372-12.2012. → pages 24[66] W. Dong, R. H. Lee, H. Xu, S. Yang, K. G. Pratt, V. Cao, Y.-K. Song,A. Nurmikko, and C. D. Aizenman. Visual avoidance in xenopus tadpolesis correlated with the maturation of visual responses in the optic tectum.Journal of Neurophysiology, 101(2):803–815, Feb. 2009. ISSN 0022-3077.doi:10.1152/jn.90848.2008. → pages 32[67] E. R. Dowski Jr and W. T. Cathey. Extended depth of field throughwave-front coding. Applied Optics, 34(11):1859–1866, 1995. → pages 17[68] S. Druckmann, L. Feng, B. Lee, C. Yook, T. Zhao, J. C. Magee, and J. Kim.Structured synaptic connectivity between hippocampal regions. Neuron, 81(3):629–640, 2014. → pages 104147[69] D. Dunfield and K. Haas. Metaplasticity governs natural experience-drivenplasticity of nascent embryonic brain circuits. Neuron, 64(2):240–250,2009. → pages 25, 28, 29, 31, 32, 33, 34, 39, 41, 49, 51, 52, 53, 92, 98, 99[70] M. Eigen and R. Rigler. Sorting single molecules: application todiagnostics and evolutionary biotechnology. Proceedings of the NationalAcademy of Sciences of the United States of America, 91(13):5740–5747,June 1994. ISSN 0027-8424. → pages 74[71] F. Engert, H. W. Tao, L. I. Zhang, and M.-m. Poo. Moving visual stimulirapidly induce direction sensitivity of developing tectal neurons. Nature,419(6906):470–475, 2002. → pages 26, 28, 31, 32, 39, 42, 51, 109[72] A. L. Fairhall, G. D. Lewen, W. Bialek, and R. R. d. R. van Steveninck.Efficiency and ambiguity in an adaptive neural code. Nature, 412(6849):787–792, 2001. → pages 3[73] J. Falk, J. Drinjakovic, K. Leung, A. Dwivedy, A. G. Regan, M. Piper, andC. E. Holt. Electroporation of cDNA/morpholinos to targeted areas ofembryonic CNS in xenopus. BMC Developmental Biology, 7(1):107, 2007.ISSN 1471-213X. doi:10.1186/1471-213X-7-107. → pages 21[74] B. J. Farley, H. Yu, D. Z. Jin, and M. Sur. Alteration of visual input resultsin a coordinated reorganization of multiple visual cortex maps. The Journalof Neuroscience: The Official Journal of the Society for Neuroscience, 27(38):10299–10310, Sept. 2007. ISSN 1529-2401.doi:10.1523/JNEUROSCI.2257-07.2007. → pages 45[75] M. Fill and J. A. Copello. Ryanodine receptor calcium release channels.Physiological Reviews, 82(4):893–922, Oct. 2002. ISSN 0031-9333.doi:10.1152/physrev.00013.2002. → pages 7[76] C. C. Fink, K.-U. Bayer, J. W. Myers, J. E. Ferrell Jr, H. Schulman, andT. Meyer. Selective regulation of neurite extension and synapse formationby the but not the isoform of CaMKII. Neuron, 39(2):283–297, 2003. →pages 24[77] J. G. Flanagan. Neural map specification by gradients. Current opinion inneurobiology, 16(1):59–66, 2006. → pages 20, 21[78] U. Frey and R. G. Morris. Synaptic tagging and long-term potentiation.Nature, 385(6616):533–536, 1997. → pages 110148[79] N. A. Frost, H. Shroff, H. Kong, E. Betzig, and T. A. Blanpied.Single-molecule discrimination of discrete perisynaptic and distributedsites of actin filament assembly within dendritic spines. Neuron, 67(1):86–99, 2010. → pages 14[80] M. Fu, X. Yu, J. Lu, and Y. Zuo. Repetitive motor learning inducescoordinated formation of clustered dendritic spines in vivo. Nature, 483(7387):92–95, 2012. → pages 97, 98, 114[81] Y. Fukazawa, Y. Saitoh, F. Ozawa, Y. Ohta, K. Mizuno, and K. Inokuchi.Hippocampal LTP is accompanied by enhanced f-actin content within thedendritic spine that is essential for late LTP maintenance in vivo. Neuron,38(3):447–460, 2003. → pages 103[82] C. R. Gallistel and L. D. Matzel. The neuroscience of learning: beyond thehebbian synapse. Annual review of psychology, 64:169–200, 2013. →pages 2[83] T. J. Gawne and B. J. Richmond. How independent are the messagescarried by adjacent inferior temporal cortical neurons? The Journal ofneuroscience, 13(7):2758–2771, 1993. → pages 32, 34, 35, 36, 38, 47, 109[84] R. M. Gaze, M. J. Keating, and S. H. Chung. The evolution of theretinotectal map during development in xenopus. Proceedings of the RoyalSociety of London. Series B, Biological Sciences, pages 301–330, 1974. →pages 25, 98[85] R. Gens and P. Domingos. Discriminative learning of sum-productnetworks. In Advances in Neural Information Processing Systems, pages3248–3256, 2012. → pages 4[86] K. B. Godfrey, S. J. Eglen, and N. V. Swindale. A multi-component modelof the developing retinocollicular pathway incorporating axonal andsynaptic growth. PLoS computational biology, 5(12):e1000600, 2009. →pages 25[87] C. S. Goodman and C. J. Shatz. Developmental mechanisms that generateprecise patterns of neuronal connectivity. Cell, 72 Suppl:77–98, Jan. 1993.ISSN 0092-8674. → pages 31[88] L. A. Goodman and P. G. Model. Eliminating afferent impulse activitydoes not alter the dendritic branching of the amphibian mauthner cell.Journal of neurobiology, 21(2):283–294, 1990. → pages 24149[89] A. B. A. Graf, A. Kohn, M. Jazayeri, and J. A. Movshon. Decoding theactivity of neuronal populations in macaque primary visual cortex. NatureNeuroscience, 14(2):239–245, Feb. 2011. ISSN 1546-1726.doi:10.1038/nn.2733. → pages 31, 38[90] W. T. Greenough, J. E. Black, and C. S. Wallace. Experience and braindevelopment. Child development, pages 539–559, 1987. → pages 22[91] B. F. Grewe, D. Langer, H. Kasper, B. M. Kampa, and F. Helmchen.High-speed in vivo calcium imaging reveals neuronal network activity withnear-millisecond precision. Nature Methods, 7(5):399–405, May 2010.ISSN 1548-7091. doi:10.1038/nmeth.1453. → pages 90[92] D. A. Gutnisky and V. Dragoi. Adaptive coding of visual information inneural populations. Nature, 452(7184):220–224, Mar. 2008. ISSN1476-4687. doi:10.1038/nature06563. → pages 32, 36[93] K. Haas, W.-C. Sin, A. Javaherian, Z. Li, and H. T. Cline. Single-cellelectroporationfor gene transfer in vivo. Neuron, 29(3):583–591, 2001. →pages 78, 99[94] K. Haas, J. Li, and H. T. Cline. AMPA receptors regulateexperience-dependent dendritic arbor growth invivo. Proceedings of theNational Academy of Sciences, 103(32):12127–12131, 2006. → pages 24,91[95] D. H. Hall and R. L. Russell. The posterior nervous system of thenematode caenorhabditis elegans: serial reconstruction of identifiedneurons and complete pattern of synaptic interactions. The Journal ofneuroscience, 11(1):1–22, 1991. → pages 12[96] J.-H. Han, S. A. Kushner, A. P. Yiu, C. J. Cole, A. Matynia, R. A. Brown,R. L. Neve, J. F. Guzowski, A. J. Silva, and S. A. Josselyn. Neuronalcompetition and selection during memory formation. Science (New York,N.Y.), 316(5823):457–460, Apr. 2007. ISSN 1095-9203.doi:10.1126/science.1139438. → pages 47[97] R. M. Harland and R. M. Grainger. Xenopus research: metamorphosed bygenetics and genomics. Trends in genetics: TIG, 27(12):507–515, Dec.2011. ISSN 0168-9525. doi:10.1016/j.tig.2011.08.003. → pages 20[98] M. T. Harnett, J. K. Makara, N. Spruston, W. L. Kath, and J. C. Magee.Synaptic amplification by dendritic spines enhances input cooperativity.150Nature, 491(7425):599–602, Nov. 2012. ISSN 0028-0836.doi:10.1038/nature11554. → pages 10[99] C. D. Harvey and K. Svoboda. Locally dynamic synaptic learning rules inpyramidal neuron dendrites. Nature, 450(7173):1195–1200, Dec. 2007.ISSN 1476-4687. doi:10.1038/nature06416. → pages 11, 110[100] C. D. Harvey, R. Yasuda, H. Zhong, and K. Svoboda. The spread of rasactivity triggered by activation of a single dendritic spine. Science, 321(5885):136–140, 2008. → pages 109, 110[101] D. O. Hebb. The organization of behavior; a neuropsychological theory.1949. → pages 2[102] S. W. Hell and J. Wichmann. Breaking the diffraction resolution limit bystimulated emission: stimulated-emission-depletion fluorescencemicroscopy. Optics letters, 19(11):780–782, 1994. → pages 13[103] F. Helmchen and W. Denk. Deep tissue two-photon microscopy. Naturemethods, 2(12):932–940, 2005. → pages 74[104] M. Helmstaedter, K. L. Briggman, S. C. Turaga, V. Jain, H. S. Seung, andW. Denk. Connectomic reconstruction of the inner plexiform layer in themouse retina. Nature, 500(7461):168–174, 2013. → pages 12[105] S. T. Hess, T. P. K. Girirajan, and M. D. Mason. Ultra-high resolutionimaging by fluorescence photoactivation localization microscopy.Biophysical Journal, 91(11):4258–4272, Dec. 2006. ISSN 0006-3495.doi:10.1529/biophysj.106.091116. → pages 14[106] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly,A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath, and others. Deep neuralnetworks for acoustic modeling in speech recognition: The shared views offour research groups. Signal Processing Magazine, IEEE, 29(6):82–97,2012. → pages 4[107] J. C. Hocking, C. L. Hehr, G. E. Bertolesi, J. Y. Wu, and S. McFarlane.Distinct roles for robo2 in the regulation of axon and dendrite growth byretinal ganglion cells. Mechanisms of development, 127(1):36–48, 2010.→ pages 23[108] A. L. Hodgkin and A. F. Huxley. A quantitative description of membranecurrent and its application to conduction and excitation in nerve. TheJournal of physiology, 117(4):500, 1952. → pages 6151[109] R. M. Hoffman. Visualization of GFP-expressing tumors and metastasis invivo. BioTechniques, 30(5):1016–1022, 1024–1026, May 2001. ISSN0736-6205. → pages 89[110] C. E. Holt and W. A. Harris. Order in the initial retinotectal map inxenopus: a new technique for labelling growing nerve fibres. 1983. →pages 25[111] J. J. Hopfield. Neural networks and physical systems with emergentcollective computational abilities. Proceedings of the national academy ofsciences, 79(8):2554–2558, 1982. → pages 4[112] H. W. Horch and L. C. Katz. BDNF release from single cells elicits localdendritic growth in nearby neurons. Nature Neuroscience, 5(11):1177–1184, Nov. 2002. ISSN 1097-6256. doi:10.1038/nn927. → pages 23[113] H. W. Horch, A. Krttgen, S. D. Portbury, and L. C. Katz. Destabilization ofcortical dendrites and spines by BDNF. Neuron, 23(2):353–364, June1999. ISSN 0896-6273. → pages 23[114] N. G. Horton, K. Wang, D. Kobat, C. G. Clark, F. W. Wise, C. B. Schaffer,and C. Xu. In vivo three-photon microscopy of subcortical structures withinan intact mouse brain. Nature photonics, 7(3):205–209, 2013. → pages 15[115] S. Hossain, D. Sesath Hewapathirane, and K. Haas. Dynamicmorphometrics reveals contributions of dendritic growth cones andfilopodia to dendritogenesis in the intact and awake embryonic brain.Developmental neurobiology, 72(4):615–627, 2012. → pages 21, 22, 120[116] B. Hu, A. M. Nikolakopoulou, and S. Cohen-Cory. BDNF stabilizessynapses and maintains the structural complexity of optic axons in vivo.Development, 132(19):4285–4298, 2005. → pages 23[117] J. Y. Hua and S. J. Smith. Neural activity and the dynamics of centralnervous system development. Nature Neuroscience, 7(4):327–332, Apr.2004. ISSN 1097-6256. doi:10.1038/nn1218. → pages 23[118] D. H. Hubel and T. N. Wiesel. Receptive fields, binocular interaction andfunctional architecture in the cat’s visual cortex. The Journal of physiology,160(1):106, 1962. → pages 7, 97[119] D. H. Hubel and T. N. Wiesel. The period of susceptibility to thephysiological effects of unilateral eye closure in kittens. The Journal ofPhysiology, 206(2):419–436, Feb. 1970. ISSN 0022-3751. → pages 31152[120] J. Huxter, N. Burgess, and J. O’Keefe. Independent rate and temporalcoding in hippocampal pyramidal cells. Nature, 425(6960):828–832, Oct.2003. ISSN 0028-0836. doi:10.1038/nature02058. → pages 8[121] V. Iyer, T. M. Hoogland, and P. Saggau. Fast functional imaging of singleneurons using random-access multiphoton (RAMP) microscopy. Journal ofneurophysiology, 95(1):535–545, 2006. → pages 94, 98[122] V. Jain, H. S. Seung, and S. C. Turaga. Machines that learn to segmentimages: a crucial technology for connectomics. Current opinion inneurobiology, 20(5):653–666, 2010. → pages 12[123] H. Jia, N. L. Rochefort, X. Chen, and A. Konnerth. Dendritic organizationof sensory input to cortical neurons in vivo. Nature, 464(7293):1307–1312,2010. → pages 10, 18, 107, 114[124] H. Jia, Z. Varga, B. Sakmann, and A. Konnerth. Linear integration of spineca2+ signals in layer 4 cortical neurons in vivo. Proceedings of theNational Academy of Sciences, 111(25):9277–9282, 2014. → pages 7, 10,97, 107[125] B. Judkewitz, M. Rizzi, K. Kitamura, and M. Husser. Targeted single-cellelectroporation of mammalian neurons in vivo. Nature Protocols, 4(6):862–869, 2009. ISSN 1750-2799. doi:10.1038/nprot.2009.56. → pages 80[126] E. Jurrus, M. Hardy, T. Tasdizen, P. T. Fletcher, P. Koshevoy, C.-B. Chien,W. Denk, and R. Whitaker. Axon tracking in serial block-face scanningelectron microscopy. Medical image analysis, 13(1):180–188, 2009. →pages 12[127] E. R. Kandel. The molecular biology of memory: cAMP, PKA, CRE,CREB-1, CREB-2, and CPEB. Mol Brain, 5(1):14, 2012. → pages 2[128] T. Keck, V. Scheuss, R. I. Jacobsen, C. J. Wierenga, U. T. Eysel,T. Bonhoeffer, and M. Hbener. Loss of sensory input causes rapidstructural changes of inhibitory neurons in adult mouse visual cortex.Neuron, 71(5):869–882, 2011. → pages 24, 98[129] S. Kishore and J. R. Fetcho. Homeostatic regulation of dendritic dynamicsin a motor map in vivo. Nature communications, 4, 2013. → pages 93[130] K. Kitamura, B. Judkewitz, M. Kano, W. Denk, and M. Husser. Targetedpatch-clamp recordings and single-cell electroporation of unlabeledneurons in vivo. Nature methods, 5(1):61–67, 2007. → pages 99, 119153[131] T. Kleindienst, J. Winnubst, C. Roth-Alpermann, T. Bonhoeffer, andC. Lohmann. Activity-dependent clustering of functional synaptic inputson developing hippocampal dendrites. Neuron, 72(6):1012–1024, Dec.2011. ISSN 1097-4199. doi:10.1016/j.neuron.2011.10.015. → pages 11,97, 107, 110, 114[132] G. Knott, H. Marchman, D. Wall, and B. Lich. Serial section scanningelectron microscopy of adult brain tissue using focused ion beam milling.The Journal of Neuroscience, 28(12):2959–2964, 2008. → pages 12[133] S. B. Kodandaramaiah, G. T. Franzesi, B. Y. Chow, E. S. Boyden, and C. R.Forest. Automated whole-cell patch-clamp electrophysiology of neurons invivo. Nature methods, 9(6):585–587, 2012. → pages 9[134] A. Kohn and M. A. Smith. Stimulus dependence of neuronal correlation inprimary visual cortex of the macaque. The Journal of Neuroscience: TheOfficial Journal of the Society for Neuroscience, 25(14):3661–3673, Apr.2005. ISSN 1529-2401. doi:10.1523/JNEUROSCI.5106-04.2005. →pages 32, 36[135] T. Komiyama, T. R. Sato, D. H. OConnor, Y.-X. Zhang, D. Huber, B. M.Hooks, M. Gabitto, and K. Svoboda. Learning-related fine-scale specificityimaged in motor cortex circuits of behaving mice. Nature, 464(7292):1182–1186, 2010. → pages 26, 29, 32, 33, 39, 49[136] S. Konur and A. Ghosh. Calcium signaling and the control of dendriticdevelopment. Neuron, 46(3):401–405, May 2005. ISSN 0896-6273.doi:10.1016/j.neuron.2005.04.022. → pages 23[137] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification withdeep convolutional neural networks. In Advances in neural informationprocessing systems, pages 1097–1105, 2012. → pages 4[138] V. A. Kulkarni and B. L. Firestein. The dendritic tree and brain disorders.Molecular and Cellular Neuroscience, 50(1):10–20, 2012. → pages 91[139] S. Kwok, C. Lee, S. A. Snchez, T. L. Hazlett, E. Gratton, and Y. Hayashi.Genetically encoded probe for fluorescence lifetime imaging of CaMKIIactivity. Biochemical and biophysical research communications, 369(2):519–525, 2008. → pages 89[140] H.-B. Kwon and B. L. Sabatini. Glutamate induces de novo growth offunctional spines in developing cortex. Nature, 474(7349):100–104, 2011.→ pages 24, 110154[141] C. S. W. Lai, T. F. Franke, and W.-B. Gan. Opposite effects of fearconditioning and extinction on dendritic spine remodelling. Nature, 483(7387):87–91, 2012. → pages 112[142] D. R. Larson, W. R. Zipfel, R. M. Williams, S. W. Clark, M. P. Bruchez,F. W. Wise, and W. W. Webb. Water-soluble quantum dots for multiphotonfluorescence imaging in vivo. Science (New York, N.Y.), 300(5624):1434–1436, May 2003. ISSN 1095-9203. doi:10.1126/science.1083780.→ pages 74[143] P. E. Latham and S. Nirenberg. Synergy, redundancy, and independence inpopulation codes, revisited. The Journal of Neuroscience: The OfficialJournal of the Society for Neuroscience, 25(21):5195–5206, May 2005.ISSN 1529-2401. doi:10.1523/JNEUROSCI.5319-04.2005. → pages 36[144] P.-m. Lau, R. S. Zucker, and D. Bentley. Induction of filopodia by directlocal elevation of intracellular calcium ion concentration. The Journal ofcell biology, 145(6):1265–1276, 1999. → pages 110[145] S. B. Laughlin, R. R. de Ruyter van Steveninck, and J. C. Anderson. Themetabolic cost of neural information. Nature Neuroscience, 1(1):36–41,May 1998. ISSN 1097-6256. doi:10.1038/236. → pages 49[146] M. Lavzin, S. Rapoport, A. Polsky, L. Garion, and J. Schiller. Nonlineardendritic processing determines angular tuning of barrel cortex neurons invivo. Nature, 490(7420):397–401, Oct. 2012. ISSN 0028-0836.doi:10.1038/nature11451. → pages 11, 109[147] G. Y. Lazar. The development of the optic tectum in xenopus laevis: a golgistudy. Journal of anatomy, 116(Pt 3):347, 1973. → pages 99, 116, 119[148] J. D. Lechleiter, D.-T. Lin, and I. Sieneart. Multi-photon laser scanningmicroscopy using an acoustic optical deflector. Biophysical journal, 83(4):2292–2299, 2002. → pages 122[149] Y. LeCun and Y. Bengio. Convolutional networks for images, speech, andtime series. The handbook of brain theory and neural networks, 3361,1995. → pages 4[150] S.-J. R. Lee, Y. Escobedo-Lozoya, E. M. Szatmari, and R. Yasuda.Activation of CaMKII in single dendritic spines during long-termpotentiation. Nature, 458(7236):299–304, 2009. → pages 110155[151] J. Li, A. Erisir, and H. Cline. In vivo time-lapse imaging and serial sectionelectron microscopy reveal developmental synaptic rearrangements.Neuron, 69(2):273–286, 2011. → pages 98, 104, 107, 112[152] Y. Li, S. D. Van Hooser, M. Mazurek, L. E. White, and D. Fitzpatrick.Experience with moving visual stimuli drives the early development ofcortical direction selectivity. Nature, 456(7224):952–956, 2008. → pages26, 29, 31, 32, 33, 43, 49[153] C.-C. Lien, Y. Mu, M. Vargas-Caballero, and M.-m. Poo. Visualstimuli-induced LTD of GABAergic synapses mediated by presynapticNMDA receptors. Nature Neuroscience, 9(3):372–380, Mar. 2006. ISSN1097-6256. doi:10.1038/nn1649. → pages 41[154] Y.-C. Lin and A. J. Koleske. Mechanisms of synapse and dendritemaintenance and their disruption in psychiatric and neurodegenerativedisorders. Annual review of neuroscience, 33:349, 2010. → pages 24[155] X. Liu. Role of PKMz in morphological and synaptic development of optictectal neurons in xenopus laevis tadpoles in vivo. 2010. → pages 25[156] C. Lohmann and T. Bonhoeffer. A role for local calcium signaling in rapidsynaptic partner selection by dendritic filopodia. Neuron, 59(2):253–260,2008. → pages 110[157] C. Lohmann, K. L. Myhr, and R. O. Wong. Transmitter-evoked localcalcium release stabilizes developing dendrites. Nature, 418(6894):177–181, 2002. → pages 24, 97[158] C. Lohmann, A. Finski, and T. Bonhoeffer. Local calcium transientsregulate the spontaneous motility of dendritic filopodia. Natureneuroscience, 8(3):305–312, 2005. → pages 24, 97, 109, 110[159] L. Luo and D. D. O’Leary. Axon retraction and degeneration indevelopment and disease. Annu. Rev. Neurosci., 28:127–156, 2005. →pages 22[160] C. Luscher and R. C. Malenka. NMDA receptor-dependent long-termpotentiation and long-term depression (LTP/LTD). Cold Spring Harborperspectives in biology, 4(6):a005710, 2012. → pages 7, 103[161] A. Majewska, G. Yiu, and R. Yuste. A custom-made two-photonmicroscope and deconvolution system. Pflgers Archiv: European Journal156of Physiology, 441(2-3):398–408, Dec. 2000. ISSN 0031-6768. → pages90[162] H. Makino and R. Malinow. Compartmentalized versus global synapticplasticity on dendrites controlled by experience. Neuron, 72(6):1001–1011,Dec. 2011. ISSN 1097-4199. doi:10.1016/j.neuron.2011.09.036. → pages11, 110[163] R. C. Malenka and M. F. Bear. LTP and LTD: an embarrassment of riches.Neuron, 44(1):5–21, Sept. 2004. ISSN 0896-6273.doi:10.1016/j.neuron.2004.09.012. → pages 51[164] R. C. Malenka, J. A. Kauer, R. S. Zucker, and R. A. Nicoll. Postsynapticcalcium is sufficient for potentiation of hippocampal synaptic transmission.Science, 242(4875):81–84, 1988. → pages 110[165] M. Maletic-Savatic, R. Malinow, and K. Svoboda. Rapid dendriticmorphogenesis in CA1 hippocampal dendrites induced by synaptic activity.Science, 283(5409):1923–1927, 1999. → pages 103, 110[166] S. Marshak, A. M. Nikolakopoulou, R. Dirks, G. J. Martens, andS. Cohen-Cory. Cell-autonomous TrkB signaling in presynaptic retinalganglion cells mediates axon arbor growth and synapse maturation duringthe establishment of retinotectal synaptic connectivity. The Journal ofNeuroscience, 27(10):2444–2456, Mar. 2007. ISSN 0270-6474,1529-2401. doi:10.1523/JNEUROSCI.4434-06.2007. → pages 23[167] J. S. Marvin, B. G. Borghuis, L. Tian, J. Cichon, M. T. Harnett,J. Akerboom, A. Gordus, S. L. Renninger, T.-W. Chen, C. I. Bargmann, andothers. An optimized fluorescent probe for visualizing glutamateneurotransmission. Nature methods, 10(2):162–170, 2013. → pages 9, 139[168] M. Matsuzaki, N. Honkura, G. C. Ellis-Davies, and H. Kasai. Structuralbasis of long-term potentiation in single dendritic spines. Nature, 429(6993):761–766, 2004. → pages 110[169] T. J. McBride, A. Rodriguez-Contreras, A. Trinh, R. Bailey, and W. M.DeBello. Learning drives differential clustering of axodendritic contacts inthe barn owl auditory system. The Journal of Neuroscience, 28(27):6960–6973, 2008. → pages 2, 97, 104, 114[170] J. T. McIlwain. Distributed spatial coding in the superior colliculus: areview. Visual neuroscience, 6(01):3–13, 1991. → pages 21157[171] I. L. Medintz, H. T. Uyeda, E. R. Goldman, and H. Mattoussi. Quantumdot bioconjugates for imaging, labelling and sensing. Nature Materials, 4(6):435–446, June 2005. ISSN 1476-1122. doi:10.1038/nmat1390. →pages 74[172] L. Mei and W.-C. Xiong. Neuregulin 1 in neural development, synapticplasticity and schizophrenia. Nature Reviews Neuroscience, 9(6):437–452,2008. → pages 91[173] X. Michalet, F. F. Pinaud, L. A. Bentolila, J. M. Tsay, S. Doose, J. J. Li,G. Sundaresan, A. M. Wu, S. S. Gambhir, and S. Weiss. Quantum dots forlive cells, in vivo imaging, and diagnostics. Science (New York, N.Y.), 307(5709):538–544, Jan. 2005. ISSN 1095-9203.doi:10.1126/science.1104274. → pages 89[174] K. D. Micheva and S. J. Smith. Array tomography: a new tool for imagingthe molecular architecture and ultrastructure of neural circuits. Neuron, 55(1):25–36, 2007. → pages 13[175] K. D. Micheva, B. Busse, N. C. Weiler, N. O’Rourke, and S. J. Smith.Single-synapse analysis of a diverse synapse population: proteomicimaging methods and markers. Neuron, 68(4):639–653, 2010. → pages 14[176] J. P. Miller, G. A. Jacobs, and F. E. Theunissen. Representation of sensoryinformation in the cricket cercal sensory system. i. response properties ofthe primary interneurons. Journal of Neurophysiology, 66(5):1680–1689,Nov. 1991. ISSN 0022-3077. → pages 7, 32, 47[177] M. Minsky. Confocal scanning microscope. Rapport technique, Patent, 3,1955. → pages 14[178] M. Minsky and S. Papert. Perceptron: an introduction to computationalgeometry. The MIT Press, Cambridge, expanded edition, 19:88, 1969. →pages 4[179] F. Montani, A. Kohn, M. A. Smith, and S. R. Schultz. The role ofcorrelations in direction and contrast coding in the primary visual cortex.The Journal of Neuroscience: The Official Journal of the Society forNeuroscience, 27(9):2338–2348, Feb. 2007. ISSN 1529-2401.doi:10.1523/JNEUROSCI.3417-06.2007. → pages 38[180] E. I. Moser, E. Kropff, and M.-B. Moser. Place cells, grid cells, and thebrain’s spatial representation system. Neuroscience, 31(1):69, 2008. →pages 8158[181] J. R. Mourant, J. P. Freyer, A. H. Hielscher, A. A. Eick, D. Shen, and T. M.Johnson. Mechanisms of light scattering from biological cells relevant tononinvasive optical-tissue diagnostics. Applied Optics, 37(16):3586–3593,1998. → pages 17[182] T. D. Mrsic-Flogel, S. B. Hofer, K. Ohki, R. C. Reid, T. Bonhoeffer, andM. Hubener. Homeostatic regulation of eye-specific responses in visualcortex during ocular dominance plasticity. Neuron, 54(6):961–972, 2007.→ pages 97[183] Y. Mu and M.-m. Poo. Spike timing-dependent LTP/LTD mediates visualexperience-dependent plasticity in a developing retinotectal system.Neuron, 50(1):115–125, 2006. → pages 98, 109, 110[184] M. Munz, D. Gobert, A. Schohl, J. Poqurusse, K. Podgorski, P. Spratt, andE. S. Ruthazer. Rapid hebbian axonal remodeling mediated by visualstimulation. Science, 344(6186):904–909, 2014. → pages 21, 98, 114[185] T. H. Murphy, J. M. Baraban, and W. G. Wier. Mapping miniature synapticcurrents to single synapses using calcium imaging reveals heterogeneity inpostsynaptic output. Neuron, 15(1):159–168, 1995. → pages 10[186] S. Nabavi, H. W. Kessels, S. Alfonso, J. Aow, R. Fox, and R. Malinow.Metabotropic NMDA receptor function is required for NMDAreceptor-dependent long-term depression. Proceedings of the NationalAcademy of Sciences, 110(10):4027–4032, 2013. → pages 7[187] U. V. Nagerl, N. Eberhorn, S. B. Cambridge, and T. Bonhoeffer.Bidirectional activity-dependent morphological plasticity in hippocampalneurons. Neuron, 44(5):759–767, 2004. → pages 103, 110[188] C. M. Niell and S. J. Smith. Functional imaging reveals rapid developmentof visual response properties in the zebrafish tectum. Neuron, 45(6):941–951, Mar. 2005. ISSN 0896-6273. doi:10.1016/j.neuron.2005.01.047.→ pages 22, 28, 32, 33, 43[189] C. M. Niell, M. P. Meyer, and S. J. Smith. In vivo imaging of synapseformation on a growing dendritic arbor. Nature neuroscience, 7(3):254–260, 2004. → pages 25, 98, 104, 112[190] K. Ohki, S. Chung, Y. H. Ch’ng, P. Kara, and R. C. Reid. Functionalimaging with cellular resolution reveals precise micro-architecture in visualcortex. Nature, 433(7026):597–603, Feb. 2005. ISSN 1476-4687.doi:10.1038/nature03274. → pages 33, 43, 49159[191] M. Oizumi, T. Ishii, K. Ishibashi, T. Hosoya, and M. Okada. Mismatcheddecoding in the brain. The Journal of Neuroscience: The Official Journalof the Society for Neuroscience, 30(13):4815–4826, Mar. 2010. ISSN1529-2401. doi:10.1523/JNEUROSCI.4360-09.2010. → pages 39, 51[192] K.-I. Okamoto, T. Nagai, A. Miyawaki, and Y. Hayashi. Rapid andpersistent modulation of actin dynamics regulates postsynapticreorganization underlying bidirectional plasticity. Nature neuroscience, 7(10):1104–1112, 2004. → pages 103[193] A. F. Oliveira and R. Yasuda. Neurofibromin is the major ras inactivator indendritic spines. The Journal of Neuroscience: The Official Journal of theSociety for Neuroscience, 34(3):776–783, Jan. 2014. ISSN 1529-2401.doi:10.1523/JNEUROSCI.3096-13.2014. → pages 103[194] A. M. Packer, L. E. Russell, H. W. Dalgleish, and M. Husser. Simultaneousall-optical manipulation and recording of neural circuit activity withcellular resolution in vivo. Nature methods, 2014. → pages 17[195] S. Panzeri, S. R. Schultz, A. Treves, and E. T. Rolls. Correlations and theencoding of information in the nervous system. Proceedings. BiologicalSciences / The Royal Society, 266(1423):1001–1012, May 1999. ISSN0962-8452. doi:10.1098/rspb.1999.0736. → pages 36, 38[196] I. P. Pavlov and W. H. Thompson. The work of the digestive glands.Charles Griffin, 1902. → pages 1[197] J. B. Pawley and B. R. Masters. Handbook of biological confocalmicroscopy, third edition. Journal of Biomedical Optics, 13:9902, 2008.ISSN 1083-3668. doi:10.1117/1.2911629. → pages 89[198] K. L. Perkins. Cell-attached voltage-clamp and current-clamp recordingand stimulation techniques in brain slices. Journal of NeuroscienceMethods, 154(1-2):1–18, June 2006. ISSN 0165-0270.doi:10.1016/j.jneumeth.2006.02.010. → pages 54[199] A. J. Peters, S. X. Chen, and T. Komiyama. Emergence of reproduciblespatiotemporal activity during motor learning. Nature, 2014. → pages 97,98[200] C. Petzold. The annotated Turing: a guided tour through Alan Turing’shistoric paper on computability and the Turing machine. Wiley Publishing,2008. → pages 3160[201] J. W. Pillow, L. Paninski, V. J. Uzzell, E. P. Simoncelli, and E. J.Chichilnisky. Prediction and decoding of retinal ganglion cell responseswith a probabilistic spiking model. The Journal of Neuroscience, 25(47):11003–11013, 2005. → pages 6[202] J. W. Pillow, J. Shlens, L. Paninski, A. Sher, A. M. Litke, E. J.Chichilnisky, and E. P. Simoncelli. Spatio-temporal correlations and visualsignalling in a complete neuronal population. Nature, 454(7207):995–999,2008. → pages 6, 8, 9, 32[203] S. Pinker. Language learnability and language development, with newcommentary by the author. Harvard University Press, 2009. → pages 1[204] K. Podgorski and K. Haas. Fast non-negative temporal deconvolution forlaser scanning microscopy. Journal of biophotonics, 6(2):153–162, 2013.→ pages iii, 88, 118, 123[205] K. Podgorski, D. Dunfield, and K. Haas. Functional clustering drivesencoding improvement in a developing brain network during awake visuallearning. PLoS biology, 10(1):e1001236, 2012. → pages iii, 74, 89, 97, 98,99, 117, 118[206] K. Podgorski, E. Terpetschnig, O. P. Klochko, O. M. Obukhova, andK. Haas. Ultra-bright and-stable red and near-infrared squarainefluorophores for in vivo two-photon imaging. PloS one, 7(12):e51980,2012. → pages iii[207] P. Poirazi and B. W. Mel. Impact of active dendrites and structuralplasticity on the memory capacity of neural tissue. Neuron, 29(3):779–796,2001. → pages 109[208] T. A. Pologruto, B. L. Sabatini, and K. Svoboda. ScanImage: flexiblesoftware for operating laser scanning microscopes. BiomedicalEngineering Online, 2:13, May 2003. ISSN 1475-925X.doi:10.1186/1475-925X-2-13. → pages 90[209] H. Poon and P. Domingos. Sum-product networks: A new deeparchitecture. In Computer Vision Workshops (ICCV Workshops), 2011IEEE International Conference on, pages 689–690. IEEE, 2011. → pages 4[210] Y. A. Povrozin, L. I. Markova, A. L. Tatarets, V. I. Sidorov, E. A.Terpetschnig, and L. D. Patsenker. Near-infrared, dual-ratiometricfluorescent label for measurement of pH. Analytical Biochemistry, 390(2):161136–140, July 2009. ISSN 1096-0309. doi:10.1016/j.ab.2009.03.047. →pages 75[211] K. G. Pratt, W. Dong, and C. D. Aizenman. Development and spiketiming-dependent plasticity of recurrent excitation in the xenopus optictectum. Nature Neuroscience, 11(4):467–475, Apr. 2008. ISSN 1097-6256.doi:10.1038/nn2076. → pages 32, 42, 51[212] D. o. P. W. U. S. F. Professor. Minds behind the Brain : A History of thePioneers and Their Discoveries: A History of the Pioneers and TheirDiscoveries. Oxford University Press, Mar. 2000. ISBN 9780198024682.→ pages 12[213] R. Quian Quiroga and S. Panzeri. Extracting information from neuronalpopulations: information theory and decoding approaches. Nature Reviews.Neuroscience, 10(3):173–185, Mar. 2009. ISSN 1471-0048.doi:10.1038/nrn2578. → pages 38[214] S. Quirin, D. S. Peterka, and R. Yuste. Instantaneous three-dimensionalsensing using spatial light modulator illumination with extended depth offield imaging. Optics express, 21(13):16007–16021, 2013. → pages 17[215] R. Q. Quiroga, L. Reddy, G. Kreiman, C. Koch, and I. Fried. Invariantvisual representation by single neurons in the human brain. Nature, 435(7045):1102–1107, 2005. → pages 7, 32, 33, 36[216] I. Rajan and H. T. Cline. Glutamate receptor activity is required for normaldevelopment of tectal cell dendrites in vivo. The Journal of neuroscience,18(19):7836–7846, 1998. → pages 24[217] I. Rajan, S. Witte, and H. T. Cline. NMDA receptor activity stabilizespresynaptic retinotectal axons and postsynaptic optic tectal cell dendrites invivo. Journal of Neurobiology, 38(3):357–368, Feb. 1999. ISSN0022-3034. → pages 41[218] G. D. Reddy, K. Kelleher, R. Fink, and P. Saggau. Three-dimensionalrandom access multiphoton microscopy for functional imaging of neuronalactivity. Nature neuroscience, 11(6):713–720, 2008. → pages 18, 98[219] T. A. Reh and M. Constantine-Paton. Eye-specific segregation requiresneural activity in three-eyed rana pipiens. The Journal of Neuroscience, 5(5):1132–1143, 1985. → pages 21162[220] C. S. Rex, C. F. Gavin, M. D. Rubio, E. A. Kramar, L. Y. Chen, Y. Jia, R. L.Huganir, N. Muzyczka, C. M. Gall, C. A. Miller, and others. Myosin IIbregulates actin dynamics during synaptic plasticity and memory formation.Neuron, 67(4):603–617, 2010. → pages 103[221] R. V. Riccio and M. A. Matthews. Effects of intraocular tetrodotoxin on thepostnatal development of the dorsal lateral geniculate nucleus of the rat: agolgi analysis. Journal of neuroscience research, 17(4):440–451, 1987. →pages 24[222] B. A. Richards, O. P. Voss, and C. J. Akerman. GABAergic circuits controlstimulus-instructed receptive field development in the optic tectum. Natureneuroscience, 13(9):1098–1106, 2010. → pages 29[223] T. F. Roberts, K. A. Tschida, M. E. Klein, and R. Mooney. Rapid spinestabilization and synaptic enhancement at the onset of behaviourallearning. Nature, 463(7283):948–952, 2010. → pages 97, 98[224] E. Robles and H. Baier. Assembly of synaptic laminae by axon guidancemolecules. Current opinion in neurobiology, 22(5):799–804, 2012. →pages 23[225] F. Rosenblatt. The perceptron: a probabilistic model for informationstorage and organization in the brain. Psychological review, 65(6):386,1958. → pages 3[226] R. Roy, S. Hohng, and T. Ha. A practical guide to single-molecule FRET.Nature Methods, 5(6):507–516, June 2008. ISSN 1548-7105.doi:10.1038/nmeth.1208. → pages 74[227] B. Rozsa, G. Katona, E. S. Vizi, Z. Varallyay, A. Saghy, L. Valenta,P. Maak, J. Fekete, A. Banyasz, and R. Szipocs. Random accessthree-dimensional two-photon microscopy. Applied optics, 46(10):1860–1865, 2007. → pages 94[228] D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internalrepresentations by error propagation. Technical report, DTIC Document,1985. → pages 4[229] M. J. Rust, M. Bates, and X. Zhuang. Sub-diffraction-limit imaging bystochastic optical reconstruction microscopy (STORM). Nature methods, 3(10):793–796, 2006. → pages 13163[230] E. S. Ruthazer, C. J. Akerman, and H. T. Cline. Control of axon branchdynamics by correlated activity in vivo. Science (New York, N.Y.), 301(5629):66–70, July 2003. ISSN 1095-9203. doi:10.1126/science.1082545.→ pages 41, 45, 49, 52[231] A. Sahay, K. N. Scobie, A. S. Hill, C. M. O’Carroll, M. A. Kheirbek, N. S.Burghardt, A. A. Fenton, A. Dranovsky, and R. Hen. Increasing adulthippocampal neurogenesis is sufficient to improve pattern separation.Nature, 472(7344):466–470, 2011. → pages 2[232] D. S. Sakaguchi and R. K. Murphey. Map formation in the developingxenopus retinotectal system: an examination of ganglion cell terminalarborizations. The Journal of neuroscience, 5(12):3228–3245, 1985. →pages 25, 32[233] M. Sakamoto, N. Ieki, G. Miyoshi, D. Mochimaru, H. Miyachi, T. Imura,M. Yamaguchi, G. Fishell, K. Mori, R. Kageyama, and others. Continuouspostnatal neurogenesis contributes to formation of the olfactory bulb neuralcircuits and flexible olfactory associative learning. The Journal ofNeuroscience, 34(17):5788–5799, 2014. → pages 2[234] B. Sakmann and E. Neher. Patch clamp techniques for studying ionicchannels in excitable membranes. Annual review of physiology, 46(1):455–472, 1984. → pages 8[235] J. M. Samonds, J. D. Allison, H. A. Brown, and A. B. Bonds. Cooperationbetween area 17 neuron pairs enhances fine discrimination of orientation.The Journal of Neuroscience: The Official Journal of the Society forNeuroscience, 23(6):2416–2425, Mar. 2003. ISSN 1529-2401. → pages 36[236] A. L. Sanchez, B. J. Matthews, M. M. Meynard, B. Hu, S. Javed, andS. Cohen-Cory. BDNF increases synapse density in dendrites ofdeveloping tectal neurons in vivo. Development, 133(13):2477–2486,2006. → pages 23[237] J. Sanders, K. Cowansage, K. Baumgrtel, and M. Mayford. Elimination ofdendritic spines with long-term memory is specific to active circuits. TheJournal of Neuroscience, 32(36):12570–12578, 2012. → pages 112[238] T. Saneyoshi, D. A. Fortin, and T. R. Soderling. Regulation of spine andsynapse formation by activity-dependent intracellular signaling pathways.Current opinion in neurobiology, 20(1):108–115, 2010. → pages 104164[239] M. C. Schmid, S. W. Mrowka, J. Turchi, R. C. Saunders, M. Wilke, A. J.Peters, Q. Y. Frank, and D. A. Leopold. Blindsight depends on the lateralgeniculate nucleus. Nature, 466(7304):373–377, 2010. → pages 21[240] E. Schneidman, W. Bialek, and M. J. Berry. Synergy, redundancy, andindependence in population codes. The Journal of Neuroscience: TheOfficial Journal of the Society for Neuroscience, 23(37):11539–11553,Dec. 2003. ISSN 1529-2401. → pages 32, 47[241] E. Schneidman, M. J. Berry, R. Segev, and W. Bialek. Weak pairwisecorrelations imply strongly correlated network states in a neuralpopulation. Nature, 440(7087):1007–1012, 2006. → pages 6, 36[242] M. N. Shadlen and W. T. Newsome. Noise, neural codes and corticalorganization. Current Opinion in Neurobiology, 4(4):569–579, Aug. 1994.ISSN 0959-4388. → pages 49[243] C. E. Shannon. The Mathematical Theory of Communication. University ofIllinois Press, Jan. 1949. → pages 3[244] S. C. Sharma, A. Memic, C. N. Rupasinghe, A.-C. E. Duc, and M. R.Spaller. T7 phage display as a method of peptide ligand discovery for PDZdomain proteins. Biopolymers, 92(3):183–193, 2009. ISSN 0006-3525.doi:10.1002/bip.21172. → pages 80, 84[245] T. O. Sharpee, H. Sugihara, A. V. Kurgansky, S. P. Rebrik, M. P. Stryker,and K. D. Miller. Adaptive filtering enhances information transmission invisual cortex. Nature, 439(7079):936–942, Feb. 2006. ISSN 1476-4687.doi:10.1038/nature04519. → pages 32[246] M. E. Sheffield and D. A. Dombeck. Calcium transient prevalence acrossthe dendritic arbour predicts place field properties. Nature, 2014. → pages109, 110[247] J. Shi and J. Malik. Normalized cuts and image segmentation. PatternAnalysis and Machine Intelligence, IEEE Transactions on, 22(8):888–905,2000. → pages 60[248] T. J. Siddiqui and A. M. Craig. Synaptic organizing complexes. Currentopinion in neurobiology, 21(1):132–143, 2011. → pages 25[249] E. P. Simoncelli and B. A. Olshausen. Natural image statistics and neuralrepresentation. Annual review of neuroscience, 24(1):1193–1216, 2001. →pages 3, 52165[250] W. C. Sin, K. Haas, E. S. Ruthazer, and H. T. Cline. Dendrite growthincreased by visual activity requires NMDA receptor and rho GTPases.Nature, 419(6906):475–480, 2002. → pages 20, 23, 24, 32, 41, 97, 98, 103[251] B. F. Skinner. Science and human behavior. Simon and Schuster, 1953. →pages 1[252] S. L. Smith, I. T. Smith, T. Branco, and M. Husser. Dendritic spikesenhance stimulus selectivity in cortical neurons in vivo. Nature, 503(7474):115–120, Nov. 2013. ISSN 0028-0836. doi:10.1038/nature12600. →pages 11[253] A. Sobczyk, V. Scheuss, and K. Svoboda. NMDA receptorsubunit-dependent [ca2+] signaling in individual hippocampal dendriticspines. The Journal of neuroscience, 25(26):6037–6046, 2005. → pages139[254] M. Soiza-Reilly and K. G. Commons. Quantitative analysis ofglutamatergic innervation of the mouse dorsal raphe nucleus using arraytomography. Journal of Comparative Neurology, 519(18):3802–3814,2011. → pages 14[255] S. Song, K. D. Miller, and L. F. Abbott. Competitive hebbian learningthrough spike-timing-dependent synaptic plasticity. Nature Neuroscience,3(9):919–926, Sept. 2000. ISSN 1097-6256. doi:10.1038/78829. → pages52[256] H. Spemann and H. Mangold. Uber induktion von embryonalanlagen durchimplantation artfremder organisatoren. Development Genes and Evolution,100(3):599–638, 1924. → pages 20[257] R. W. Sperry. Effect of 180 degree rotation of the retinal field onvisuomotor coordination. Journal of Experimental Zoology, 92(3):263–279, 1943. ISSN 1097-010X. doi:10.1002/jez.1400920303. → pages21[258] R. W. Sperry. Chemoaffinity in the ororder growth of nerve fiber patternsand connections. Proceedings of the National Academy of Sciences of theUnited States of America, 50:703–710, Oct. 1963. ISSN 0027-8424. →pages 32[259] J. M. Squirrell, D. L. Wokosin, J. G. White, and B. D. Bavister. Long-termtwo-photon fluorescence imaging of mammalian embryos without166compromising viability. Nature Biotechnology, 17(8):763–767, Aug. 1999.ISSN 1087-0156. doi:10.1038/11698. → pages 74[260] S. Sreejith, P. Carol, P. Chithra, and A. Ajayaghosh. Squaraine dyes: amine of molecular materials. Journal of Materials Chemistry, 18(3):264–274, 2008. → pages 75[261] M. I. Stefan, S. Pepke, S. Mihalas, T. Bartol, T. J. Sejnowski, and M. B.Kennedy. Multi-stage modeling of the kinetics of activation of CaMKII.Frontiers in Neuroinformatics, (00033), 2011. → pages 7[262] G. Stuart, N. Spruston, and M. Husser. Dendrites. Oxford University PressNew York, 2008. → pages 7, 11[263] G. Sumbre, A. Muto, H. Baier, and M.-m. Poo. Entrained rhythmicactivities of neuronal ensembles as perceptual memory of time interval.Nature, 456(7218):102–106, Nov. 2008. ISSN 1476-4687.doi:10.1038/nature07351. → pages 32[264] L. B. Sweeney, Y.-H. Chou, Z. Wu, W. Joo, T. Komiyama, C. J. Potter,A. L. Kolodkin, K. C. Garcia, and L. Luo. Secreted semaphorins fromdegenerating larval ORN axons direct adult projection neuron dendritetargeting. Neuron, 72(5):734–747, 2011. → pages 23[265] N. Takahashi, K. Kitamura, N. Matsuo, M. Mayford, M. Kano, N. Matsuki,and Y. Ikegaya. Locally synchronized synaptic inputs. Science, 335(6066):353–356, 2012. → pages 10[266] K. Tanaka, L. Khiroug, F. Santamaria, T. Doi, H. Ogasawara, G. C.Ellis-Davies, M. Kawato, and G. J. Augustine. Ca2+ requirements forcerebellar long-term synaptic depression: Role for a postsynaptic leakyintegrator. Neuron, 54(5):787–800, 2007. → pages 110[267] S. Tanaka, J. Ribot, K. Imamura, and T. Tani. Orientation-restrictedcontinuous visual exposure induces marked reorganization of orientationmaps in early life. NeuroImage, 30(2):462–477, Apr. 2006. ISSN1053-8119. doi:10.1016/j.neuroimage.2005.09.056. → pages 45[268] H. W. Tao and M.-m. Poo. Activity-dependent matching of excitatory andinhibitory inputs during refinement of visual receptive fields. Neuron, 45(6):829–836, Mar. 2005. ISSN 0896-6273.doi:10.1016/j.neuron.2005.01.046. → pages 22167[269] H. W. Tao, L. I. Zhang, F. Engert, and M.-m. Poo. Emergence of inputspecificity of ltp during development of retinotectal connections in vivo.Neuron, 31(4):569–580, 2001. → pages 98[270] E. Terpetschnig, H. Szmacinski, A. Ozinskas, and J. R. Lakowicz.Synthesis of squaraine-n-hydroxysuccinimide esters and their biologicalapplication as long-wavelength fluorescent labels. Analytical Biochemistry,217(2):197–204, Mar. 1994. ISSN 0003-2697.doi:10.1006/abio.1994.1109. → pages 75[271] P. Thevenaz, U. E. Ruttimann, and M. Unser. A pyramid approach tosubpixel registration based on intensity. IEEE transactions on imageprocessing: a publication of the IEEE Signal Processing Society, 7(1):27–41, 1998. ISSN 1057-7149. doi:10.1109/83.650848. → pages 55[272] G. Tkacik, E. Schneidman, I. I. Berry, J. Michael, and W. Bialek. Isingmodels for networks of real neurons. arXiv preprint q-bio/0611072, 2006.→ pages 60[273] G. Tkacik, J. S. Prentice, V. Balasubramanian, and E. Schneidman.Optimal population coding by noisy spiking neurons. Proceedings of theNational Academy of Sciences of the United States of America, 107(32):14419–14424, Aug. 2010. ISSN 1091-6490.doi:10.1073/pnas.1004906107. → pages 31, 32, 36, 38, 45, 52[274] J. T. Trachtenberg, B. E. Chen, G. W. Knott, G. Feng, J. R. Sanes,E. Welker, and K. Svoboda. Long-term in vivo imaging ofexperience-dependent synaptic plasticity in adult cortex. Nature, 420(6917):788–794, Dec. 2002. ISSN 0028-0836. doi:10.1038/nature01273.→ pages 74[275] A. Treves and S. Panzeri. The upward bias in measures of informationderived from limited data samples. Neural Computation, 7(2):399–407,1995. → pages 57[276] M. C. W. Van Rossum and R. G. Smith. Noise removal at the rod synapseof mammalian retina. Visual neuroscience, 15(05):809–821, 1998. →pages 7[277] Z. Varga, H. Jia, B. Sakmann, and A. Konnerth. Dendritic coding ofmultiple sensory inputs in single cortical neurons in vivo. Proceedings ofthe National Academy of Sciences, 108(37):15420–15425, 2011. → pages107, 114168[278] J. E. Vaughn, R. P. Barber, and T. J. Sims. Dendritic development andpreferential growth into synaptogenic fields: A quantitative study ofgolgi-impregnated spinal motor neurons. Synapse, 2(1):69–78, 1988. →pages 25[279] J. T. Vogelstein, B. O. Watson, A. M. Packer, R. Yuste, B. Jedynak, andL. Paninski. Spike inference from calcium imaging using sequential montecarlo methods. Biophysical Journal, 97(2):636–655, July 2009. ISSN1542-0086. doi:10.1016/j.bpj.2008.08.005. → pages 33, 56[280] F. R. Volkmar and W. T. Greenough. Rearing complexity affects branchingof dendrites in the visual cortex of the rat. Science, 176(4042):1445–1447,1972. → pages 23[281] C.-K. Wang, P. Macak, Y. Luo, and H. \AAgren. Effects of centers andsymmetry on two-photon absorption cross sections of organicchromophores. The Journal of Chemical Physics, 114(22):9813–9820,2001. → pages 16[282] K. Wang, D. E. Milkie, A. Saxena, P. Engerer, T. Misgeld, M. E. Bronner,J. Mumm, and E. Betzig. Rapid adaptive optical recovery of optimalresolution over large volumes. Nature methods, 2014. → pages 15[283] X. Wang, Y. Pang, G. Ku, X. Xie, G. Stoica, and L. V. Wang. Noninvasivelaser-induced photoacoustic tomography for structural and functional invivo imaging of the brain. Nature biotechnology, 21(7):803–806, 2003. →pages 15[284] S. Ward, N. Thomson, J. G. White, and S. Brenner. Electron microscopicalreconstruction of the anterior sensory anatomy of the nematodecaenorhabditis elegans. Journal of Comparative Neurology, 160(3):313–337, 1975. → pages 12[285] G. Wei, X. Xiao, S. Wang, J. D. Zimmerman, K. Sun, V. V. Diev, M. E.Thompson, and S. R. Forrest. Arylamine-based squaraine donors for use inorganic solar cells. Nano Letters, 11(10):4261–4264, Oct. 2011. ISSN1530-6992. doi:10.1021/nl2022515. → pages 75[286] T. N. Wiesel, D. H. Hubel, and others. Single-cell responses in striatecortex of kittens deprived of vision in one eye. J Neurophysiol, 26(6):1003–1017, 1963. → pages 23, 26169[287] W. T. Wong, B. E. Faulkner-Jones, J. R. Sanes, and R. O. Wong. Rapiddendritic remodeling in the developing retina: dependence onneurotransmission and reciprocal regulation by rac and rho. The Journal ofNeuroscience, 20(13):5024–5036, 2000. → pages 24, 91[288] G. Y. Wu, D. J. Zou, I. Rajan, and H. Cline. Dendritic dynamics in vivochange during neuronal maturation. The Journal of Neuroscience: TheOfficial Journal of the Society for Neuroscience, 19(11):4472–4483, June1999. ISSN 1529-2401. → pages 32[289] G.-Y. Wu, K. Deisseroth, and R. W. Tsien. Spaced stimuli stabilize MAPKpathway activation and its effects on dendritic morphology. Natureneuroscience, 4(2):151–158, 2001. → pages 24[290] N.-l. Xu, M. T. Harnett, S. R. Williams, D. Huber, D. H. OConnor,K. Svoboda, and J. C. Magee. Nonlinear dendritic integration of sensoryand motor input during an active sensing task. Nature, 492(7428):247–251,2012. → pages 8, 11[291] A. P. Yiu, V. Mercaldo, C. Yan, B. Richards, A. J. Rashid, H.-L. L. Hsiang,J. Pressey, V. Mahadevan, M. M. Tran, S. A. Kushner, and others. Neuronsare recruited to a memory trace based on relative neuronal excitabilityimmediately before training. Neuron, 83(3):722–735, 2014. → pages 97[292] R. Zalk, O. B. Clarke, A. des Georges, R. A. Grassucci, S. Reiken,F. Mancia, W. A. Hendrickson, J. Frank, and A. R. Marks. Structure of amammalian ryanodine receptor. Nature, 517(7532):44–49, Jan. 2015.ISSN 0028-0836. doi:10.1038/nature13950. → pages 7[293] S. Zhai, E. D. Ark, P. Parra-Bueno, and R. Yasuda. Long-distanceintegration of nuclear ERK signaling triggered by activation of a fewdendritic spines. Science, 342(6162):1107–1111, 2013. → pages 109[294] J. Zhang. Method of unconfounding orientation and direction tunings inneuronal response to moving bars and gratings. Journal of the OpticalSociety of America. A, Optics, Image Science, and Vision, 22(10):2246–2256, Oct. 2005. ISSN 1084-7529. → pages 57[295] L. I. Zhang, H. W. Tao, C. E. Holt, W. A. Harris, and M.-m. Poo. A criticalwindow for cooperation and competition among developing retinotectalsynapses. Nature, 395(6697):37–44, 1998. → pages 26, 28, 32, 41, 52170[296] L. I. Zhang, W. T. Hui-zhong, and M.-m. Poo. Visual input induceslong-term potentiation of developing retinotectal synapses. Natureneuroscience, 3(7):708–715, 2000. → pages 28[297] Y. Zhao, S. Araki, J. Wu, T. Teramoto, Y.-F. Chang, M. Nakano, A. S.Abdelfattah, M. Fujiwara, T. Ishihara, T. Nagai, and others. An expandedpalette of genetically encoded ca2+ indicators. Science, 333(6051):1888–1891, 2011. → pages 116[298] S. Zhong, D. Navaratnam, and J. Santos-Sacchi. A genetically-encodedYFP sensor with enhanced chloride sensitivity, photostability and reducedpH interference demonstrates augmented transmembrane chloridemovement by gerbil prestin (SLC26a5). PLoS ONE, 9(6):e99095, June2014. ISSN 1932-6203. doi:10.1371/journal.pone.0099095. → pages 9[299] Q. Zhou, H. W. Tao, and M.-m. Poo. Reversal and stabilization of synapticmodifications in a developing visual system. Science, 300(5627):1953–1957, 2003. → pages 26[300] W. R. Zipfel, R. M. Williams, and W. W. Webb. Nonlinear magic:multiphoton microscopy in the biosciences. Nature Biotechnology, 21(11):1369–1377, Nov. 2003. ISSN 1087-0156. doi:10.1038/nbt899. → pages 74[301] Y. Ziv, L. D. Burns, E. D. Cocker, E. O. Hamel, K. K. Ghosh, L. J. Kitch,A. El Gamal, and M. J. Schnitzer. Long-term dynamics of CA1hippocampal place codes. Nature Neuroscience, 16(3):264–266, Mar.2013. ISSN 1546-1726. doi:10.1038/nn.3329. → pages 15[302] P. Zou, Y. Zhao, A. D. Douglass, D. R. Hochbaum, D. Brinks, C. A.Werley, D. J. Harrison, R. E. Campbell, and A. E. Cohen. Bright and fastmulticoloured voltage reporters via electrochromic FRET. Naturecommunications, 5, 2014. → pages 9[303] Y. Zuo, G. Yang, E. Kwon, and W.-B. Gan. Long-term sensory deprivationprevents dendritic spine loss in primary somatosensory cortex. Nature, 436(7048):261–265, 2005. → pages 22171


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items