{"title": "Silicon growth cones map silicon retina", "book": "Advances in Neural Information Processing Systems", "page_first": 1329, "page_last": 1336, "abstract": "", "full_text": "Silicon Growth Cones Map Silicon Retina\n\nBrian Taba and Kwabena Boahen\u2217\n\nDepartment of Bioengineering\n\nUniversity of Pennsylvania\n\nPhiladelphia, PA 19104\n\n{btaba,boahen}@seas.upenn.edu\n\nAbstract\n\nWe demonstrate the \ufb01rst fully hardware implementation of retinotopic\nself-organization, from photon transduction to neural map formation.\nA silicon retina transduces patterned illumination into correlated spike\ntrains that drive a population of silicon growth cones to automatically\nwire a topographic mapping by migrating toward sources of a diffusible\nguidance cue that is released by postsynaptic spikes. We varied the pat-\ntern of illumination to steer growth cones projected by different retinal\nganglion cell types to self-organize segregated or coordinated retinotopic\nmaps.\n\n1\n\nIntroduction\n\nEngineers have long admired the brain\u2019s ability to effortlessly adapt to novel situations\nwithout instruction, and sought to endow digital computers with a similar capacity for un-\nsupervised self-organization. One prominent example is Kohonen\u2019s self-organizing map\n[1], which achieved popularity by distilling neurophysiological insights into a simple set\nof mathematical equations. Although these algorithms are readily simulated in software,\nprevious hardware implementations have required high precision components that are ex-\npensive in chip area (e.g. [2, 3]). By contrast, neurobiological systems can self-organize\ncomponents possessing remarkably heterogeneous properties. To pursue this biological ro-\nbustness against component mismatch, we designed circuits that mimic neurophysiological\nfunction down to the subcellular level. In this paper, we demonstrate topographic re\ufb01ne-\nment of connections between a silicon retina and the \ufb01rst neuromorphic self-organizing\nmap chip, previously reported in [5], which is based on axon migration in the developing\nbrain.\n\nDuring development, neurons wire themselves into their mature circuits by extending ax-\nonal and dendritic precursors called neurites. Each neurite is tipped by a motile sensory\nstructure called a growth cone that guides the elongating neurite based on local chemi-\ncal cues. Growth cones move by continually sprouting and retracting \ufb01nger-like exten-\nsions called \ufb01lopodia whose dynamics can be biased by diffusible ligands in an activity-\ndependent manner [4]. Based on these observations, we designed and fabricated the Neu-\nrotrope1 chip to implement a population of silicon growth cones [5]. We interfaced Neu-\n\n\u2217\n\nwww.neuroengineering.upenn.edu/boahen\n\n\fN\n\n\u0006\u001cN\u001d\n\n?\u001c=\u001d\n\n=\n\nJ\nA\nC\nH\n=\nJ\n\nA\n?\nH\nK\n\u0006\nI\n\na\n\nb\n\nc\n\nd\n\nFigure 1: Neurotropic axon guidance. a. Active source cells (grey) relay spikes down\ntheir axons to their growth cones, which excite nearby target cells. b. Active target cell\nbodies secrete neurotropin. c. Neurotropin spreads laterally, establishing a spatial concen-\ntration gradient that is sampled by active growth cones. d. Active growth cones climb the\nlocal neurotropin gradient, translating temporal activity coincidence into spatial position\ncoincidence. Growth cones move by displacing other growth cones.\n\nrotrope1 directly to a spiking silicon retina to illustrate its applicability to larger neuromor-\nphic systems.\n\nThis paper is organized as follows. In Section 2, we present an algorithm for axon mi-\ngration under the guidance of a diffusible chemical whose release and uptake is gated by\nactivity. In Section 3, we describe our hardware implementation of this algorithm. In Sec-\ntion 4, we examine the Neurotrope1 system\u2019s performance on a topographic re\ufb01nement\ntask when driven by spike trains generated by a silicon retina in response to several types\nof illumination stimuli.\n\n2 Neurotropic axon guidance\n\nWe model the self-organization of connections between two layers of neurons (Fig. 1).\nCells in the source layer innervate cells in the target layer with excitatory axons that are\ntipped by motile growth cones. Growth cones tow their axons within the target layer as\ndirected by a diffusible guidance factor called neurotropin that they bind from the local\nextracellular environment. Neurotropin is released by postsynaptically active target cell\nbodies and bound by presynaptically active growth cones, so the retrograde transfer of\nneurotropin from a target cell to a source cell measures the temporal coincidence of their\nspike activities. Growth cones move to maximize their neurotropic uptake, a Hebbian-like\nlearning rule that causes cells that \ufb01re at the same time to wire to the same place. To\nprevent the population of growth cones from attempting to trivially maximize their uptake\nby all exciting the same target cell, we impose a synaptic density constraint that requires a\nmigrating growth cone to displace any other growth cone occupying its path.\n\nTo state the model more formally, source cell bodies occupy nodes of a regular two-\ndimensional (2D) lattice embedded in the source layer, while growth cones and target cell\nbodies occupy nodes on separate 2D lattices that are interleaved in the target layer. We in-\ndex nodes by their positions in their respective layers, using Greek letters for source layer\npositions (e.g., \u03b1 \u2208 Z2) and Roman letters for target layer positions (e.g., x, c \u2208 Z2).\nEach source cell \u03b1 \ufb01res spikes at a rate aSC(\u03b1) and conveys this presynaptic activity down\nan axon that elaborates an excitatory arbor in the target layer centered on c(\u03b1). In principle,\nevery branch of this arbor is tipped by its own motile growth cone, but to facilitate ef\ufb01cient\n\n\f5E\u0006E?\u0006\u0006\u0014HAJE\u0006=\n\n\u00144)\u0004\n\n\u0004AKH\u0006JH\u0006FA\u001f\n\n\u001e\n\n\u001f\n\n \n\n!\n\n\"\n\n\u001f\n\n!\n\n\"\n\n\u001e\n\n \n\n!\n\n\u001e\n\n\"\n\n\u001f\n\n \n\n\u001e\n\n\u001f\n\n \n\n!\n\n\"\n\n/+\n\n\u0004\n\n\u0004\n\n/+\n\n/+\n\n/+\n\n\u0004\n\n)-4\n\n/+\n\n\u0004\n\n/+\n\n)-4\n\n75*\n\n+\u0006\u0006FKJAH\n\n\u0005+\n\na\n\nb\n\nc\n\nFigure 2: a. Neurotrope1 system. Spike communication is by address-events (AER). b.\nNeurotrope1 cell mosaic. The extracellular medium (grey) is laid out as a monolithic hon-\neycomb lattice. Growth cones (GC) occupy nodes of this lattice and extend \ufb01lopodia to the\nadjacent nodes. Neurotropin receptors (black) are located at the tip of each \ufb01lopodium and\nat the growth cone body. Target cells (N) occupy nodes of an interleaved triangular lattice.\nc. Detail of chip layout.\n\nhardware implementation, we abstract the collection of branch growth cones into a single\ncentral growth cone that tows the arbor\u2019s trunk around the target layer, dragging the rest of\nthe arbor with it. The arbor overlaps nearby target cells with a branch density A(x \u2212 c(\u03b1))\nthat diminishes with distance kx \u2212 c(\u03b1)k from the arbor center. The postsynaptic activity\naTC(x) of target cell x is proportional to the linear sum of its excitation.\n\naSC(\u03b1)A(x \u2212 c(\u03b1))\n\n(1)\n\naTC(x) =X\n\n\u03b1\n\nPostsynaptically active target cell bodies release neurotropin, which spreads laterally until\nconsumed by constitutive decay processes. The neurotropin n(x0) present at target site\nx0 is assembled from contributions from all active release sites. The contribution of each\ntarget cell x is proportional to its postsynaptic activity and weighted by a spreading kernel\nN(x \u2212 x0) that is a decreasing function of its distance kx\u2212 x0k from the measurement site\nx0.\n\naTC(x)N(x \u2212 x\n\n0)\n\n(2)\n\n0) =X\n\nn(x\n\nx\n\nA presynaptically active growth cone located at c(\u03b1) computes the direction of the local\nneurotropin gradient by identifying the adjacent lattice node c0(\u03b1) \u2208 C(c(\u03b1)) with the most\nneurotropin, where C(c(\u03b1)) includes c(\u03b1) and its nearest neighbors.\n\n0(\u03b1) = arg maxx0\u2208C(c(\u03b1))n(x\nc\n\n(3)\nOnce the growth cone has identi\ufb01ed c0(\u03b1), it swaps positions with the growth cone al-\nready located at c0(\u03b1), increasing its own neurotropic uptake while preserving a con-\nstant synaptic density. Growth cones compute position updates independently, at a rate\n\u03bb(\u03b1) \u221d aSC(\u03b1) maxy\u2208C(c(\u03b1)) n(x0). Updates are executed asynchronously, in order of\ntheir arrival.\n\n0)\n\nSoftware simulation of a similar set of equations generates self-organized feature maps\nwhen driven by appropriately correlated source cell activity [6]. Here, we illustrate to-\npographic map formation in hardware using correlated spike trains generated by a silicon\nretina.\n\n\f4\u0006KJA\n\n7F@=JA\n\n4\u0006KJA\n\n\u001e!\n\n\" \u001f  \n\n\u001f\u001e\n\n  ! \"\n\n\u001e!\n\n\" \u001f  \n\n\u001f\u001e\n\n  ! \"\n\n\u001e \n\n\" \u001f !\n\n\u001f\u001e\n\n  ! \"\n\n\u001f\u001e\n\n  ! \"\n\n!\u001f\n\n\" \u001e  \n\n\u001f\u001e\n\n  ! \"\n\n!\u001f\n\n\" \u001e  \n\n\u001f\u001e\n\n  ! \"\n\n!\u001f\n\n\u001e \"  \n\na\n\nb\n\nc\n\nFigure 3: Virtual axon remapping. a. Cell bodies tag their spikes with their own source\nlayer addresses, which the forward lookup table translates into target layer destinations.\nb. Axon updates are computed by growth cones, which decode their own target layer\naddresses through the reverse lookup table to obtain the source layer addresses of their cell\nbodies that identify their entries in the forward lookup table. c. Growth cones move by\nmodifying their entries in the forward and reverse lookup tables to reroute their spikes to\nupdated locations.\n\n3 Neurotrope1 system\n\nOur hardware implementation splits the model into three stages: the source layer, the target\nlayer, and the intervening axons (Fig. 2a). Any population of spiking neurons can act\nas a source layer; in this paper we employ the silicon retina of [7]. The target layer is\nimplemented by a full custom VLSI chip that interleaves a 48 \u00d7 20 array of growth cone\ncircuits with a 24 \u00d7 20 array of target cell circuits. There is also a spreading network\nthat represents the intervening medium for propagating neurotropin. The Neurotrope1 chip\nwas fabricated by MOSIS using the TSMC 0.35\u00b5m process and has an area of 11.5 mm2.\nConnections are speci\ufb01ed as entries in a pair of lookup tables, stored in an off-chip RAM,\nthat are updated by a Ubicom ip2022 microcontroller as instructed by the Neurotrope1 chip.\nThe ip2022 also controls a USB link that allows a computer to write and read the contents\nof the RAM. Subsection 3.1 explains how updates are computed by the Neurotrope1 chip\nand Subsection 3.2 describes the procedure for executing these updates.\n\n3.1 Axon updates\n\nAxon updates are computed by the Neurotrope1 chip using the transistor circuits described\nin [5]. Here, we provide a brief description. The Neurotrope1 chip represents neurotropin\nas charge spreading through a monolithic transistor channel laid out as a honeycomb lat-\ntice. Each growth cone occupies one node of this lattice and extends \ufb01lopodia to the three\nadjacent nodes, expressing neurotropin receptors at all four locations (Fig. 2b-c). When\na growth cone receives a presynaptic spike, its receptor circuits tap charge from all four\nnodes onto separate capacitors. The \ufb01rst capacitor voltage to integrate to a threshold resets\nall of the growth cone\u2019s capacitors and transmits a request off-chip to update the growth\ncone\u2019s position by swapping locations with the growth cone currently occupying the win-\nning node.\n\n3.2 Address-event remapping\n\nChips in the Neurotrope1 system exchange spikes encoded in the address-event represen-\ntation (AER) [8], an asynchronous communication protocol that merges spike trains from\nevery cell on the same chip onto a single shared data link instead of requiring a dedicated\nwire for each connection. Each spike is tagged with the address of its originating cell for\ntransmission off-chip. Between chips, spikes are routed through a forward lookup table\nthat translates their original source layer addresses into their destined target layer addresses\n\n\fRetina color map\n\nn=0\n\nn=85\n\na\n\nb\n\nc\n\n<F\u0001 n\u0001>\n\n25\n20\n15\n10\n5\n\nd\n\n25 50 75\n\nn\n\nFigure 4: Retinotopic self-organization of ON-center RGCs. a. Silicon retina color map\nof ON-center RGC body positions. A representative RGC body is outlined in white, as\nare the RGC neighbors that participate in its topographic order parameter \u03a6(n). b. Target\nlayer color map of growth cone positions for sample n = 0, colored by the retinal positions\nof their cell bodies. Growth cones projected by the representative RGC and its nearest\nneighbors are outlined in white. Grey lines denote target layer distances used to compute\n\u03a6(n). c. Target layer color map at n = 85. d. Order parameter evolution.\n\non the receiving chip (Fig. 3a). An axon entry in this forward lookup table is indexed by\nthe source layer address of its cell body and contains the target layer address of its growth\ncone. The virtual axon moves by updating this entry.\n\nAxon updates are computed by growth cone circuits on the Neurotrope1 chip, encoded as\naddress-events, and sent to the ip2022 for processing. Each update identi\ufb01es a pair of axon\nterminals to be swapped. These growth cone addresses are translated through a reverse\nlookup table into the source layer addresses that index the relevant forward lookup table\nentries (Fig. 3b). Modi\ufb01cation of the affected entries in each lookup table completes the\naxon migration (Fig. 3c).\n\n4 Retinotopic self-organization\n\nWe programmed the growth cone population to self-organize retinotopic maps by driving\nthem with correlated spike trains generated by the silicon retina. The silicon retina trans-\nlates patterned illumination in real-time into spike trains that are fed into the Neurotrope1\nchip as presynaptic input from different retinal ganglion cell (RGC) types. An ON-center\nRGC is excited by a spot of light in the center of its receptive \ufb01eld and inhibited by light in\nthe surrounding annulus, while an OFF-center RGC responds analogously to the absence\nof light. There is an ON-center and an OFF-center RGC located at every retinal coordinate.\n\nTo generate appropriately correlated RGC spike trains, we illuminated the silicon retina\nwith various mixtures of light and dark spot stimuli. Each spot stimulus was presented\nagainst a uniformly grey background for 100 ms and covered a contiguous cluster of RGCs\ncentered on a pseudorandomly selected position in the retinal plane, eliciting overlapping\nbursts of spikes whose coactivity established a spatially restricted presynaptic correlation\nkernel containing enough information to instruct topographic ordering [9]. Strongly driven\nRGCs could \ufb01re at nearly 1 kHz, which was the highest mean rate at which the silicon retina\ncould still be tuned to roughly balance ON- and OFF-center RGC excitability. We tracked\nthe evolution of the growth cone population by reading out the contents of the lookup table\nevery \ufb01ve minutes, a sampling interval selected to include enough patch stimuli to allow\neach of the 48 \u00d7 20 possible patches to be activated on average at least once per sample.\nWe \ufb01rst induced retinotopic self-organization within a single RGC cell type by illuminat-\ning the silicon retina with a sequence of randomly centered spots of light presented against\na grey background, selectively activating only ON-center RGCs. Each of the 960 growth\n\n\fRate\u0001 kHz\u0001\n\nRGC stimulus\n\nn=0\n\nn=310\n\nr\ne\nt\nn\ne\nc\n-\nN\nO\n\nr\ne\nt\nn\ne\nc\n-\nF\nF\nO\n\n1\n\n1\n\nx\n\nx\n\n<F\u0001 n\u0001>\n\n25\n\n20\n\n15\n\n10\n\n5\n\n100\n\n200\n\n300\n\nn\n\na\n\nb\n\nc\n\nd\n\ne\n\nFigure 5: Segregation by cell type under separate light and dark spot stimulation. Top:\nON-center; bottom: OFF-center. a. Silicon retina image of representative spot stimulus.\nLight or dark intensity denotes relative ON- or OFF-center RGC output rate. b. Spike rates\nfor ON-center (grey) and OFF-center (black) RGCs in column x of a cross-section of a\nrepresentative spot stimulus. c. Target layer color maps of RGC growth cones at sample\nn = 0. Black indicates the absence of a growth cone projected by an RGC of this cell type.\nOther colors as in Fig. 4. d. Target layer color maps at n = 310. e. Order parameter\nevolution for ON-center (grey) and OFF-center (black) RGCs.\n\ncones was randomly assigned to a different ON-center RGC, creating a scrambled map\nfrom retina to target layer (Fig. 4a-b). The ON-center RGC growth cone population visibly\nre\ufb01ned the topography of the nonretinotopic initial state (Fig. 4c). We quantify this ob-\nservation by introducing an order parameter \u03a6(n) whose value measures the instantaneous\nretinotopy for an RGC at the nth sample. The de\ufb01nition of retinotopy is that adjacent RGCs\ninnervate adjacent target cells, so we de\ufb01ne \u03a6(n) for a given RGC to be the average target\nlayer distance separating its growth cone from the growth cones projected by the six adja-\ncent RGCs of the same cell type. The population average h\u03a6(n)i converges to a value that\nrepresents the achievable performance on this task (Fig. 4d).\n\nWe next induced growth cones projected by each cell type to self-organize disjoint topo-\ngraphic maps by illuminating the silicon retina with a sequence of randomly centered light\nor dark spots presented against a grey background (Fig. 5a-b). Half the growth cones\nwere assigned to ON-center RGCs and the other half were assigned to the corresponding\nOFF-center RGCs. We seeded the system with a random projection that evenly distributed\ngrowth cones of both cell types across the entire target layer (Fig. 5c). Since only RGCs of\nthe same cell type were coactive, growth cones segregated into ON- and OFF-center clus-\nters on opposite sides of the target layer (Fig. 5d). OFF-center RGCs were slightly more\nexcitable on average than ON-center RGCs, so their growth cones re\ufb01ned their topography\nmore quickly (Fig. 5e) and clustered in the right half of the target layer, which was also\nmore excitable due to poor power distribution on the Neurotrope1 chip.\n\nFinally, we induced growth cones of both cell types to self-organize coordinated retino-\ntopic maps by illuminating the retina with center-surround stimuli that oscillate radially\nfrom light to dark or vice versa (Fig. 6). The light-dark oscillation injected enough coac-\ntivity between neighboring ON- and OFF-center RGCs to prevent their growth cones from\nsegregating by cell type into disjoint clusters. Instead, both subpopulations developed and\nmaintained coarse retinotopic maps that cover the entire target layer and are oriented in\nregister with one another, properties suf\ufb01cient to seed more interesting circuits such as\noriented receptive \ufb01elds [10].\n\nPerformance in this hardware implementation is limited mainly by variability in the behav-\n\n\fRate\u0001 kHz\u0001\n\nRGC stimulus\n\nn=0\n\nn=335\n\nr\ne\nt\nn\ne\nc\n-\nN\nO\n\nr\ne\nt\nn\ne\nc\n-\nF\nF\nO\n\n1\n\n1\n\nx\n\nx\n\n<F\u0001 n\u0001>\n\n25\n\n20\n\n15\n\n10\n\n5\n\n100\n\n200\n\n300\n\nn\n\na\n\nb\n\nc\n\nd\n\ne\n\nFigure 6: Coordinated retinotopy under center-surround stimulation. Top: ON-center; bot-\ntom: OFF-center. a. Silicon retina image of a representative center-surround stimulus.\nLight or dark intensity denotes relative ON- or OFF-center RGC output rate. b. Spike rates\nfor ON-center (grey) and OFF-center (black) RGCs in column x of a cross-section of a\nrepresentative center-surround stimulus. c. Target layer color maps of RGC growth cones\nfor sample n = 0. Colors as in Fig. 5. d. Target layer color maps at n = 335. e. Order\nparameter evolution for ON-center (grey) and OFF-center (black) RGCs.\n\nior of nominally identical circuits on the Neurotrope1 chip and the silicon retina. In the\nsilicon retina, the wide variance of the RGC output rates [7] limits both the convergence\nspeed and the \ufb01nal topographic level achieved by the spot-driven growth cone population.\nGrowth cones move faster when stimulated at higher rates, but elevating the mean output\nrate of the RGC population allows more excitable RGCs to \ufb01re spontaneously at a sus-\ntained rate, swamping growth cone-speci\ufb01c guidance signals with stimulus-independent\npostsynaptic activity that globally attracts all growth cones. The mean RGC output rate\nmust remain low enough to suppress these spontaneous distractors, limiting convergence\nspeed. Variance in the output rates of neighboring RGCs also distorts the shape of the spot\nstimulus, eroding the \ufb01delity of the correlation-encoded instructions received by the growth\ncones.\n\nVariability in the Neurotrope1 chip further limits topographic convergence. Migrating\ngrowth cones are directed by the local neurotropin landscape, which forms an image of\nrecent presynaptic activity correlations as \ufb01ltered through the postsynaptic activation of\nthe target cell population. This image is distorted by variations between the properties of\nindividual target cell and neurotropin circuits that are introduced during fabrication. In\nparticular, poor power distribution on the Neurotrope1 chip creates a systematic gradient\nin target cell excitability that warps a growth cone\u2019s impression of the relative coactivity of\nits neighbors, attracting it preferentially toward the more excitable target cells on the right\nside of the array.\n\n5 Conclusions\n\nIn this paper, we demonstrated a completely neuromorphic implementation of retinotopic\nself-organization. This is the \ufb01rst time every stage of the process has been implemented\nentirely in hardware, from photon transduction through neural map formation. The only\ncomparable system was described in [11], which processed silicon retina data of\ufb02ine using\na software model of neurotrophic guidance running on a workstation. Our system computes\nresults in real time at low power, two prerequisites for autonomous mobile applications.\n\nThe novel infrastructure developed to implement virtual axon migration allows silicon\ngrowth cones to directly interface with an existing family of AER-compliant devices,\n\n\fenabling a host of multimodal neuromorphic self-organizing applications. In particular,\nthe silicon retina\u2019s ability to translate arbitrary visual stimuli into growth cone-compatible\nspike trains in real-time opens the door to more ambitious experiments such as using natural\nvideo correlations to automatically wire more complicated visual feature maps.\n\nOur faithful adherence to cellular level details yields an algorithm that is well suited to\nphysical implementation. In contrast to all previous self-organizing map chips (e.g. [2, 3]),\nwhich implemented a global winner-take-all function to induce competition, our silicon\ngrowth cones compute their own updates using purely local information about the neu-\nrotropin gradient, a cellular approach that scales effortlessly to larger populations. Perfor-\nmance might be improved by supplementing our purely morphogenetic model with addi-\ntional physiologically-inspired mechanisms to prune outliers and consolidate well-placed\ngrowth cones into permanent synapses.\n\nAcknowledgments\n\nWe would like to thank J. Arthur for developing a USB system to facilitate data collection.\nThis project was funded by the David and Lucille Packard Foundation and the NSF/BITS\nprogram (EIA0130822).\n\nReferences\n\n[1] T. Kohonen (1982), \u201cSelf-organized formation of topologically correct feature maps,\u201d Biol.\n\nCybernetics, vol. 43, no. 1, pp. 59-69.\n\n[2] W.-C. Fang, B.J. Sheu, O.T.-C. Chen, and J. Choi (1992), \u201cA VLSI neural processor for im-\nage data compression using self-organization networks,\u201d IEEE Trans. Neural Networks, vol. 3,\nno. 3, pp. 506-518.\n\n[3] S. Rovetta and R. Zunino (1999), \u201cEf\ufb01cient training of neural gas vector quantizers with analog\n\ncircuit implementation,\u201d IEEE Trans. Circ. & Sys. II, vol. 46, no. 6, pp. 688-698.\n\n[4] E.W. Dent and F.B. Gertler (2003), \u201cCytoskeletal dynamics and transport in growth cone mo-\n\nbility and axon guidance,\u201d Neuron, vol. 40, pp. 209-227.\n\n[5] B. Taba and K. Boahen (2003), \u201cTopographic map formation by silicon growth cones,\u201d in:\nAdvances in Neural Information Processing Systems 15 (MIT Press, Cambridge, eds. S. Becker,\nS. Thrun, and K. Obermayer), pp. 1163-1170.\n\n[6] S.Y.M. Lam, B.E. Shi, and K.A. Boahen (2005), \u201cSelf-organized cortical map formation by\n\nguiding connections,\u201d Proc. 2005 IEEE Int. Symp. Circ. & Sys., in press.\n\n[7] K.A. Zaghloul and K. Boahen (2004), \u201cOptic nerve signals in a neuromorphic chip I: Outer and\n\ninner retina models,\u201d IEEE Trans. Bio-Med. Eng., vol. 51, no. 4, pp. 657-666.\n\n[8] K. Boahen (2000), \u201cPoint-to-point connectivity between neuromorphic chips using address-\n\nevents,\u201d IEEE Trans. Circ. & Sys. II, vol. 47, pp. 416-434.\n\n[9] K. Miller (1994), \u201cA model for the development of simple cell receptive \ufb01elds and the ordered\narrangement of orientation columns through activity-dependent competition between on- and\noff-center inputs,\u201d J. Neurosci., vol. 14, no. 1, pp. 409-441.\n\n[10] D. Ringach (2004), \u201cHaphazard wiring of simple receptive \ufb01elds and orientation columns in\n\nvisual cortex,\u201d J. Neurophys., vol. 92, no. 1, pp. 468-476.\n\n[11] T. Elliott and J. Kramer (2002), \u201cCoupling an aVLSI neuromorphic vision chip to a neu-\nrotrophic model of synaptic plasticity: the development of topography,\u201d Neural Comp., vol. 14,\nno. 10, pp. 2353-2370.\n\n\f", "award": [], "sourceid": 2773, "authors": [{"given_name": "Brian", "family_name": "Taba", "institution": null}, {"given_name": "Kwabena", "family_name": "Boahen", "institution": null}]}