{"title": "Bio-inspired Real Time Sensory Map Realignment in a Robotic Barn Owl", "book": "Advances in Neural Information Processing Systems", "page_first": 713, "page_last": 720, "abstract": "The visual and auditory map alignment in the Superior Colliculus (SC) of barn owl is important for its accurate localization for prey behavior. Prism learning or Blindness may interfere this alignment and cause loss of the capability of accurate prey. However, juvenile barn owl could recover its sensory map alignment by shifting its auditory map. The adaptation of this map alignment is believed based on activity dependent axon developing in Inferior Colliculus (IC). A model is built to explore this mechanism. In this model, axon growing process is instructed by an inhibitory network in SC while the strength of the inhibition adjusted by Spike Timing Dependent Plasticity (STDP). We test and analyze this mechanism by application of the neural structures involved in spatial localization in a robotic system.", "full_text": "Bio-inspired Real Time Sensory Map Realignment in\n\na Robotic Barn Owl\n\nJuan Huo, Zhijun Yang and Alan Murray\n\nDTC, School of Informatics, Schoolf of Electronics & Engineering\n\nThe University of Edinburgh\n\nEdinburgh, UK\n\n{J.Huo, Zhijun.Yang, Alan.Murray}@ed.ac.uk\n\nAbstract\n\nThe visual and auditory map alignment in the Superior Colliculus (SC) of barn\nowl is important for its accurate localization for prey behavior. Prism learning or\nBlindness may interfere this alignment and cause loss of the capability of accurate\nprey. However, juvenile barn owl could recover its sensory map alignment by\nshifting its auditory map. The adaptation of this map alignment is believed based\non activity dependent axon developing in Inferior Colliculus (IC). A model is\nbuilt to explore this mechanism. In this model, axon growing process is instructed\nby an inhibitory network in SC while the strength of the inhibition adjusted by\nSpike Timing Dependent Plasticity (STDP). We test and analyze this mechanism\nby application of the neural structures involved in spatial localization in a robotic\nsystem.\n\n1 Introduction\n\nBarn owl is a nocturnal predator with strong able auditory and visual localization system. During\nlocalization, the sensory stimuli are translated into neuron response, the visual and auditory maps\nare formed. In the deep Superior Colliculus (SC), visual and auditory information are integrated\ntogether. Normally, the object localization of visual map and auditory map are aligned with each\nother. But this alignment can be disrupted by wearing a prism or blindness [1, 2]. The juvenile barn\nowl is able to adapt so that it can foveates correctly on the source of auditory stimuli. A model based\non the newest biological discoveries and account for a large amount of biological observations has\nbeen developed to explore the adaptation in map alignment [3].\nThis model is applied to a robotic system emulating the behavior of heading of the barn owl, so\nas to provide a real-time visual and auditory information integration and map realignment. It also\nprovides a new mechanism for the hardware to mimic some of brain\u2019s abilities, adapt to novel\nsituation without instruction.\n\n1.1 Biological Background\n\nSuperior Colliculus (SC) gets different sensory inputs and it sends its outputs to effect behavior.\nAs a hub of sensory information, SC neurons access the auditory stimuli from Inferior Colliculus\n(IC) [4, 1], which includes external Inferior Colliculus (ICx) and central Inferior Colliculus (ICc).\nICx wraps around ICc. As revealed by anatomical and physiological experiments, the main site of\nmap adaptation is in two areas, one is axon connection between ICc and ICx, the other area is an\ninhibitory network in SC.\nLarge amounts of evidence has shown axon sprouting and retraction between ICc and ICx are guided\nby inhibitory network in SC during prism learning [5, 6, 7]. Axons do not extend spontaneously,\n\n1\n\n\f(a)\n\n(b)\n\nFigure 1: (a) The simulation environment. (b) The information projection between ICc, ICx and SC.\n\nthey\u2019re promoted by neurotrophin (one kind of nerve growth factor) release and electrical activity of\nthe cell body [8]. The release of neurotrophin is triggered by guiding signal from SC. In this paper\nwe call the guiding signal, Map Adaptation Cue (MAC), as shown in Fig. 1(b). In the inhibitory\nnetwork, MAC is assumed to be introduced by inter neuron, which is plausible to be bimodal neuron\n[7]. Bimodal neuron can be potentiated by both visual input (from retina) and auditory input (from\nICx). Its neuron response is obviously strenthened when visual and auditory input are correlated [9].\nPrevious work has pointed out Hebbian Learning plays a main role in sensory information integra-\ntion on bimodal neuron [4]. This paper includes a closer representation of biological structure.\n\n2 Neural spike train\n\nNeurons in nervous system process and transmit information by neural spikes. Sensory stimulus\nis coded by the spatiotemporal spike pattern before applied to the spiking neural network [10]. In\nthis study, the input spike pattern was applied repeatedly and frequently, similar as the input stimuli.\nSpike patterns within which the \ufb01xed time intervals between spikes are set mannually, with two\ndiscrete values of mean \ufb01ring rate, high and low. As the neuron response in visual map (retina layer)\nand auditory map (ICx layer) has a center surround pro\ufb01le, the receptive center has the highest\n\ufb01ring rate, e.g.\nthe \u201dmexican hat\u201d, [11]. The spike patterns of visual center and auditory center,\ncorresponding to a same target, are highly correlated with each other. Adjacent neurons respond\nwith template spike trains of low \ufb01ring rate. The spike patterns of center neuron and ajacent neuron\nare independent with each other. The remaining neurons have negligible activity. Another possilbe\nspike train generating method and its simulation result for this model can be found in paper [12].\n\n3 Neural Model\n\nThe simulation is to emulate a virtual barn owl at the center of a \ufb01xed, head-centered reference\nsystem with the origin centered on the perch as in Fig. 1(a). Fig. 2(b) schematically illustrates the\nmodel, 4-layer (ICc, ICx, SC, retina), 10-pathway. Each pathway corresponds to 18\u25e6 in azimuth.\nThe single pathway is composed of two basic sections shown in Fig. 2(a). Block I comprises the\nICc, ICx and the axon connections that map between them. Block II is both the detector of any\nshift between visual and auditory cues and the controller of the ICx/ICc mapping in block I. The\nconnection between ICc and ICx in block I is instructed by Map Adaptation Cue (MAC), which is\ngenerated by the inter neuron in block II.\nIn block II, both bimodal and inter neurons in this model are Leaky Integrate-and-Fire (LIF) neuron\n(Equation 1). ge is the excitatory synaptic conductance, which is associated with excitatory rever-\nsal potential Vexc. Similarly, gi, the inhibitory conductance, is associated with inhibitory reversal\n\n2\n\n\f(a)\n\n(b)\n\nFigure 2: (a) Schematic of the auditory and visual signal processing pathway. (b) Schematic of the\nnetwork. Each single pathway represents 18\u25e6 in azimuth. The visual stimulus arrives in the retina at\nN42, N22 receives the strongest MAC. The active growthcone from N13 is attracted by neurotrophin.\nThe dashed line is the new connection built when the growthcone reaches its threshold. The old\nconnection between N13 and N23 is thus eliminated due to lack of alignment between the auditory\nand visual stimuli.\n\npotential Vinh. gl is the membrane conductance, the membrane resistance in this case is given by\nRm = 1/gl. When the membrane potential V (t) reaches the threshold value of about -50 to -55mV ,\nV (t) is reset to a value Vreset [13]. In this model, Vreset is chosen to be equal to Vrest, the rest mem-\nbrane potential, here Vrest = Vreset = \u221270mV . The other paprameters of the neuron model are as\nfollows: Vexc = 0mV , Vinh = \u221270mV , \u03c4m = CmRm = 5ms.\n\nCm\n\ndV (t)\n\ndt\n\n= \u2212gl(V (t) \u2212 Vrest) \u2212 ge(V (t) \u2212 Vexc) \u2212 gi(V (t) \u2212 Vinh)\n\n(1)\n\nThe synapses connecting the sensory signals with the bimodel neuron are excitatory while the\nsynapse between bimodal neuron and inter neuron is inhibitory. The synaptic weight change in this\nmodel is mediated by Spike Timing Dependent Plasticity (STDP). STDP is a learning rule in which\nthe synaptic weight is strengthened or weakened by the paired presynaptic spikes and postsynaptic\nspikes in a time window [14].\nThe whole network is shown in Fig. 2(b), neuron Nij indicates the neuron location in layer i and\npathway j. The developing of axon growthcone is activated by presynaptic spikes from its source\nlayer ICc (layer 1). The target layer ICx (layer 2) releases neurotrophin when it is excited by MAC\nspikes. The concentration of neurotrophin c2j is set to be linearly proportional to the total MAC\ninduced synaptic activity, P2j, which sums the MAC spikes of ICx layer neurons. In Fig. 2(b),\nN2j(cen) is the ICx neuron that receives strongest stimulation from the visual signal, via the retina\nand SC. The concentrations of neurotrophin released by neurons N2j depend upon the distance\nbetween neuron N2j and N2j(cen), kN2j \u2212 N2j(cen)k. c2j is contributed by all active release sites,\nhowever, this contribution decays with distance. To represent the effect of neighbouring neurons, a\nspreading kernel D(N2j \u2212 N2j(cen)) is used to weight P2j. D(N2j \u2212 N2j(cen)) is an exponential\ndecay function with the decay variable kN2j \u2212 N2j(cen)k. The concentration of neurotrophin also\ndecays with time step .\n\n3\n\n\fc(N2j(cen)) = X\n= X\n\nN2j\n\nN2j\n\nP (N2j)D(N2j \u2212 N2j(cen))\n\nP (N2j)e\u2212\u03bbkN2j\u2212N2j (cen)k\n\n(2)\n\n(3)\n\nWhen there is neurotrophin released, the growth cone begins to grow induced by neural activity.\nThe growth cone activity is bounded by the presynaptic factor which is a summation \ufb01lter repre-\nsenting the linear sum of the presynaptic spikes of the corresponding neuron N1j. The most active\ngrowth cone from source neuron N1j(sou) has the highest possibility to be extended. If N2j(tag)\nis the target direction of growth cone, N2j(tag) is identi\ufb01ed when the accumulated neurotrophin\nc2j(tag) exceeds the threshold, the new connection between N1j(sou) and N2j(tag) is validated,\nmeanwhile the neurotrophin is reset to the initial state. When the new connection is completed, the\nold connection which is bifurcated from the same neuron is blocked [15, 16].\n\nN2j (tag) = argmaxN2j (tag)\u2208Y (N2j )c2j\n\n(4)\n\n4 Real-time Learning and Adaptation\n\n4.1 Experiments\n\nTo analyze the capability of the model in a real-time robotic system, an e-puck robot equipped\nwith two lateral microphones, and a camera with a 30\u25e6 prism is shown in Fig. 3. E-puck robot\ncommunicates with PC through bluetooth interface. We use e-puck robot to emulate barn owl head.\nThe visual and auditory target (LED and loudspeaker) was \ufb01xed in one location and the owl-head\nrobot moves into different directions manually or by motor command. The high \ufb01ring rate spike\npattern was fed into center neurons, which correspond to the target localization in space, in ICc or\nretina layer. In the network model, each pathway represents 18\u25e6 \ufb01eld in space. We label the neurons\ncorresponding to the azimuth angle \u221290\u25e6 \u223c \u221272\u25e6, pathway 1, so that azimuth angle 0\u25e6 \u223c 18\u25e6 is\nrepresented by pathway 6.\nThe chirp from the loudspeaker is 1K Hz sine wave. The sound signal is processed by Fast Fourier\nTransform (FFT). When the average amplitude of the input signal above a threshold, the character-\nistic frequency f and phase \u2206\u03c6 between the left and right ear are calculated. With Equation 5 and\n6, we get the interaural time difference \u2206t and the target direction \u03b8 in azimuth. In this equation, V\nis sound speed, L is the diameter of the robot head.\n\n\u2206t =\n\n\u03b8 =\n\n\u2206\u03c6\n2\u03c0f\n\u2206tV\nL\n\n(5)\n\n(6)\n\n4.2 Experiment Results\n\nThe experiment consisted of two steps: \ufb01rst, the owl-head robot without prism was positioned to\nhead towards different directions in a random sequence. For every stimulation, a visual or an audio-\nvisual target was presented at one of the 10 available locations. Secondly, the owl-head robot wear-\ning prism with azimuth angle 36\u25e6 was presented to randomly selected direction in azimuth. In each\ndirection, the target stimuli repeated 75 times. Each stimuli introduce spike cluster in 40 time units.\nThese spikes are binary signals with equal amplitude. Experiment results have shown that the system\nwas able to adjust itself in different initial conditions.\nThe results of 0\u25e6 target localization in the \ufb01rst experiment are shown in Fig. 4. Since visual and\nauditory signals are registered with each other, both the visual excitatory synapse (the arrow between\nN4j and N3j in Fig. 2(b)) and auditory excitatory synapse (the arrow between N2j and N3j in\nFig. 2(b)) are strengthened. This means the bimodal neuron becomes more active. Because of\n\n4\n\n\f(a)\n\n(b)\n\n(c)\n\nFigure 3: (a) E-puck robot wearing a prism. (b) Real-time experiment. (c) Visual and auditory input.\nWe get the visual direction from the luminous image by identifying the position of the brightest pixel.\nThe auditory signal is processed by FFT to identify the phase difference between left and right ear,\nso as to \ufb01nd the auditory direction.\n\nthe inhibitory relationship between bimodal neuron and the interneuron, the interneuron is strongly\ninhibited and its output is close to zero. Therefore, no neurotrophin is released in the ICx neuron, as\nshown in Fig. 4(a). The growthcone does not grow in the auditory layer, so there is no change to the\noriginal axon connection, Fig. 4(b).\nThe results of 0\u25e6 target localization in the second experiment are shown in Fig. 5 and Fig. 6. Be-\ncause of the prism wearing, the visual receptive center and auditory receptive center are in different\npathways, pathway 8 and pathway 6. Visual and auditory input spike trains are independent of each\nother in pathway 8. Thus both visual and auditory synapses connected to the bimodal neuron are\nweakened. The reduced inhibition increases the spike output of interneuron. This stimulates neu-\nrotrophin release in pathway 8. With high neurotrophin value and high \ufb01ring rate spike train input,\nthe pathway 6 growthcone is the most active one at the source layer. When the growthcone grows to\ncertain level, the axon connection is updated, as shown in Fig. 6(b).\nFor the camera is limited by its visual angle \u221230\u25e6 \u223c 30\u25e6, the real-time robot experiment only tested\npathway 4 \u223c 7. The rest of the pathway test is simulated in PC in terms of data accessed from\npathway 4 \u223c 7. The last map realignment result is shown in Fig.7.\n\n5 Conclusion\n\nAdaptability is a crucial issue in the design of autonomous systems. In this paper, we demonstrate\na robust model to eliminate the visual and auditory localization disparity. This model explains the\nmechanism behind visual and auditory signal integration. Spike-Timing Dependent Plasticity is\naccompanied by modulation of the signals between ICc and ICx neurons. The model also provides\nthe \ufb01rst clear indication of the possible role of a \u201dMap Adaptation Cue\u201d in map alignment. The\nreal-time application in a robotic barn owl head shows the model can work in real world which the\nbrains of animals have to face.\nBy studying the brain wiring mechanism in superior colliculus, we can better understand the maps\nalignment in brain. Maps alignment also exists in hippacampus and cortex. It is believed maps\nalignment plays an important role for learning, perception and memory which is our future work.\n\n5\n\n\fFigure 4: Visual and auditory localization signals from a same target are registered with each other.\n(a) There\u2019s no neurotrophin released from ICx layer at any time during the experiment. (b) The\naxon connection between ICc and ICx doesn\u2019t change. (c)(d) Here the target direction is in 0\u25e6. Both\nvisual and auditory receptive center corresponds to pathway 6 and their synaptic weight increases\nsimultaneously.\n\nFigure 5: Visual and auditory localization signals are misaligned with each other. (a) Neurotrophin\nreleased from the target ICx neurons is accumulated. (b) The axon connection between ICc and\nICx doesn\u2019t change before the neurontrophin and growthcone reach a threshold. Here the visual\nreceptive center is in pathway 8, while the auditory receptive center is in pathway 6. (c)(d) Both the\nvisual and auditory synapses are weakened because the input spike trains are independent with each\nother.\n\n6\n\n\fFigure 6: New axon connection is built. (a) After the axon connection is updated, neurotrophin is\nreset to its original status. (b) The new axon connection is built while the old connection is inhibited.\n(c)(d) Both visual and auditory synapses begin to increase after visual and auditory signal registered\nwith each other again.\n\nFigure 7: The arrangement of axon connection between maps. The small square represents the\noriginal point to point connection. The black blocks represent the new connection after adaptation.\n\n7\n\n\fAnother issue for further discussion is MAC. Although it is clear MAC is generated by an inhibitory\nnetwork in SC, whether it comes from bimodal neuron or not remains unclear.\nThe video of the experiment can be found on website:\n\nhttp://www.see.ed.ac.uk/ s0454392/\n\nAcknowledgments\n\nFor this research, we are grateful to Barbara Webb\u2019s suggestion for using e-puck robot. We would\nlike to thank the support of the EPSRC Doctoral Training Center in Neuroinformatics. We also\nthank Leslie Smith for advice and assistance in model building.\n\nReferences\n[1] E. Knudsen, \u201cAuditory and visual maps of space in the optic tectum of the owl,\u201d The Journal\n\nof Neuroscience, vol. 2, pp. 1177\u20131194, 1982.\n\n[2] K. EI, \u201cEarly blindness results in a degraded auditory map of space in the optic tectum of the\n\nbarn owl,\u201d Proc Natl Acad Sci, vol. 85, no. 16, pp. 6211\u20134, 1988.\n\n[3] J. Huo, A. Murray, L. Smith, and Z. Yang, \u201cAdaptation of barn owl localization system with\nIEEE World Congress on Computational Intelligence,\n\nspike timing dependent plasticity.\u201d\nJune 2008.\n\n[4] M. Rucci, G. Tononi, and G. M. Edelman, \u201cRegistration of neural maps through value-\ndependent learning: modeling the alignment of auditory and visual maps in the barn owl\u2019s\noptic tectum,\u201d J Neurosci., vol. 17, no. 1, pp. 334\u201352, 1997.\n\n[5] J. I.Gold and K. EI, \u201cAdaptive adjustment of connectivity in the inferior colliculus revealed by\n\nfocal pharmacological inactivation,\u201d J Neurophysiol., vol. 85(4), pp. 1575\u201384, 2001.\n\n[6] P. S. Hyde and E. I. Knudsen, \u201cTopographic projection from the optic tectum to the auditory\nspace map in the inferior colliculus of the barn owl,\u201d J Comp Neurol, vol. 421, pp. 146\u2013160,\n2000.\n\n[7] Y. Gutfreund, W. Zheng, and E. I. Knudsen, \u201cGated visual input to the central auditory system,\u201d\n\nScience, vol. 297, pp. 1556\u20131559, 2002.\n\n[8] J. L. Goldberg, J. S. Espinosa, Y. Xu, N. Davidson, G. T. Kovacs, and B. A. Barres, \u201cReti-\nnal ganglion cells do not extend axons by default: Promotion by neurotrophic signaling and\nelectrical activity,\u201d Neuron, vol. 33, pp. 689\u2013702, 2002.\n\n[9] M. Meredith, J. Nemitz, and B. Stein, \u201cDeterminants of multisensory integration in superior\n\ncolliculus neurons. i. temporal factors.\u201d J Neurosci., vol. 10, pp. 3215\u201329, 1987.\n\n[10] R. Hosaka, O. Araki, and T. Ikeguchi, \u201cStdp provides the substrate for igniting syn\ufb01re chains\n\nby spatiotemporal input patterns,\u201d Neural Computation, vol. 20, pp. 415\u2013435, 2008.\n\n[11] J. Sneyd, Mathematical Modeling in Physiology, Cell Biology, and Immunology. New Or-\n\nleans, Louisiana: American Mathematical Society, January 2001, vol. 59.\n\n[12] J. Huo, Z. Yang, and A. Murray, \u201cModeling visual and auditory integration of barn owl superior\n\ncolliculus with stdp.\u201d\n\nIEEE CIS&RAM, June 2008.\n\n[13] L.F.Abbott and P. Dayan, Theoretical Neuroscience. Cambridge: MIT Press, August 2001,\n\nch. 6.\n\n[14] S. Song, K. D. Miller, and L.F.Abbott, \u201cCompetitive hebbian learning through spike-timing-\n\ndependent synaptic plasticity,\u201d Nature Neurosci, vol. 3, pp. 919\u2013926, 2000.\n\n[15] E. W.Dent and F. B.Gertler, \u201cCytoskeletal dynamics and transport in growth cone mobility and\n\naxon guidance,\u201d Neuron, vol. 40, pp. 209\u2013227, 2003.\n\n[16] H. Hatt and D. O. Smith, \u201cSynaptic depression related to presynaptic axon conduction block,\u201d\n\nJ. Physiol., vol. 259, no. 2, pp. 367\u201393, 1976.\n\n8\n\n\f", "award": [], "sourceid": 893, "authors": [{"given_name": "Juan", "family_name": "Huo", "institution": null}, {"given_name": "Zhijun", "family_name": "Yang", "institution": null}, {"given_name": "Alan", "family_name": "Murray", "institution": null}]}