{"title": "Neuronal Maps for Sensory-Motor Control in the Barn Owl", "book": "Advances in Neural Information Processing Systems", "page_first": 366, "page_last": 374, "abstract": null, "full_text": "366 \n\nNEURONAL MAPS FOR SENSORY -MOTOR \n\nCONTROL IN THE BARN OWL \n\nC.D. Spence, J.C. Pearson, JJ. Gelfand, and R.M. Peterson \n\nDavid Sarnoff Research Center \nSubsidiary of SRI International \n\nCN5300 \n\nPrinceton, New Jersey 08543-5300 \n\nW.E. Sullivan \n\nDepartment of Biology \nPrinceton University \n\nPrinceton, New Jersey 08544 \n\nABSTRACT \n\nThe bam owl has fused visual/auditory/motor representations of \nspace in its midbrain which are used to orient the head so that visu(cid:173)\nal or auditory stimuli are centered in the visual field of view. We \npresent models and computer simulations of these structures which \naddress various problems, inclu<lln~ the construction of a map of \nspace from auditory sensory information, and the problem of driv(cid:173)\ning the motor system from these maps. We compare the results \nwith biological data. \n\nINTRODUCTION \n\nMany neural network models have little resemblance to real neural structures, part(cid:173)\nly because the brain's higher functions, which they attempt to imitate, are not yet \nexperimentally accessible. Nevertheless, some neural-net researchers are finding that \nthe accessible structures are interesting, and that their functions are potentially use(cid:173)\nful. Our group is modeling a part of the barn owl's nervous system which orients \nthe head to visual and auditory stimuli. \n\nThe bam owl's brain stem and midbrain contains a system that locates visual and \nauditory stimuli in space. The system constructs an auditory map of spatial direction \nfrom the non-spatial information in the output of the two cochlea. This map is in \nthe external nucleus of the inferior colliculus, or ICx [Knudsen and Konishi, 1978]. \nThe ICx, along with the visual system, projects to the optic tectum, producing a \nfused visual and auditory map of space [Knudsen and Knudsen, 1983]. The map in the \ntectum is the source of target position used by the motor system for orienting the \nhead. \n\nIn the last futeen years, biologists have determined all of the structures in the sys(cid:173)\ntem which produces the auditory map of space in the ICx. This system provides sev-\n\n\fNeuronal Maps for Sensory-Motor Coritrol in the Barn Owl \n\n367 \n\neral examples of neuronal maps, regions of tissue in which the response properties \nof neurons vary continuously with position in the map. (For reviews, see Knudsen, \n1984; Knudsen, du Lac, and Esterly, 1987; and Konishi, 1986.) Unfortunately, the \nmotor system and the projections from the tectum are not well known, but experi(cid:173)\nmental study of them has recently begun [Masino and Knudsen, 1988]. We should \neventually be able to model a complete system, from sensory input to motor output \n\nIn this paper we present several models of different parts of the head orientation sys(cid:173)\ntem. Fig. 1 is an overview of the structures we'll discuss. In the second section of \nthis paper we discuss models for the construction of the auditory space map ~ the \nlex. In the third section we discuss how the optic tectum might drive the motor \nsystem. \n\nCONSTRUCTION OF AN AUDITORY MAP OF SPACE \n\nThe bam owl uses two binaural cues to locate objects in space: azimuth is derived \nfrom inter-aural time or phase delay (not onset time difference), while elevation is \nderived from inter-aural intensity difference (due to vertical asymmetries in sensitiv-\n\nAcoustic \n\nCore of the ICc \n\n~ \nLateral Shell \nof the ICc \n\nExternal \nNucleus of \nthe IC (ICx) \n\n~ \nOptic \nTectum \n\n~ \n\nRetina \n\nFigure 1. Overview of the neuronal system for target localization in the barn owl \n(head orients towards potential targets for closer scrutiny). \nThe illustration \nfocuses on the functional representations of the neuronal computation, and does \nnot show all of the relevant connections. The grids represent the centrally synthe(cid:173)\nsized neuronal maps and the patterns within them indicate possible patterns of neu(cid:173)\nronal activation in response to acoustic stimuli. \n\n\f368 \n\nSpence, Pearson, Gelfand, Peterson and Sullivan \n\nity). Corresponding to these two cues are two separate processing streams which pro(cid:173)\nduce maps of the binamal cues, which are shown in Figs. 2-5. The information on \nthese maps must be merged in order to construct the map of space in the ICx. \n\nlTD ---. \n\nlex :~:~: \n\nf \n\n.~ \n~ \ntiS \n\n~-----~ \n\nAzimuth \n\nt \n\n~ _ __ .-...IID \n\nFigure 2. Standard Model for the construction of an auditory map of space from \nmaps of the binaural cues. Shading represents activity level. lID is Inter-aural \nIntensity Difference, lTD is Inter-aural Time Delay. \n\nA simple model for combining the two maps is shown in Fig. 2. It has not been \ndescribed explicitly in the literature, but it has been hinted at [Knudsen, et ai, \n1987]. For this reason we have called it the standard model. Here all of the neurons \nrepresenting a given time delay or azimuth in the lTD vs. frequency map project to \nall of the neurons representing that azimuth in the space map. Thus a stimulus with \na certain lTD would excite a strip of cells representing the associated azimuth and \nall elevations. Similarly, all of the neurons representing a given intensity difference \nor elevation in the lID vs. frequency map project to all of the neurons representing \nthat elevation in the space map. (TIle map of lID vs. frequency is constructed in the \nnucleus ventralis lemnisci lateralis pars posterior, or VL Vp. VL Vp neurons are said \nto be sensitive to intensity difference, that is they fire if the intensity difference is \ngreat enough. Neurons in the VL Vp are spatially organized by their intensity differ(cid:173)\nence threshold [Manley. et ai, 1988]. Thus, intensity difference has a bar-chart-like \nrepresentation, and our model needs some mechanism to pick out the ends of the \nbars.) Only the neurons at the intersection of the two strips will fire if lateral inhi(cid:173)\nbition allows only those neurons receiving the most stimulation to fue. In the third \nsection we will present a model for connections of inhibitory inter-neurons which \ncan be applied to this model. \nPart of the motivation for the standard model is the problem with phase ghosts. \nPhase ghosts occur when the barn owl's nervous system incorrectly associates the \nwave fronts arriving at the two ears at high frequency. In this case, neurons in the \nmap of lTD vs. frequency become active at locations representing a time delay which \n\n\fNeuronal Maps for Sensory-Motor Control in the Barn Owl \n\n369 \n\nthose neurons representing \n\ndiffers from the true time delay by an integer multiple of the period of the sound. \nBecause the period varies inversely with the frequency, these phase ghosts will have \napparent time delays that vary with frequency. Thus, for stimuli that are not pure \ntones, if the bam owl can compare the activities in the map at different frequencies, \nit can eliminate the ghosts. The standard model does this (Fig. 2). In the lTD vs. fre(cid:173)\nquency map there are more neurons fIring at the position of the true lTD than at the \nghost positions, so space map neurons representing the true position will receive the \nmost stimulation. Only \ntrue position will fIre \nbecause of the lateral inhibition. \nThere is another kind of ghost which we call the multiple-source ghost (Fig. 3). If \ntwo sounds occur simultaneously, then space map neurons representing the time \ndelay of one source and the intensity difference of the other will receive a large \namount of stimulation. Lateral inhibition may suppress these ghosts, but if so, the \nowl should only be able to locate one source at a time. In addition, the true sources \nmight be suppressed. The bam owl may actually suffer from this problem, although \nit seems unlikely if the owl has to function in a noisy environment The relevant \nbehavioral experiments have not yet been done. \nExperimental evidence does not support the standard model. The ICx receives most \nof its input from the lateral shell of the central nucleus of the inferior colliculus \n\nthe \n\nlTD --.. \n\n~Cx \n\n4~ \n\n1- ~\" \n\n\"' i/\" \n\n. ::. \u2022 \n\n-\n\nt \n\n/ \n\n/ \n\n., \n\nI \n\n/: ~ \n\n\" ',/ \n\nt \n\nliD \n\nAzimuth \n\nFigure 3. Multiple-source ghosts in the standard model for the construction of \nthe auditory space map. For clarity, only two pure tone stimuli are represented, \nand their frequencies and locations are such that the \"phase ghost\" problem is \nnot a factor. The black squares represent regions of cells that are above thresh(cid:173)\nold. The circled regions are those that are fIring in response to the lTD of one \nstimulus and the lID of another. These regions correspond to phantom targets. \n\n\f370 \n\nSpence, Pearson, Gelfand, Peterson and Sullivan \n\n(lateral shell of the ICc) [Knudsen, 1983]. Neurons in the lateral shell are tuned to \nfrequency and time delay, and these parameters are mapped [Wagner. Takahashi. and \nKonishi, 1987]. However, they are also affected by intensity difference [Knudsen and \nKonishi. 1978, I. Fujita, private communication]. Thus the lateral shell does not fit \nthe picture of the input to the ICx in the standard model. rather it is some interme(cid:173)\ndiate stage in the processing. \n\nWe have a model, called the lateral shell model. which does not suffer from multi(cid:173)\nple-source ghosts (Fig. 4). In this model, the lateral shell of the ICc is a tbree(cid:173)\ndimensional map of frequency vs. intensity difference vs. time delay. A neuron in \nthe map of time delay vs. frequency in the ICc core projects to all of the neurons in \n\nlTD \n\nlID \n\nlex \n\nAzimuth \n\nFigure 4. Lateral shell model for the construction of the auditory map of space \nin the ICx. f: frequency. lTD: inter-aural time delay. lID: inter-aural intensity \ndifference. \n\n\fNeuronal Maps for Sensory-Motor Control in the Barn Owl \n\n371 \n\nthe three-dimensional map which represent the same time delay and frequency. As in \nthe standard model, a strip of neurons is stimulated, but now the frequency tuning \nis preserved. The map of intensity difference vs. frequency in the nucleus ventralis \nlemnisci lateralis pars posterior (VLVp) [Manley, et al, 1988] projects to the three(cid:173)\ndimensional map in a similar fashion. Lateral inhibition picks out the regions of \nintersection of the strips. Neurons in the space map in the ICx receive input from \nthe strip of neurons in the three-dimensional map which represent the appropriate \ntime delay and intensity difference, or equivalently azimuth and elevation. Phase \nghosts will be present in the three-dimensional map, but in the ICx lateral inhibi(cid:173)\ntion will suppress them. \nMultiple-source ghosts are eliminated in the lateral shell model because the sources \nare essentially tagged by their frequency spectra. If two sources with no common \nfrequencies are present, there are no neurons which represent the time delay of one \nsource, the intensity difference of another, and a frequency common to both. In the \nmore likely case in which some frequencies are common to both sources, there will \nbe fewer neurons fIring at the ghost positions than at the real positions, so again lat(cid:173)\neral inhibition in the ICx can suppress fIring at the ghost positions, exactly as it \nsuppresses the phase ghosts. The fact that intensity and time delay information is \ncombined before frequency tuning is lost in the lex suggests that the owl handles \nmultiple-source ghosts by frequency tagging. A three dimensional map is not essen(cid:173)\ntial, but it is conceptually simple. \nBefore ending this section, we should mention that others have independently \nthought of this model. In particular, M. Konishi and co-workers have looked for a \nspatial organization or mapping of intensity response properties in the lateml shell, \nbut they have not found it They also have said that they can't yet rule it out [M. \nKonishi, I. Fujita, private communication]. \n\nDRMNG THE MOTOR SYSTEM FROM MAPS \n\nAs mentioned before, all of the parts of the auditory system in the bmin stem and \nmidbrain are known, up to the optic tectum. The optic tectum has a map of spatial \ndirection which is common to both the visual and auditory systems. In addition, it \ndrives the motor system, so if the tectum is stimulated at a point, the bam owl's \nhead will move to face a certain direction. The new orientation is mapped in register \nwith the auditory/visual map of spatial direction, e.g., stimulating a location which \nrepresents a stimulus eight degrees to the right of the current orientation of the head \nwill cause the head to turn eight degrees to the right \nLittle is known about the projections of the tectum, although work has started \n[Masino and Knudsen, 1988]. There was one earlier experiment. Two electrodes were \nplaced in one of the tecta at positions representing sensory stimuli eight degrees and \nsixty degrees toward the side of the head opposite to this tectum. When either posi(cid:173)\ntion was stimulated by itself, the alert owl moved its head as expected, by eight or \n\n\f372 \n\nSpence, Pearson, Gelfand, Peterson and Sullivan \n\nsixty degrees. When both were stimulated together. the head moved about forty \ndegrees [du Lac and Knudsen. 1987]. \nThe averaging of activity in the tectum is easy to explain. In some motor models it \nshould be produced naturally by the activation of an agonist-antagonist muscle pair \n(see. for example. Grossberg and Kuperstein. 1986). In the presence of two stimuli. \nthe tension in each muscle is the sum of the tensions it would have for either stimu(cid:173)\nlus alone. so the equilibrium position should be about the average position. \nWe have a different model. which produces a map of the average position. The con(cid:173)\nnection strengths from tectal cells to an averaging map cell decrease quadratically \nwith the difference in represented direction. A quadratic distribution of stimulation \nis very broad. however. so lateral inhibition is required to make the active region \nfairly narrow. \n\n1-D tectum -. \n25 \n\no \no. _ \n\u2022 \no \n\n0 \n\n\u2022 \n\n\u2022 \n\n. \n\n\u2022 \n\n. \n. .. \n\n.-\n\n. .. \n\n0 \n\n0 \n\no \n\n. . . .-\n\u2022 \n.. \n. . '\" . \n. . \n\".;-\n'. .... \n. -.rI' ... ' I.' . \n. .... . ' \n. -. \n. .' ' . \n-- . . \n.- .' . \n. -.. . \n.. \n... . .... ,. .. \n. . \n. . \n'. .. . . \n. . '\" \n'. .... _\". . '.. \n\n. \n\n0 0 \n\no \n\no \n\n0 \n\n0 \n\n, \n\nI . . . . . . ' . _ \n\n\u2022 \n\n\u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \n\no~----------------~~~------------------------~ \n\nPosition \n\nFigure 5. Averaging map simulation. The upper thin rectangle is a one-dimen(cid:173)\nsional version of the space map in the optic tectum. The two marks represent the \nposition of stimulating electrodes that are simultaneously active. The lower \nrectangle is the averaging map. with position represented horizontally and time \nincreasing vertically. The squares represent cell ftrings. Note that the activity \nquickly becomes centered at the average position of the two active positions in \nthe tectum. \n\nWe have simulated a one-dimensional version of this model. with 128 cells in the \ntectum. and in the averaging map. 128 excitatory and 128 inhibitory cells. The exci(cid:173)\ntatory cells and the inhibitory cells both receive the same quadratically weighted \ninput from the tectum. Each inhibitory cell inhibits all of the other cells in the aver(cid:173)\naging map. both excitatory and inhibitory. except for those in a small local neigh(cid:173)\nborhood. (The weights are actually proportional to one minus a gaussian with a max(cid:173)\nimum of one.) An excitatory cell receives exactly the same input as the inhibitory \n\n\fNeuronal Maps for Sensory-Motor Control in the Barn Owl \n\n373 \n\ncell at the same location, so their voltages are the same. Because of this we only \nshow the excitatory cells in Fig. 5. This figure shows cell position horizontally and \ntime increasing vertically. The black squares are plotted at the time a given cell frres. \nWe are interested in whether an architecture like this is biologically plausible. For \nthis reason we have tried to be fairly realistic in our model. The cells obey a mem(cid:173)\nbrane equation: \n\ndv \ndt \n\nC - = -gl (v - VI) - g (v - v ) - g. (v - v.) \nee l I \n\nin which C is the capacitance, the g's are conductances, I refers to leakage quantities. \ne refers to excitatory quantities, and i refers to inhibitory quantities. The output of \na cell is not a real number, but spikes that occur randomly at an average rate that is \na monotonic function of the voltage. We used the usual sigmoidal function in this \nsimulation, although the membrane equation automatically limits the voltage and \nhence the firing rate. A cell that spikes on a given time step affects other cells by \naffecting the appropriate conductance. To get the effect of a post synaptic potential, \nthe conductances obey the equation of a damped harmonic oscillator: \n\ndg \n\n_y \n-2 \n-- (If-ro g \n\nd 2g \n-\ndt2 \n\nWhen a spike from an excitatory cell arrives, we increment the time derivative of g \nby some amount If the oscillator is overdamped or critically damped, the conduc(cid:173)\ntance goes up for a time and then decreases, approaching zero exponentially. \nWe are not suggesting that a damped harmonic oscillator exists in the membranes of \nneurons, but it efficiently models the dynamics of synaptic transmission. The equa(cid:173)\ntions for the conductances also have the nice property that the effects of multiple \nspikes at different times add. \nWith values for the cell parameters that agree well with experimental data, it takes \nabout twenty milliseconds for the simulated map to settle into a fairly steady state, \nwhich is a reasonable time for the function of this map. Also, there was no need to \nfrne tune the parameters; within a fairly wide range the effect of changing them is to \nchange the width of the region of activity. \nWe tried another architecture for the inhibitory intemeurons, in which they received \ntheir input from the excitatory neurons and did not inhibit other inhibitory neurons. \nThe voltages in this architecture oscillated for a very long time, without picking out \na maximum. The architecture we are now using is apparently superior. Since it is \nquick to pick out a maximum of a broad distribution of stimulation, it should work \nvery well in other models requiring lateral inhibition. such as the lateral shell mod(cid:173)\nel discussed earlier. \n\n\f374 \n\nSpence, Pearson, Gelfand, Peterson and Sullivan \n\nCONCLUSION \n\nWe have presented models for two parts of the barn owl's visual/auditory localiza(cid:173)\ntion and head orientation system. These models make experimentally testable predic(cid:173)\ntions. and suggest architectures for artificial systems. One model constructs a map \nof stimulus position from maps of inter-aural intensity and timing differences. This \ni.e.. the representation of false \nmodel solves potential problems with ghosts. \nsources in the presence of certain kinds of real sources. \n\nAnother model computes the average value of a quantity represented on a neural map \nwhen the activity on the map has a complex distribution. This model explains recent \nphysiological experiments. A simulation with fairly realistic model neurons has \nshown that a biological structure could perform this function in this way. \n\nA common feature of these models is the use of neuronal maps. We have only men(cid:173)\ntioned a few of the maps in the barn owl. and they are extremely common in other \nnervous systems. We think this architecture shows great promise for applications in \nartificial processing systems. \n\nAcknowledgments \n\nThis work was supported by internal funds of the David Sarnoff Research Center. \n\nReferences \n\nGrossberg. S. and M. Kuperstein (1986) Neural dynamics of adaptive motor control. \n\nNorth-Holland. \n\nKnudsen. E.I. (1983) J. Compo Neurology. 218:174-186. \n\nKnudsen. E.I. (1984) in Dynamical Aspects of Neocortical Function. G.M. Edel-\n\nman, W.E. Gall, and W.M. Cowan, editors, Wiley, New York. \n\nKnudsen, E.I. andP.F. Knudsen (1983) J. Compo Neurology, 218:187-196. \nKnudsen, E.I. and M. Konishi (1978) J. Neurophys .\u2022 41:870-884. \nKnudsen. E.!.. S. du Lac. and S. Esterly (1987) Ann. Rev. Neurosci., 10:41-56. \n\nKonishi. M. (1986) Trends in Neuroscience. April. \n\ndu Lac. S. and E.!. Knudsen (1987) Soc. for Neurosci. Abstr., 112.10. \nManley. G.A .\u2022 A.C. Koeppl. and M. Konishi (1988) J. Neurosci. 8:2665-2676 \nMasino. T .\u2022 and E.I. Knudsen (1988) Soc. for Neurosci. Abstr., 496.16. \n\nWagner. H .\u2022 T. Takahashi. and M. Konishi (1987) J. Neurosci. 10:3105-3116 \n\n\f", "award": [], "sourceid": 171, "authors": [{"given_name": "Clay", "family_name": "Spence", "institution": null}, {"given_name": "John", "family_name": "Pearson", "institution": null}, {"given_name": "J.", "family_name": "Gelfand", "institution": null}, {"given_name": "R.", "family_name": "Peterson", "institution": null}, {"given_name": "W.", "family_name": "Sullivan", "institution": null}]}