The Dynamics of Population Responses in Visual Cortex

The control of circadian rhythms in the mammalian brain can be localized to a brain region known as the suprachiasmatic nucleus (SCN). The SCN consists of approximately 20,000 neurons whose firing rates oscillate on an approximately 24 hour rhythm. A large portion of these neurons are capable of maintaining a circadian rhythm in isolation in vitro. The firing rate of each neuron seems to be driven by the cycling of a genetic clock. The precision of the SCN as a whole is a result of the synchronization of these individual clocks, but specific mechanism by which neurons of the SCN synchronize their activity remains unknown. We provide the first biological model for synchronizing these clocks through neurotransmitter mediated communication. The model uses a network of integrate and fire neurons, each with their own independently isolating genetic clock. In this model the firing rate of a neuron is determined by the inputs from other neurons, and the level of mRNA in the genetic clock cycle. In the other direction, the clock cycle is determined by its own dynamics and pushes to the mRNA production rates based on inputs from other neurons. We use this model to show synchronization through this form of communication possible. We further investigate the model by varying the effects of the neurotransmitter on the clock and neuron, and find that only certain neurotransmitter mediated effects lead to synchronization. The results lend support to findings that activity dependent synaptic communication plays a major role in SCN synchronization, and suggest a future path of research focusing on the communication path between a neuron's synapses and it's genetic clock. 70. Shape representation in V4: Investigating position-specific tuning for boundary conformation with the standard model of object recognition Charles Cadieu, Minjoon Kouh, Maximilian Riesenhuber, Tomaso Poggio 1 M.I.T. 2 Georgetown University The computational processes in the intermediate stages of the ventral pathway responsible for visual object recognition are not well understood. A recent physiological study by A. Pasupathy and C. Connor (2001) in intermediate area V4 using contour stimuli, proposes that a population of V4 neurons display object-centered, position-specific curvature tuning. Our simulations of a feedforward, network level mechanism (Riesenhuber and Poggio 1999) exhibit selectivity and invariance properties that correspond to the responses of the V4 cells described by Pasupathy and Connor (2001). These results suggest how object-centered, position-specific curvature tuning of V4 cells may arise from combinations of complex V1 cell responses. Furthermore, the model makes predictions about the responses of the same V4 cells studied by Pasupathy and Connor to novel gray level patterns, such as gratings and natural images. These predictions suggest specific experiments to further explore shape representation in V4. Many physiological studies have shed light on the ventral pathway, which is thought to mediate object recognition processes in primate visual cortex. Neurons in the early stages of the ventral pathway have small receptive fields and are selective to simple features, while neurons far along the pathway in inferotemporal cortex (IT) have large receptive fields and are selective to complex features and objects. The general selectivities at these two stages are relatively well understood. However, neurons at intermediate stages, between V1 and IT, have not been fully characterized, and the computational roles and mechanisms of these stages have yet to be resolved. In one recent study of an intermediate stage in the ventral pathway, Pasupathy and Connor (2001) investigated the nature of shape representation in area V4. Building on a previous study (1999) in which they found tuning of V4 neurons to angle orientations, Pasupathy and Connor examined shape representation more extensively using a set of simple closed shapes formed by combining convex and concave boundary elements. They characterized a subpopulation of V4 neurons as having selectivity for object-centered position-specific boundary conformation, such as neurons that were tuned to multiple curvatures at specific angular positions from the object’s center of mass. We have used the standard model, a computational model, to obtain quantitative characterizations, interpretive models, and predictions of V4 neural responses, which are highly nonlinear and can not be easily analyzed or predicted with other approaches. The standard model, developed by Riesenhuber and Poggio (1999), combines many experimental data and standard assumptions about the ventral pathway into a hierarchical computational model of object recognition. The two basic cognitive requirements of object recognition, invariance and specificity, are evident at the earliest and highest stages within the ventral pathway. Hubel and Wiesel proposed that the translation-invariance of V1 complex cells in the early stages of the visual pathway could be created by pooling together simple cells with similar selectivities but translated receptive fields (1965). Perrett and Oram proposed a similar mechanism within IT (1993), the highest stage of the ventral pathway, to describe invariance as pooling over afferents tuned to transformed versions of the same stimuli. Riesenhuber and Poggio extended these proposals in a quantitative model to describe the mechanisms that achieve invariance and specificity throughout the ventral stream (2002). Kouh and Poggio (Consyne, 2005) discuss the computational mechanisms for selectivity used in the standard model. Our results show that V4 wiring of the type envisaged in the framework of the standard model, can exhibit selectivity and invariance properties that correspond to the responses of V4 neurons. Simulated responses show high correlation with the raw V4 responses over the same stimulus set used by Pasupathy and Connor (2001). As a result, the model neurons show tuning properties similar to those described by Pasupathy and Connor under their methodology (2001). These results suggest how the object-centered, position-specific curvature tuning of V4 neurons as defined and described by Pasupathy and Connor may arise from combinations of V1 subunits. In addition, these model neurons show tuning characteristics consistent with previous physiological studies using gratings (Gallant et. al. 1996). Further preliminary work examines how V4 neuron selectivity can be learned from viewing natural scenes. Work by Serre and Poggio (Consyne, 2005) has examined biologically plausible mechanisms for learning neural selectivity. Features learned using this methodology are also consistent with the tuning properties of V4 neurons, indicating that selectivity may be learned from experience. In conclusion, the standard model can be used to quantitatively predict how the V4 neurons studied by Pasupathy and Connor would respond to novel stimuli, such as gratings and natural images. These predictions suggest specific experiments to further explore shape representation in V4. 71. Time-scales of temporal response in regular and fast-spiking cortical neurons Giancarlo La Camera, Alexander Rauch, Walter Senn, Stefano Fusi, Dave Thurbon, Hans R. Luescher 1 Lab of Neuropsychology, NIMH, NIH, Bethesda, USA 2 Max Planck Institute for Biological Cybernetics, Tuebingen, Germany 3 Institute of Physiology, University of Bern, Switzerland 4 The Scripps Research Institute, La Jolla, CA 92037, USA Recently, the use of a noisy input current to investigate single cell and network behavior in vitro and in cultures is becoming a standard and appreciated tool. Such an approach aims at describing the response of cortical neurons as if they were embedded in an intact brain, with the bonus that the input current can be manipulated, and the response easily recorded intracellularly, allowing quantitative modelling under in vivo-like conditions. We used this approach to study and compare the firing patterns of fast spiking (FS) and pyramidal neurons from rat somatosensory cortex on time scales of the order of tens of seconds. FS interneurons showed a pronounced sensitivity to input fluctuations, much larger than pyramidal neurons (50Hz vs 10Hz in response to a subthreshold noisy stimulus). Moreover, although no adaptation of the firing rate seems to occur in the first few hundred milliseconds, cellular processes are at play which reduce the firing rate over time in a slow fashion, as reported recently in somatosensory cortex (Reutimann et al, J. Neurosci. 24: 3295-3303, 2004) and in visual cortex (Descalzo et al, J. Neurophysiol, doi:10.1152/jn.00658.2004). The stationary response (both the firing rate and the variability of the interspike intervals) was characterized as a function of the average and the standard deviation (SD) of the input current, chosen as a Gaussian process. The firing rate in the final seconds of the stimulation interval (i.e. where it varies very slowly) could be fitted by a modified leaky integrate-and-fire model with 5 effective parameters (one representing spike frequency adaptation), as shown previously for pyramidal neurons (Rauch et al, J. Neurophysiol. 90: 1598-1612, 2003). The same model could reproduce well the statistics of the interspike intervals (ISIs), even though its parameters were tuned to fit the firing rates only. The temporal properties of the response are rich even though the stimuli were stationary. Several processes characterize the time course of the instantaneous frequency, which could be reduced to a small number (1 to 3) of phenomenological mechanisms. These mechanisms are to be interpreted as frequency-dependent modulations of the input current, either reducing (adapting) or increasing (facilitating) the neuron's firing rate. FS interneurons could be described either by a single adapting process (time constant of seconds), or by two adapting processes (time constants of hundreds of milliseconds and seconds respectively). For pyramidal neurons, a third process representing a strong initial adaptation, and an intermediate-duration process facilitating the firing rate, were also required. Slow adaptation was not disrupted by the presence of noise. An extended IF model including these processes provided an excellent fit to the data, providing a mechanistic description of the statistics of the temporal processes, i.e. their time constants and magnitudes. The parameters were usually input-dependent; for those which correlated with input current, we derived a model of their dependence. This makes it possible to build a time-dependent model out of data taken in stationary conditions, by allowing the current to change over time with its characteristic way, and letting e.g. the currentdependent parameters follow instantaneously the input dynamics. Such an approach has proved legitimate in very similar models (see e.g. La Camera et al, Neural Comp 16: 2101-2124, 2004). All the parameters describing the neural dynamics vary from cell to cell and the distribution is broad. The different adaptation mechanisms cover a wide range of time scales, ranging from initial adaptation (10-20 ms), to fast adaptation (50-200 ms), early facilitation (0.5-1 s), and slow (or late) adaptation (order of seconds). Although for single cells, the processes are usually distinct and their time scales well separated, populations of different neurons can practically cover all possible time scales, without any gap. Were our study extended to longer stimulation protocols, we would probably find other mechanisms operating on longer time scales. In conclusion, our results indicate that multiple time scales are at play in cortical neurons, even in response to stationary stimuli and in the presence of noise. The time-dependent processes and the distribution of their time scales across different neurons might partially explain the long range correlations in firing rate observed in vivo over many different time scales (S.B. Lowen, et al., Methods 24(4):377-394, 2001). Interestingly these correlations were present also in the absence of any sensory stimulus (the experiment was done in the darkness and the activity was recorded in the cat lateral geniculate nucleus), indicating that they are not a mere reflection of the complexity of the external world. Recently proposed computational consequences of multiple time scales (P.J. Drew, PhD thesis, Brandeis University) range from estimating the envelope of a signal, to discriminate between the responses to rare and common stimuli, to connect and link stimuli which are separated in time by intervals of a few seconds. 72. Parallel processing of multiple stimulus parameters and the emergence of combination sensitivity Bruce A. Carlson, Masashi Kawasaki University of Virginia Information theoretic approaches to neural coding have proven extremely useful in understanding the encoding and processing of time-varying stimuli by sensory systems [1]. The focus of such studies has been on the encoding and initial processing of a single time-varying stimulus parameter [2]. However, natural stimuli frequently vary in several different behaviorally relevant parameters. The task of any sensory system is to encode each of these parameters, extract the relevant features, and then integrate the resulting computations to appropriately guide behavior. We studied the encoding and processing of multiple stimulus parameters at three different stages of the electrosensory system in the weakly electric fish Gymnarchus niloticus. Gymnarchus generates a continuous quasi-sinusoidal electric organ discharge (EOD) for communication and active electrolocation. Only two time-varying stimulus parameters are relevant for these behaviors: amplitude modulation (AM) and phase modulation (PM) of the EOD carrier signal. These two parameters are thought to be processed in separate amplitudeand phase-coding pathways [3]. Two distinct types of primary afferent have been described: S-afferents show tight phase locking to each cycle of the EOD, while O-afferents fire sporadically, in a loosely phase locked manner, at rates that are proportional to EOD amplitude. The S-afferent pathway gives rise to PM-sensitive pyramidal neurons in the electrosensory lateral line lobe (ELL) of the hindbrain, while the O-afferent pathway gives rise to AM-sensitive pyramidal neurons in ELL [3]. Both types of pyramidal neurons project to the torus semicircularis, where the two pathways converge at the single neuron level onto combination-sensitive neurons that show strong selectivity for particular temporal patterns of AM relative to PM [4]. The simplicity of electric stimuli, the reliance on just two stimulus parameters, and our extensive knowledge of the electrosensory system in Gymnarchus makes it an ideal system in which to apply information theoretic approaches to the encoding, processing, and integration of multiple stimulus parameters. Electrosensory stimuli consisted of low-pass filtered (cutoff = 1-50 Hz), Gaussian-distributed, random AM and PM presented separately and simultaneously (RAM, RPM and joint RAM/RPM). We recorded intracellularly from single Sand O-afferent fibers using sharp electrodes, and used the spike times for stimulus estimation following established methods [2, 5]. In response to RAM stimulation (std. dev. = 10-25%), Oafferents encoded 57.5 +/7.62% (mean +/st. dev.) of the stimulus time course and S-afferents encoded 24.2 +/11.4%. In response to RPM stimulation (std. dev. = 5-20°), S-afferents encoded 73.6 +/11.8% and O-afferents encoded 41.2 +/10.2%. The encoding of both stimulus parameters by both types of afferent suggests a certain degree of ambiguity in the information conveyed by the spike trains. However, selectivity for a particular stimulus parameter improved considerably in the presence of joint RAM/RPM: O-afferents still encoded 54.8 +/9.10% of the RAM stimulus time course, but only 6.81 +/1.94% of the RPM, while S-afferents still encoded 69.5 +/12.9% of the RPM, but only 2.86 +/2.39% of the RAM. We also made intracellular recordings from ELL pyramidal neurons and toral neurons using the whole cell recording technique. The spike trains of AM-sensitive ELL neurons contained significantly less information (7.97 +/4.33%) on the detailed time course of RAM stimuli compared to O-afferents (Mann-Whitney U Test, z=5.04; p<0.000001). Similarly, the spike trains of PM-sensitive ELL neurons contained significantly less information (13.2 +/12.7%) about RPM stimuli than S-afferents (z=4.56; p<0.00001). We tested the alternative hypothesis that these neurons perform a non-linear feature extraction task in which individual spikes signal the occurrence of specific stimulus features, using a modification of previous methods [2]. AM-sensitive ELL neurons reliably signaled the occurrence of either upstrokes or downstrokes in stimulus amplitude during RAM stimulation (probability of misclassification, Perror = 29.5 +/5.25%), while PM-sensitive ELL neurons reliably signaled the occurrence of either phase advances or delays during RPM stimulation (Perror = 24.0 +/9.00%). However, AM-sensitive neurons discriminated stimulus features in RPM above chance levels (Perror = 38.4 +/2.75%), and PM-sensitive neurons similarly discriminated stimulus features in RAM (Perror = 39.2 +/8.47%). In response to joint RAM/RPM stimulation, however, ELL neurons provided reliable discrimination of stimulus features for only one parameter. Thus, the discrimination performance of AM-sensitive neurons in response to joint RAM/RPM stimulation was strongly affected by randomly shuffling the RAM stimulus waveform (sign test, z=3.47; p<0.001), but not the RPM stimulus waveform (z=1.66; p>0.05). Similarly, the discrimination performance of PM-sensitive neurons was strongly affected by randomly shuffling the RPM stimulus waveform (z=4.13; p<0.0001), but not the RAM stimulus waveform (z=0.92; p>0.05). Similar to the primary afferents, then, the selectivity of ELL neurons was enhanced when both stimulus parameters were simultaneously modulated. Some toral neurons responded similarly to ELL neurons, signaling the occurrence of specific features in only one stimulus parameter. By contrast, combination-sensitive toral neurons acted like neuronal AND gates, reliably signaling the occurrence of specific joint modulations in both amplitude and phase. As a result, their discrimination performance during joint RAM/RPM stimulation was greater than during modulation of either parameter alone. In addition, their discrimination performance was strongly affected by randomly shuffling either the RAM or RPM stimulus waveforms. This study is the first to apply information theoretic approaches to the encoding of multiple stimulus parameters in parallel sensory pathways. Our findings underscore the importance of considering each potential source of variation in natural signals, and reveal that limiting stimuli to variation in just one dimension may provide misleading results. [1] Borst, A. and F. Theunissen (1999) Information theory and neural coding. Nature Neurosci 2: 947-957. [2] Gabbiani, F., et al. (1996) From stimulus encoding to feature extraction in weakly electric fish. Nature 384: 564-567. [3] Kawasaki, M. and Y. Guo (1998) Parallel projection of amplitude and phase information from the hindbrain to the midbrain of the African electric fish Gymnarchus niloticus. J Neurosci 18: 75997611. [4] Carlson, B.A. and M. Kawasaki (2004) Non-linear response properties of combination-sensitive electrosensory neurons in the midbrain of Gymnarchus niloticus. J Neurosci 24: 8039-8048. [5] Bialek, W., et al. (1991) Reading a neural code. Science 252: 1854-1857. 73. Responses of Anteroventral Cochlear Nucleus Cells to Tones in the Presence of Noise: Support for Temporal Models of Masked Detection Laurel H. Carney Syracuse University This study focused on responses of cells in the gerbil anteroventral cochlear nucleus (AVCN) to tones added to wideband Gaussian noise at masker levels typical for psychophysical studies of masked detection (e.g. 30 dB SPL spectrum level). Energy-based models for masked detection suggest that neurons encoding tones in the presence of noise should respond with increased rate upon addition of the tone to the masker. However, at mid to high masker levels, most AVCN cells have discharge rates that are saturated, especially at low frequencies. Although some cells do show an increased rate in response to the added tone, this typically occurs at tone levels that far exceed detection thresholds. The most sensitive neural responses to the added tones were decreases in rate observed in 30-35% of AVCN neurons. For some cells, rates significantly decreased at tone levels comparable to detection thresholds. These neural responses were consistent with at least two temporal models for mask ed detection, phase-opponency and envelope-based models. The phase-opponency model (Carney et al., 2002, Acta Acustica United with Acustica 88: 334) predicts a decrease in rate of a coincidence-detecting cell that receives inputs from lowfrequency auditory-nerve (AN) fibers tuned below and above the target tone frequency. The decreased rate is due to a reduction in coincidences between fibers that are phase-locked to different phases of the target frequency. Envelope-based models, such as Richards’ envelope-slope based model (1992, J. Acoust. Soc. Am. 91:3424) also use a decreased value of a decision variable to signal the presence of a tone. A neural implementation of this model involves coincidence detection of matched-frequency inputs combined with a more sluggish same-frequency inhibitory input (Carney et al., 2003, ARO abstract; Nelson and Carney, 2004, J. Acoust. Soc. Am., 116:2173). Because addition of a tone to a noise reduces envelope fluctuations in the output of a narrowband filter tuned to the tone frequency, cells that respond t o envelope fluctuations will have decreased rates upon addition of a tone to the masker noise. Further studies are required to determine which, if either, of these models successfully describes other aspects of AVCN responses, and whether different models are required to explain responses at low and high frequencies and for different cell types. 74. Experience-Dependent Frequency Map Reorganization Endures for at Least 20 Days Rafael A. Carrasco, Roshini Jain, Amanda C. Puckett, Michael P. Kilgard, Pritesh K. Pandya, Christopher L Heydrick, Alyssa McMenamy, Joanna Gibbons, Raluca Moucha University of Texas at Dallas One of the goals of our research effort has been to determine the rules and principles that govern remodeling of tonotopic maps in primary auditory cortex (A1) of young adult rats. It has previously been documented that massive and progressive reorganization of primary auditory cortex (A1) can occur with daily episodic activation of neuromodulatory inputs paired with tonal stimuli. In previous studies, the magnitude of representational reorganization in A1 was determined twenty-four to forty-eight hours after the last conditioning session. An open question is how long these experimental manipulations on cortical representation endure. Therefore, the objective of this experiment was to determine the duration and decay of cortical map reorganization after the cessation of one month of daily conditioning. To determine how long A1 map reorganization endures, a 19 kHz tone was repeatedly paired with electrical activation of the nucleus basalis, located in the basal forebrain, ~350 times a day for one month. Cortical representation of tones was determined by conducting acute mapping experiments at several pre-designated times after the termination of the basal forebrain-acoustic stimulus pairing procedure. In addition to confirming robust map reorganization twenty-four hours after the last conditioning session, our preliminary results reveal long lasting physiological changes in map topography. Specifically, it appears that cortical remodeling induced by our pairing protocol endures for at least twenty days and reestablishes normal tonotopic organization 100 days post stimulation. In conclusion, frequency map reorganization enabled by basal forebrain activation endures for at least twenty days and it returns to baseline 100 days post stimulation. 75. Activity-induced slow disinhibition in a cortical model of spatial working memory. Eugene S Carter, Xiao-Jing Wang Volen Center. Brandeis University. Activity-induced slow disinhibition in a cortical model of spatial working memory. Eugene Carter and Xiao-Jing Wang Volen Center for Complex Systems, Brandeis University, Waltham, MA 02254 Many cognitive functions operate on the timescale of many seconds, subserved by neural dynamics that are intrinsic to the brain rather than driven by sensory inputs. Although electrical events in neurons and synapses are fast (tens of milliseconds), much slower (tens of seconds) intracellular signaling processes interact with the neural spiking activity and play a critical role in cognition, but this important topic has been rarely theoretically studied. One molecular phenomenon that takes place at cognitive time scales is the recently discovered Depolarizationinduced Suppression of Inhibition (DSI), which has been shown to be mediated by endocannabinoid CB1 receptors (Wilson and Nicoll, 2001). CB1 represents one of the most prevalent neuromodulatory receptors in the brain, and cannabinoids (the active ingredient of marijuana, for instance) are known to influence cognitive abilities. The objective of our work is to explore the cellular mechanism and cortical network dynamics through which DSI contributes to cognition. In DSI, depolarization of a principal cell causes a decrease in its GABAergic input (reviewed in Freund et al, 2003). Specifically, in the hippocampus and neocortex endocannabinoid receptors (CB1-Rs) are present on GABAergic synapses that target the perisomatic region of pyramidal cells. A rise in intracellular calcium in a pyramidal cell causes the production of endocannabinoids which travel to the presynaptic terminals from inhibitory cells, leading to a spatially localized reduction in synaptic efficacy of inhibitory transmission. Functionally, a study of behaving rats (Hampson et al, 2001) suggests that cannabinoids can modulate short-term memory in a delayed response task. We have implemented a biophysically-realistic network model of spatial working memory in order to quantitatively investigate the cellular mechanisms of the DSI effect and to explore the functional implications of disinhibition mediated by CB1-Rs in short-term memory. We have constructed a spiking neuron model endowed with a DSI mechanism that matches the experimental data on DSI (Wang and Zucker, 2001). Our main results are fourfold. First, our model successfully reproduces quantitatively the Wang-Zucker data for the time course and magnitude of the DSI-calcium relationship. Secondly, we have incorporated the DSI effect into a recurrent cortical microcircuit model of spatial working memory (Compte et al. 2000), and shown that the spatially restricted DSI produces a local disinhibition, which renders the working memory function more robust. Third, the spatial localization of the disinhibition mechanism prevents adaptation-induced traveling bumps in the working memory network which would otherwise be disruptive to the active maintenance of working memory storage. Fourth, due to its long time constant (tens of seconds), the DSI-mediated disinhibition effect does not decay away in the typical inter-trial-interval in a laboratory experiment with behaving primates. Therefore the working memory network's performance is correlated across several behavioral trials. This effect can be directly tested experimentally both at the behavioral and electrophysiological levels. To summarize, our work shows that due to its activity dependence, spatial localization and slow time constant, endocannabinoid-mediated DSI may contribute to cognition as a mechanism for modulating the robustness of working memory and for temporal integration across trials. References ---------Compte A, Brunel N, Goldman-Rakic PS, Wang XJ. Synaptic mechanisms and network dynamics underlying spatial working memory in a cortical network model. Cereb Cortex. 2000 Sep;10(9):910-23. Freund TF, Katona I, Piomelli D. Role of endogenous cannabinoids in synaptic signaling. Physiol Rev. 2003 Jul;83(3):1017-66. Hampson RE, Deadwyler SA. Cannabinoids reveal the necessity of hippocampal neural encoding for short-term memory in rats. J Neurosci. 2000 Dec 1;20(23): 8932-42. Wang J, Zucker RS. Photolysis-induced suppression of inhibition in rat hippocampal CA1 pyramidal neurons. J Physiol. 2001 Jun 15;533(Pt 3):757-63. Wilson RI, Nicoll RA. Endogenous cannabinoids mediate retrograde signalling at hippocampal synapses. Nature. 2001 Mar 29;410(6828):58892. Supported by the NIMH and the NIDA. 76. Evidence of top-down attentional control in the inferior temporal cortex Yonghong Chen, Steven L. Bressler, Charles E. Schroeder, Mingzhou Ding 1 University of Florida 2 Florida Atlantic University 3 Nathan Kline Institute for Psychiatric Research Introduction: Selective attention improves information processing in the attended domain. During the deployment of covert visual attention topdown biasing signals have been observed in both monkey and human visual cortex in the absence of visual stimulation. The origins of these signals have been the subject of intense discussion and experimentation. Studies using single unit recordings, event-related potentials, functional magnetic resonance imaging, as well as lesion techniques have implicated the prefrontal cortex as a primary source of these signals. In this work, analysis of local field potentials (LFPs) and multiunit activity (MUA) recorded from the inferior temporal (IT) cortex of behaving monkeys provides evidence that IT acts as a way station that directs influences received from prefrontal cortex to lower level visual cortical areas. Patterns of activation in, and interaction among, different layers of IT are shown to be consistent with the known top-down anatomical projection pathways that IT receives from prefrontal cortex, and sends to lower level areas. Experimental Design and Data Collection: Two male macaque monkeys were trained to perform an intermodal selective attention task. Streams of interdigitated auditory and visual stimuli were presented with an irregular interstimulus interval (ISI) (minimum of 350 ms between visual stimuli). The attended modality was alternated across trial blocks. In both auditory and visual modalities, two types of stimuli were presented: a “standard” (either a light flash or a tone) which occurred randomly, on average 86% of the time, and a “deviant,” created by a slight drop in intensity of the standard stimuli on 14% of the trials. Attention to a given sensory modality was controlled by requiring the subject to make a lever-release response to the deviant stimuli in the cued (attended) modality, while ignoring all stimuli in the other modality. Difficulty was varied to maintain responding at ~ 92% correct. During each recording session a linear array electrode with 14 contacts (150 micron spacing) spanning all six cortical laminae was inserted in different parts of the visual system. Laminar local field potentials and multiunit activities were recorded. Data from the inferior temporal cortex were analyzed for this work. Data Analysis: Continuous recordings of LFPs and MUAs were epoched from 200 ms prestimulus to 400 ms poststimulus. Only the prestimulus portion of the epoch was analyzed here to assess the top-down attention effect in the absence of visual stimulation. Since the prestimulus activity was not aligned to any external trigger, we used a sinusoidal matched filter with frequencies ranging from 5 to 60 Hz to realign the trials to achieve averaged LFPs. From the averaged laminar LFPs the second order spatial derivative was computed to obtain the laminar current source density (CSD) profile, revealing the current flow configurations that generate oscillatory activity in each frequency. The generator strength was measured and compared between the attend-visual condition and the ignore-visual condition. MUAs for the attend-visual and ignore-visual conditions were also examined in the prestimulus time period by calculating the percentage modulation as: Finally, to study the patterns of influence between the different layers of IT cortex, we computed Granger Causality spectra in both directions between the superficial layers of IT, which is the main termination point of prefrontal feedback projections, and the deep layers, which is the densest origin of feedback projections into lower level visual areas (e.g. V4 and V2). Results and Discussion: (1) Elevated strength of alpha (8-12Hz) LFP oscillations in the attend-visual condition over the ignore-visual condition was observed in IT of both monkeys. CSD profiles in the alpha range revealed a generator in the superficial layers (2/3 layer) of IT. The anatomical connectivity between the prefrontal cortex and IT suggests that this increased activity may represent the effects of top-down attentional control signals biasing this sensory area to improve the fidelity of information processing. (2) Granger Causality analysis revealed that the superficial layers of IT provide a causal influence on the activities of the deep layers. Furthermore, the laminar profile of MUA activities showed increased firing in the deep layers during the attend-visual conditions. This result is consistent with the idea that top-down attentional control influences from the prefrontal cortex are disseminated to ventral-stream visual cortical areas by way of the inferior te mporal cortex. 77. More than 2: Multiple choice decision-making in humans and monkeys Churchland, Anne, Kiani, Roozbeh, Tam, Marcel, Michael N. Shadlen University of Washington Research from several laboratories is beginning to reveal the cortical machinery that may underlie perceptual decision-making in monkeys. Considerable evidence suggests that the distinctive responses of LIP neurons during a visual direction discrimination task reflect the accumulation of evidence for or against a decision when it is to be communicated by a saccadic eye movement (Roitman and Shadlen 2002; Shadlen and Newsome 2001). The responses of these neurons suggest that evidence for one direction of motion is accumulated to a threshold level, thereby terminating the decision process in a commitment to one of the alternatives. These physiological observations lend support to a class of models known as diffusion to bound or random walk (Link 1992; Ratcliff and Smith 2004). If this mechanism is to be useful for understanding decision-making in general, we must understand its extension to decisions among N>2 choices. As a first step, we have begun to examine the speed and accuracy of decisions among 4 and 8 directions of random dot motion. Monkeys and humans performed a reaction-time version of a direction-discrimination task. Dynamic random dot motion was presented at the fixation point in an aperture with a 5 deg diameter. The strength of motion (difficulty) was controlled by varying the percentage of dots that moved coherently in the same direction (from 0-76.8%). Motion was in one of 2, 4 or 8 possible directions, performed in separate blocks. Whenever ready with a response, subjects terminated the trial by making an eye movement to a peripheral choice target. Increasing motion strength led to improved accuracy and faster reaction times on 2-, 4and 8-choice tasks. Both decision speed and accuracy were affected by the number of choices. For all but the strongest motion strengths, larger numbers of choices led to reduced accuracy and slower reaction time (RT). At high coherences, accuracy and RT were asymptotic. We attempted to account for the data using a model similar to ones that were successful in the 2-choice task. In these models, the decision can be regarded as a race among N diffusion processes or random walks (Usher and McClelland 2001; Usher et al. 1995). Each accumulates noisy evidence in favor of a direction until one reaches a threshold level of accumulated evidence or decision bound. The winning process determines the decision and the decision time. Therefore, the rate of evidence accumulation and the height of the bound determine the speed and accuracy of the decisions, on average. In the 2-choice task, the evidence is thought to be a difference in spike rates between opposing motion sensors (Ditterich et al. 2003). This difference is a random number whose mean is proportional to motion strength and whose variance is approximately constant. This model explains the speed and accuracy functions in a 2-choice task (Palmer and Huk in press). For more than 2 choices, the definition of evidence is not known. We have examined a few candidate models. An attractive option is to use the same evidence used for 2 choices. Specifically, momentary evidence for the right choice is the difference in spike rates from rightward minus leftward-preferring motion-sensing neurons in area MT; evidence for an up choice is the difference in spike rates from upward minus downward neurons, and so forth. We found that this model explains the speed and accuracy data for 4 choices. For the 8-choice task, however, this simple extension fails to explain both the speed and accuracy of decisions. It overestimates reaction time at low motion strengths. Interestingly, shorter reaction times on difficult trials are expected if the bound declines as a function of elapsed viewing time. This may be a sensible strategy because as time passes, it is increasingly likely that the trial will not lead to a reward. Together, these observations imply that our current model for decision making in the 2-choice task,accumulation to bound of a quantity proportional to log likelihood ratiomay extend in a straightforward manner to the 4-choice task. Another simple extension, allowing the bound to decline over the course of a trial (Ditterich in press), may allow the model to explain behavior in the 8-choice task as well. Ditterich J. Computational approaches to visual decision making. In: Perception, Decision and Action: Bridging the Gaps: Wiley, in press. Ditterich J, Mazurek ME, and Shadlen MN. Microstimulation of visual cortex affects the speed of perceptual decisions. Nat Neurosci 6: 891898, 2003. Link SW. The Wave Theory of Difference and Similarity. Hillsdale, NJ: Earlbaum, 1992. Palmer J and Huk AC. The effects of stimulus strength on the speed and accuracy of a perceptual decision. J Vision, in press. Ratcliff R and Smith PL. A comparison of sequential sampling models for two-choice reaction time. Psychol Rev 111: 333-367, 2004. Roitman JD and Shadlen MN. Response of neurons in the lateral intraparietal area during a combined visual discrimination reaction time task. J Neurosci 22: 9475-9489, 2002. Shadlen MN and Newsome WT. Neural basis of a perceptual decision in the parietal cortex (area LIP) of the rhesus monkey. J Neurophysiol 86: 1916-1936, 2001. Usher M and McClelland JL. The time course of perceptual choice: the leaky, competing accumulator model. Psychol Rev 108: 550592, 2001. Usher M, Stemmler M, and Olami Z. Dynamic pattern formation leads to 1/f noise in neural populations. Physical Review Letters 74: 326-329, 1995. 78. The architecture of neural systems that structure sensory information Rhodri Cusack MRC CBU The processes that structure the vast array of information arriving from the senses have a profound affect on how we perceive the world. In addition to their fundamental importance in perception, characterizing these processes is of growing clinical interest, as they are implicated both in autism (“weak central coherence”) and dyslexia. Possible architectures for these processes are evaluated. In vision and audition, information best represented in many distributed neural areas can affect its perceptual structure. For example, in vision, features such as orientation, color, motion, texture, symmetry and form, and in audition, features such as frequency, pitch, location, timbre and phonetic structure, can affect perceptual organization. Information from these many different representations must be combined to determine the organization. Furthermore, as well as influencing it, many of these levels can also be affected by perceptual organization. We propose an integrated competition (IC) model, in which competition for grouping or segregation takes place at many different levels in the perceptual system, and is then integrated across representations. Behavioral and physiological evidence for such a model is reviewed. We demonstrate the architecture using an abstracted but neurally tractable model. We suggest that connections that decouple spike timing (phase) and firing rate (magnitude) allow simple implementation of the IC model. Within region competition is implemented through phase synchronizing connections between neurons representing similar features (e.g., a particular colour or orientation) and phase desynchronizing connections between those representing different features (e.g., different colours). Between region integration is implemented by phase synchronizing connections between corresponding neurons in different layers. This implementation of the IC model naturally accommodates the situation where maps at different levels of the system are organized in different ways, as would be needed, for example, to combine information across tonotopically organised frequency maps (cochlea nucleus; auditory cortex) with a spatial map (colliculus). Note that only a small subset of the neurons connected by phase synchronizing connections is likely to be active in response to any particular stimulus, but excitation will not leak to those that are not if phase and magnitude are decoupled. Such a model naturally provides a solution to the binding problem, so that for example, descending connections that lead to a phase dependent effect on magnitude allow for object-based selective attention. An interesting feature of a model that decouples phase and magnitude information is that it allows different architectures for the different paths: for example, magnitude could be feedforward while phase is interactive, or vice-versa. Although we do not specify the model down to the level of detail of physiologically plausible neurons, models of circuits with the required properties already exist. We also relate integrated competition to the hierarchical decomposition model (HDM), which suggests a relationship between perceptual organization and selective attention. The HDM proposes that initially, perceptual organization forms large scale objects. Once one of these is attended, perceptual organization may then fragment this object into smaller objects, and once one of these is attended in turn, further fragmentation may occur. This interaction between selective attention and perceptual organization is as might be expected if both are implemented via integrated competition and are distributed across similar structures. In summary, a simple integrated competition architecture, consistent with a range of behavioral and physiological knowledge about sensory systems, appears a strong candidate for the implementation of the important processes that structure our sensory world. 79. Emergence of a Dynamic Direction Selective Filter by Optimal Codeing of Natural Time Varying Images Mohammad Dastjerdi, Dawei W. Dong 1 Center for Complex Systems, Florida Atlantic University 2 Center for Complex Systems, Folrida Atlantic University We investigate the possible role of optimal transmission of information in the organization of visual cortex. In the previous studies the emerging spatiotemporal filters through independent component analysis (ICA) of TV videos (vanHateren, Ruderman 1998) are orientation and/or direction selective, similar to those of simple cells. However, because of the eye movement, the retinal input is nonstationary and furthermore the receptive field of LGN changes according to the saccade timing (Dong etal 2003 and Dastjerdi etal 2003). Therefore, we hypothesis that the optimal filters are nonstationary and predict that V1 receptive field changes dynamically relative to saccades. We use the measured eye movements and the derived retinal input of animals watching natural time varying images. We model the retina with a center-surround filter which remove the second order correlations in space. The temporal decorrelation is achieved by using temporal difference LGN filter on the retinal output. The nonstationary property of those cells is taken into account, according to the saccade timing (Dong 2000; Truccolo, Dong 2000). The ICA learning algorithm is used to remove higher order correlations in the LGN output. Our results suggest that the temporal phase difference in LGN output cause the emergence of the direction selective filters similar to those of simple cells (Saul, Hamphery 1990). More importantly, the direction selectivity of the filters changes according to the saccade timing. In particular, there is no direction selectivity right after saccades. Our study suggests that the nonstationary properties of visual input and the brain states such as the oculomotor control play an important role in the organization of visual system. 80. What Parrot Neurons Tell about in Control in Monkeys V. David, J.M. Herrmann, Theo Geisel Max-Planck Institute for Dynamics and Self-Organization, Goettingen Neural spike trains contain a substantial amount of noise and call thus for a stochastic treatment. Because of the high-dimensional data space occurring in multi-unit recordings and the relevance of temporal delays as well as the presence of potentially many hidden sources of complexity of the data, the estimation of joint or conditional probabilities quickly becomes infeasible. The approach of statistical learning theory presents itself as an interesting alternative for explanatory analysis of neural data. While, in this approach, generally it is asked for the complexity of a spike train of finite length, we derive here a complexity measure from parameters of a realistic neural model which is supposed to be capable of reproducing the given spike train. Already for formal neurons the task of minimizing the network complexity is computationally hard, but can be solved approximatively, e.g. in the framework of support vector machines or other iterative learning schemes. The present work starts with the design of a model network which represents a number of features of the microstructure of the cerebral cortex (cf. Izhikevich, 2004) such as the ratio of inhibitory and excitatory cells. In the same spirit we have incorporated short-term plasticity for facilitation and depression (Markam et al., 1998) as well receptor kinetics (Dayan & Abbott, 2001). Furthermore we implemented spike-timing dependent plasticity (STDP) in a additive fashion with a positive drift (Izhikevich, 2004), but also other versions of the STDP learning rule have been studied with respect ot stability and sensitivity. As input and drive to the model network spike-time data from off-line extracellular recordings of pyramidal neurons in somatosensory cortex have been used. The neural data triggered activity of a special group of so-called parrot neurons which served as relay to the units of the model network. Except for this inputs from the data base the parrot neurons are fully identical the N neurons in the model network consisting of N+P units in total. All neurons are subject to neuronal dynamics and plasticity of their synapses, where both the spikes produced by the neural dynamics as well as those from the off-line recording are subject to STDP. The goal is to reproduce the spike recordings after training even for switched-off external input (Maass et al.). Beyond complexity considerations the structural analysis of the evolved interactions in the model is of relevance. In particular spatiotemporal groups of activity can be identified after a period of training. These groups emerged spontaneously due to STDP and consisted out of strongly interconnected excitatory and inhibitory neurons, which fired in highly time-locked patterns. We could investigate a good stability of these responses and a competition between different ones which resulted in repeated emergence and extinction of several different groups of activity. The resulting groups as well as the methodology resembles the set up of (Izhikevich, 2004) which comprises a generalization synfire chains (Abeles et al., 1993; Herrmann et al., 1995). Instead polychronous groups (Izhikevich, 2004) or synfire braids (Bienenstock) are formed. Acknowledgement: This work has been partially supported by the German-Israeli Project Cooperation (DIP). The methods devised here were stimulated by data on monkeys motor behavior provided by M. Abeles. We are grateful to A. Morrison, M. Diesmann, T. Flash, and M. Teicher for stimulating discussions.