Hierarchical Neural Network Model of Global-to-Fine Categorization in Inferior Temporal Cortex

s / Neuroscience Research 68S (2010) e223–e334 e323 predictive coding model describes the hierarchical processing in the ventral pathway. doi:10.1016/j.neures.2010.07.1429 P2-q09 Hierarchical Neural Network Model of Global-toFine Categorization in Inferior Temporal Cortex Ryuzou Nakata 1 , Masaki Ogino 1, Minoru Asada 1,2 1 Department of Adaptive Machine Systems, Graduate School of Engineering, Osaka University 2 Japan Science and Technology Agency, ERATO, Asada Synergistic Intelligence Project Humans can classify various objects in the environment. However, the neuronal mechanism of the information processing on the categorization is not clear. Sugase et al. (99) analyzed the neuron population responses in the inferior temporal cortex of the monkeys while presenting the face images. They found that the global categorization of species was represented in the earlier part of the neuron population responses, while the fine categorization of individual identity or emotional state was represented in the later part of the neuron population responses. The results suggested that the hierarchical relationship of the visual stimulus was represented by the neuron population responses of the inferior temporal cortex. In this study, the hierarchical relationship of the global categorization and the fine categorization is modeled, based on the hierarchical neural network, Convolutional Restricted Boltzmann Machine (CRBM). The CRBM is the associative memory model and can learn from lower-order to higherorder representation in the selforganized manner. The global categorization and the fine one are formed only by changing the sparseness. In the experiment, each categorization was formed with the images of human and monkey faces that were used in Sugase’s study. As the result, the faces were classified into the human and the monkey categories in the global categorization. The monkeys were classified into the emotional category and the humans were classified into the individual identity category in the fine categorization. This results are similar to the analytical results of the neuron population responses of the monkeys. In addition, the fine information was obtained early by using the global information as a tag. This results suggest that the hierarchical relationship of the categorization contributes to the improvement of the recognition speed. doi:10.1016/j.neures.2010.07.1430 P2-q10 Rate difference coding by spike timing dependent plasticity Kazuhisa Fujita Department of Electronics and Control Engineering, Tsuyama National College of Technology Spike timing dependent plasticity (STDP) is found in various areas of the brain, visual cortex, hippocampus and hindbrain of electric fish, etc. The synaptic modification by STDP depends on time difference between presynaptic spike arrival and postsynaptic firing. If presynaptic neuron fires earlier than postsynaptic neuron dose, synaptic weight is strengthened. If postsynaptic neuron fires earlier than presynaptic neuron dose, synaptic weight is weakened. This learning rule is one example of various rules (This rule is hippocampal type). If electric fish type rule applies, reversed plasticity occurs. Changes of synaptic efficiency precisely depend on timing of preand postsynaptic spikes under STDP. Because of this precise dependence, it is thought that STDP plays important role in temporal processing of stimuli. Temporal processing by STDP is well known. However, the role of STDP on rate processing is not enough understood. In this study, we found that difference in firing rate between preand postsynaptic neurons is learned based on STDP. We show rate difference coding provide various functions for feedback network and interconnected network, using computer simulation. In case that learning rule is electric fish type, feedback network provides the function of gain control and interconnected network provides the function of spatial low pass filter. In case that learning rule is hippocampal type, feedback network provides the function of signal enhancement, and interconnected network provides the function of spatial high-pass filter. We propose that STDP provides ability of coding difference in firing rate between preand postsynaptic neurons and show that rate difference coding provides various function. These results suggest that STDP would play the important role not only in temporal processing but also in spatial processing. doi:10.1016/j.neures.2010.07.1431 P2-q11 Modulation of corticofugal receptive field induced by synaptic plasticity in bat auditory system Yoshihiro Nagase 1 , Yoshiki Kashimori 1,2 1 Department of Human Media Systems, Graduate School of Information Systems, University of Electro-Communications 2 Department of Engineering Science, University of Electro-Communications, Chofu, Tokyo 182-8585 Japan Most species of bat making echolocation use Doppler-shifted frequency of ultrasonic echo pulse to measure the velocity of target. To perform the fine-frequency analysis, the feedback signals from cortex to subcortical and peripheral areas are needed. The feedback signals are known to modulate the tuning property of subcortical neurons. Xia and Suga (2002) have shown on an intriguing property of feedback signals that the electric stimulation of cortical neurons evokes the best frequency(BF) shifts of subcortical neurons away from the BF of the stimulated cortical neuron(centrifugal BF shift) and bucuculine(an antagonist of inhibitory GABA receptors) applied to the stimulation site changes the centrifugal BF shifts into the BF shifts towards the BF of stimulated cortical neurons(centripetal BF shift). Although these BF shifts are generated by the feedback signals from cortical neurons to subcortical neurons, it is not yet clear how the feedback signals determine the direction of BF shift. To address this issue, we present a neural model for detecting Dopplershifted frequency of echo sound reflecting from a target. The network model for detecting the Doppler-shifted frequency consists of cochlea, inferior colliculus, and Doppler-shifted constant frequency area, each of which is a linear array of frequency-tuned neurons. The neurons in the three layers are tuned in to specific echo frequency ranging from 60.0 to 63.0 kHz, which corresponds to the frequency range of the second harmonics. The bat uses the Doppler-shifted frequency of echo sound to detect the relative velocity of target. We show here that the receptive field of cortical neurons is modulated by STDP learning, depending on the stimulus context. This indicates that the tuning properties of subcortical neurons change on-line. We also propose a functional role of the BF shift in extracting the information of target velocity from background signal reflecting from trees. Reference Xiao,Z.and Suga,N.PNAS99(2002)15734-15748. doi:10.1016/j.neures.2010.07.1432 P2-q12 Noise-robust realization of Turing-complete cellular automata by using an array of Hopfield models combined with multilayered perceptrons Makito Oku 1 , Kazuyuki Aihara 1,2 1 Grad School of Info Sci & Tech, Univ of Tokyo, Tokyo 2 Inst Indust Sci, Univ of Tokyo, Tokyo We show how stochastic binary-state neural network with some sort of redundancy as well as stability in the dynamics can have the computational ability comparable to Turing machines (TMs) by showing a concrete construction method. The basic idea is to make use of the Hopfield model as a mean of noise robust representation of symbols. Two orthogonal patterns are stored in a Hopfield network, which are related to 1 and 0 respectively; thus, a bit string can be represented by an array of Hopfield models. It is impossible, however, to directly regard this array of Hopfield models as the tape of a TM, because the mobile head cannot be constructed by a neural network. Therefore, we transform the computation of a TM into that of a Turing complete cellular automaton. To realize the dynamics of a cellular automaton in the framework of neural network, we consider a module which we call a ‘cell’. Each cell consists of two parts: a Hopfield neural network model and a multilayered perceptron. The first part represent a binary symbol, while the second part nonlinearly transforms the inputs from neighboring cells according to the update rule of a cellular automaton, and transmits its output to the first part in the same cell. An array of such cells is used to simulate the Rule 110 cellular automaton, which is one of the elementary cellular automata and has been proven to be Turing complete. Even when all the units of neural networks are replaced by stochastic binary ones, the proposed model can simulate the Rule 110 with high accuracy. doi:10.1016/j.neures.2010.07.1433