学習理論 Learning Theory
P1-2-217 抑制性シナプスによる神経ネットワークの非対称の補償 Inhibitory synapses compensate asymmetry of brain network ○桜井伊知郎1, 久保田繁2, 庭野道夫1 ○Ichiro Sakurai1, Shigeru Kubota2, Michio Niwano1 東北大学電気通信研究所1, 山形大学大学院理工学研究科2 Research Institute of Electrical Communication, Tohoku University, Sendai, Japan1, Graduate School of Science and Engineering, Yamagata University, Yamagata, Japan2 Introduction Spike-timing-dependent plasticity (STDP) is one possible mechanism of response selectivity. Previous animal experiment showed that some neurons in the visual cortex dominantly respond to right or left eye, and other neurons equally respond to both eyes. However, in theoretical study, network parameters of two presynaptic groups such as cell number, firing correlation, or firing rate are symmetric. In the symmetric condition, postsynaptic response after STDP learning can be always asymmetric, but it does not consistent with the actual response in the visual cortex. In this study, we constructed a computational model and examined how the network asymmetry is compensated.MethodWe used an integrate-and-fire neuron model, which receives inputs from two groups of excitatory synapses and two groups of inhibitory synapses. In the both excitatory input groups, each cell has firing correlation. One inhibitory cell group fires after one excitatory cell group or the postsynaptic cell, which we call lateral, and feedback inhibition, respectively. All parameters of synapses and inhibitory synapses, such as synapse number, correlation intensity, and activation rate are variable independently between two groups. Excitatory synapses are modified by STDP, and inhibitory synapses are fixed.ResultsWe show that when a parameter of excitatory pre synapses between two groups is different, the postsynaptic cell dominantly responds to the presynaptic group which has larger parameter. One-sided lateral inhibition has similar effect.Under the asymmetry of excitatory synapses, we show that adding weak feedback inhibition eliminates the asymmetry of postsynaptic response. Stronger feedback inhibition can be inverse the postsynaptic response, it is celled anti-Hebbian. Interestingly, this effect does not depend on whether feedback inhibition itself is symmetric or not.	P1-2-218 教師付き学習のための強化学習による教師信号の生成 Reinforcement learning of teacher signals for supervised learning ○稲葉学1, 山崎匡1 ○Manabu Inaba1, Tadashi Yamazaki1 電気通信大学 大学院 情報理工学研究科 情報・通信工学専攻1 Graduate School of Informatics and Engineering, The University of Electro-Communications, Tokyo1 Supervised learning is a learning scheme in which a learner is given a context signal and a teacher signal. The learner is asked to learn from the teacher signal and reproduce the same signal in response to the context signal. Supervised learning works properly only when good teacher signals are fed to the learner, because the learner completely relies on the teacher signals in supervised learning. Unfortunately, it is not always the case. To address this problem, in this study, we consider the combination of reinforcement learning with supervised learning. Specifically, we propose to learn teacher signals by reinforcement learning and feed them to a supervised learning machine. Temporal difference (TD) method is a reinforcement learning scheme and is thought to be implemented on the neural circuit of the basal ganglia. Electrophysiological experiments have indicated that the neural activity in ventral tegmental area (VTA), a part of the basal ganglia, encodes the value of delta, the prediction error of TD-learning (Shultz W. Predictive reward signal of dopamine neurons. J Neurophysiol 80:1-27, 1998). Furthermore, anatomical studies have indicated the existence of dopaminergic inputs from VTA to the inferior olive, which provides teacher signals to the cerebellum (Winship IR, Pakan JM, Todd KG, Wong-Wylie DR. A comparison of ventral tegmental neurons projecting to optic flow regions of the inferior olive vs. the hippocampal formation. Neuroscience 141:463-473, 2006). Therefore, the cerebellum could use information of TD-prediction error as teacher signals. We adopted the combined scheme to gain adaptation of voluntary eye movement called double-step saccade adaptation, in which both the cerebellum and basal ganglia play important roles. Our proposed scheme successfully reproduced the gain increase adaptation. These results suggest that the combination provides a powerful learning scheme.
P1-2-219 脳ネットワークからの重なりと階層とを有するコミュニティ構造の抽出 Detecting overlapping and hierarchical community structure of brain networks ○岡本洋1,2 ○Hiroshi Okamoto1,2 富士ゼロックス（株）研究技術開発本部1, 理化学研究所　脳科学総合研究センター2 Research & Technology Group, Fuji Xerox Co., Ltd., Kanagawa, Japan1, RIKEN Brain Science Institute, Saitama, Japan2 In the literature of network science, 'community' means a group of nodes that are densely connected within the group but are less connected with nodes outside the group. Community structure is a hallmark of a variety of social, biological and engineering networks. Development of methods for detecting community structure of a network has been a focus of network science in the last decade [1]. Especially in neuroscience there have been growing interests in analyzing community structure of brain networks in order to reveal functional modules in information processing in the brain. Community structure of brain networks will be characterized by the following properties: Each community has some overlaps with other communities; communities are hierarchically organized, that is, within each module there will be a set of sub-modules [2]. Although a number of mathematical and computational approaches of analyzing community structure of a network have been proposed up to now, most of them cannot deal with community structure having these properties. Here we propose a method for detecting overlapping and hierarchical community stricture of a network. This method is founded on the Bayesian formulation of PageRank algorithm. Community extraction from human cortical and C. elegans networks by this method is also discussed. [1] Newman, M. E. Communities, modules and large-scale structure in networks. Nature Physics Vol. 8, pp. 25-31 (2012). [2] Meunier, D., Lambiotte, R. & Bullmore, E. T. Modular and hierarchically modular organization of brain networks. Frontiers in Neuroscience Vol. 4, pp. 1-11 (2010).

上部に戻る

前に戻る