神経システムの学習論 Learning theory and analysis of neural systems
O3-9-5-1 培養神経回路網は電気パルス刺激の高確率な隠れ信号源に適応する Cultured neuronal networks adapt to a higher-probability hidden signal source of electrical pulse stimulations ○磯村拓哉1, 小谷潔1, 神保泰彦1 ○Takuya Isomura1, Kiyoshi Kotani1, Yasuhiko Jimbo1 東京大学大学院 新領域創成科学研究科 人間環境学1 Dept Human Envir, Univ of Tokyo, Chiba1 Principal component analysis (PCA) is a type of unsupervised learning that distinguishes signals into several components and is widely used as a signal processing method. Earlier studies reported the possibility that neurons in the cortex use PCA-like signal processing for sensory information recognition. Numerous computational studies have demonstrated that simulated neural networks are capable of performing PCA; however, only a few experimental researches have discussed neural bases of PCA-like learning. We demonstrated an example of neural adaptation to a higher-probability hidden signal source of multiple input signals using cultured neurons. Embryonic rat cortical cells were taken out and dissociated into single cells. Approximately 500,000 of the dissociated cells were seeded on microelectrode array (MEA) dishes and cultivated more than 3 weeks before experiments. We induced the neurons into electrical pulse trains with 1 s interval over 10,000 times from 32 electrodes. The trains were constructed from two independent binary signal sources u1(t) and u2(t) by the following manner: the values of a half of input trains were randomly selected by u1(t) with a probability of 3/4 or u2(t) with that of 1/4 at each time periods. The values of the rest of trains were randomly selected by u1(t) with a probability of 1/4 or u2(t) with that of 3/4. Evoked responses to the input trains were recorded from 64 electrodes of MEA. We observed the increase of evoked responses of neuronal groups at the several stimulation sites during stimulation term. Cross-correlations of responses also increased between two neuronal groups at different stimulus electrodes when the electrodes were stimulated with the same merged balance of signals. These results suggest that continuous stimulation induced synaptic plasticity of cultured neuronal networks and support the hypothesis that neuronal networks can estimate the highest-probability or largest hidden signal source of input signals.	O3-9-5-2 視覚野の複雑細胞と聴覚野のピッチ細胞が相同である可能性について A possible analogy between complex cells in visual cortex and pitch cells in auditory cortex ○寺島裕貴1,2, 岡田真人1,3 ○Hiroki Terashima1,2, Masato Okada1,3 東京大院・新領域・複雑理工1, 学振2, 理研BSI3 Dept Complex Sci and Eng, Univ of Tokyo, JAPAN1, JSPS, Tokyo, JAPAN2, RIKEN BSI, Saitama, JAPAN3 The complex cells were found in the visual cortex more than fifty years ago. The concept of nonlinear responses with phase invariance has been firmly established and successfully modelled using natural image statistics [1]. However, analogous discussions have been lacking in modalities other than vision, despite of the anatomical uniformity across sensory cortices. A kind of universality in them has been suggested by successful applications of models for visual cortex to auditory cortex [2, 3], although they studied linear receptive fields. We applied a nonlinear model of visual complex cells [1] to natural sounds; receptive fields of the "complex cells" adapted to sound statistics typically had multiple peaks at harmonic frequencies. Moreover, some of them resembled pitch cells found in auditory cortex that nonlinearly respond to harmonic complex tones in a way similar to a psychoacoustic phenomenon called "missing fundamental" [4]. The result suggests that the pitch cells might be analogous to the complex cells [5]: the visual complex cells are invariant to the phase, whereas the pitch cells are to the harmonic spectral composition. [1] A Hyvärinen and PO Hoyer (2001) Vision Research 41(18): 2413-2423. [2] H Terashima and H Hosoya (2009) Network 20(4): 253-267. [3] H Terashima et al. (2013) Neurocomputing 103: 14-21. [4] D Bendor and X Wang (2005) Nature 436(7054): 1161-1165. [5] H Terashima and M Okada (2012) Advances in Neural Information Processing Systems 25: 2321-2329.
O3-9-5-3 認知情報処理時の大域的な神経活動同期ネットワークにおける領野の重要度のマルコフ連鎖モデリングを用いた定量化 Quantifying importance of brain regions in large-scale neural synchronization networks in cognitive processing with a Markov chain modeling ○宇野裕1, 水野佑治1,2, 北城圭一1,2,3,4 ○Yutaka Uno1, Yuji Mizuno1,2, Keiichi Kitajo1,2,3,4 理研 BSI BTCC RBIP1, 農工大院 工学 電子情報工学2, 理研 BSI ABSP3, JST さきがけ4 RIKEN BSI BTCC RBIP, Japan1, Tokyo University of Agriculture and Technology, Electronic and Information Engineering, Japan.2, RIKEN BSI ABSP, Japan3, PRESTO, Japan Science and Technology Agency (JST), Japan4 Some previous studies have claimed that long-distance synchronization of neural oscillations plays an important role in dynamically linking brain regions during cognitive processing. Most previous studies have analyzed synchronization between a pair of brain regions and few studies have analyzed the structure of entire large-scale synchronization networks. We have, therefore, developed an analysis to quantify importance of brain regions from the connectivity of the large-scale synchronization networks. First, we assumed that synchronization between two brain regions represents the link mediating information flow and connection between them. Based on the assumption, we estimated the network structure constructed from synchronization links in terms of information flow associated cognitive processes. We modeled information flow in the network with the Markov chain and predicted the stationary state to estimate the important brain regions. In order to reveal whether this method is applicable to brain data, we performed experiments with simulated datasets. Simulated datasets were computed from a forward model of MEG measurement. We found that the proposed method successfully estimated important brain regions from the simulated datasets.Finally, we analyzed real EEG data in apparent motion tasks using this method. We succeed in extracting import brain regions which should be functionally relevant. We speculate that task relevant network connectivity is configured as the structure of large-scale synchronization networks.	O3-9-5-4 カルシウムイメージングデータを用いたグリア・ニューロン回路網のシステム同定 System identification of glia-neuron networks with calcium imaging data ○中江健1, 池谷裕二2, 石井信1 ○Ken Nakae1, Yuji Ikegaya2, Shin Ishii1 京都大学院・情報1, 東京大院・薬・薬品作用2 Grad Sch Infomatics, Kyoto Univ, Kyoto, Japan1, Lab. Chem. Pharmacol, Grad Sch Pharm, Univ Tokyo, Japan2 Information processing in the brain largely depends on neural communication through synaptic connections. Recently, many researchers have reported that synaptic and neuronal activities are dynamically modulated by the activity of glial cells, astrocytes that envelope neighboring synapses. In these studies, artificial stimulations to the glial cell results in exocytosis of gliotransmitter from the cell, which modulates post-synaptic current or increases neuronal excitability. This indicates that glia-neuron networks are necessary for understanding information processing in the brain. However, the glial effects in natural conditions have been controversial because of the artificial stimulations in these experiments. To discuss these effects in natural conditions, we propose a statistical method for estimating both synaptic connectivity and glial modulation by observations of glial and neuronal activities as calcium imaging data. In this study, we show that glia-neuron networks are successfully reconstructed from artificial data generated by a computational model, which is different from our generative model. Applying our method to calcium imaging data, we identify a glia-neuron network in a natural condition, where neuronal and glial activities are spontaneous. As we know, our method is a first attempt for system identification of glia-neuron networks. Moreover, our method is applicable for reading neural code and information flow in neural ensemble, when we consider the glial activity in our method as an external stimulus. Our method is closely related to reverse correlation for neural coding, dynamic causal model used in analysis of fMRI data, generalized linear model for estimating neural connectivity.

上部に戻る

前に戻る