一般口演（Oral）

Oral Neuroinformatics and large-scale simulations 一般口演神経回路モデルと人工知能

7月27日（土）8:45～9:00　第7会場（朱鷺メッセ　2F　201B） 3O-07m1-1 深層畳み込みニューラルネットワーク内ユニットの受容野空間構造の可視化 Tomoyuki Naito（内藤智之）¹，Yoshiyuki Shiraishi（白石祥之）²，Tatsuya Mori（森達哉）²，Hiromichi Sato（佐藤宏道）^1,2 ¹大阪大院医認知行動科学 ²大阪大院生命機能 Convolutional Neural Network (CNN) is a neural network model for visual pattern recognition. Alghough, Deep Convolutional Neural Network (DCNN) has been a de facto standard model for image recognition or image cathegorization, internal information processing and its representation in the DCNN are unclear. In this study, we examined systems neuroscience methods for visualizing information processing in the DCNN. First, orientation selectivity of convolutional units to sinusoidal grating rating stimuli and natural image selectivity of each unit were measured. Then, we evaluated the sharpness and multimodality of orientation selectivity by Gaussian fitting and dip test. Natural image selectivity of each unit was also assessed. We found that in the higher layer of DCNN, the more units had multimodality of orientation selectivity and natural image selectivity, which were consistent with properties of neurons in visual ventral pathway. In contrast, in the higher layer of DCNN, the more units exhibited stronger orientation selectivity, which was inconsistent with the properties of visual ventral pathway. Then, spatial profiles of receptive field (RF) of all convolutional units were reconstructed by both spike-triggered average (STA) and spike-triggered covariance (STC) analysis. Validity of reconstructed RFs was evaluated by predictability of orientation tuning and natural images category selectivity using reconstructed RFs. We found that even in the higher layers in DCNN roughly 30~50% of response variance was captured by receptive fields reconstructed by STC analysis. Taken together, our results suggested that although the properties of DCNN unit were similar with those of neurons in the visual ventral pathway, there was a clear discrepancy between the two. We also found that the methodology of systems neuroscience is useful for visualizing internal information processing and its representation in the DCNN. The reconstructe receptive field of DCNN units may allows us to understand the functional sifnificance of the unit and layers.		7月27日（土）9:00～9:15　第7会場（朱鷺メッセ　2F　201B） 3O-07m1-2 Evaluating CNNs as a model of face processing network of the macaque. Rajani Raman（Raman Rajani），Haruo Hosoya（Hosoya Haruo） Advanced Telecommunication Research Institute International (ATR) Convolutional neural networks (CNN) trained for object recognition task has shown an interesting ability to predict some stimulus-response relationship in ventral visual areas despite being not directly optimized to fit neural data (Cadieu et al., 2014; Yamins et al., 2014). However, whether such predictability holds for the face processing system, a subsystem of IT, is not clear. In this study, we evaluated CNN in the light of reported tuning properties of the face-processing network of the macaque. Specifically, we first trained an Alexnet-type CNN model with natural face images. Then, using natural and different parameterized face stimuli, we simulated past four experiments, especially related to middle face patches (ML): (1) contrast-polarity tuning (Ohayon et al., 2012), (2) view-specificity and size invariance of face-selectivity (Freiwald and Tsao, 2010), (3) Shape-features selectivity (Chang and Tsao, 2017), and (4) tuning-properties to facial geometry and its summary statistics (Freiwald et al., 2009). Furthermore, we also compare CNNs findings to the results of view-invariance (partial and full) and appearance-selectivity found in AM (and AL). We found that face-selective units in higher (in particular, fully connected) layers had size invariance and did show view-invariant identity and appearance-selectivity similar (comparable) to typical macaque AM face neurons. However, we did not find any layer equivalent to ML simultaneously showing associated properties: contrast polarity tuning, tuning to a small number of facial features (mainly related to large and global features), view specificity, and shape-features selectivity. We further repeated these simulations on high-performance pre-trained networks (vgg-face and alexnet) and a number of other CNNs, varying in architecture (deeper or shallower) or the training set (ImageNet or flower images); the results were similar. Thus, none of the CNNs that we examined simultaneously captured all the investigated properties of the face-processing network. Taken together, despite the prevailing view linking CNN and IT, our results indicate that CNN may capture only partial properties of the macaque face processing system and thus require a more comprehensive model to fully explain it.

7月27日（土）9:15～9:30　第7会場（朱鷺メッセ　2F　201B） 3O-07m1-3 スパースモデリングに基づいて神経細胞非線形ダイナミクスを推定する Toshiaki Omori（大森敏明），Shinya Otsuka（大塚慎也）神戸大学大学院工学研究科電気電子工学専攻知的学習論分野 Elucidating neural dynamics is one of the important subjects in neuroscience. In particular, nonlinear dynamical models of single neurons have been proposed using conductance-based neuron models. The conductance-based neuron models are known to be biophysically plausible models with multiple types of nonlinear membrane currents, such as sodium, potassium, and calcium currents. To elucidate nonlinear dynamics of single neurons, it is important to estimate not only the values of conductances of membrane currents but also the types of membrane currents which specific neurons under investigation have. Using high performance computing, exhaustive parameter searches using large scale simulation have been performed to reveal optimal parameters with which neuronal model can reproduce experimental data (Prinz, Bucher and Marder, Nature Neurosci. (2004); Taylor, Goillard, and Marder, J. Neurosci. (2009)). Since computational cost grows exponentially with increase of parameters, there exists difficulty in exhaustive search approaches. Data driven-approach to estimating nonlinear neuronal dynamics has received much attention and several algorithms for estimating conductance-based neuron model have been proposed (for example, Huys and Paninski (2009); Omori and Hukushima (2016)). Efficient parameter estimations are realized without setting parameter ranges in advance. However, in the previous data-driven approach, estimation of membrane conductances has been performed by assuming that the types of membrane currents are known. In this study, we propose a method to estimate neuronal dynamics based on sparse modeling (Otsuka and Omori, Neural Networks, in press.). A sparse modeling framework is derived from conductance-based neuron model with a number of nonlinear ion channels. To realize successful estimation even when scales of membrane currents and maximum conductances are different, we propose a novel sparse estimation that automatically sets different sparsity levels for respective membrane currents. Using the proposed algorithm, we showed that only necessary membrane currents are extracted from many candidate membrane currents with different scales; non-zero/zero correspondence in membrane conductances is accurately estimated in the proposed method. These results showed that our proposed sparse method successfully extracts nonlinear membrane dynamics from observable data.		7月27日（土）9:30～9:45　第7会場（朱鷺メッセ　2F　201B） 3O-07m1-4 A spiking neural network model of the whole-brain circuit linking basal ganglia, cerebellum and cortex Carlos Enrique Gutierrez（Gutierrez Carlos Enrique）¹，Jun Igarashi（Igarashi Jun）²，Zhe Sun（Sun Zhe）²，Jean Lienard（Lienard Jean）¹，Hiroshi Yamaura（Yamaura Hiroshi）⁹，Tadashi Yamazaki（Yamazaki Tadashi）⁹，Heidarinejad Morteza（Morteza Heidarinejad）²，Benoit Girard（Girard Benoit）³，Gordon Arbuthnott（Arbuthnott Gordon）⁴，Hans Plesser（Plesser Hans）^6,10，Markus Diesmann（Diesmann Markus）^5,7,8，Kenji Doya（Doya Kenji）¹ ¹Neural Computation Unit, OIST, Okinawa, Japan ²Computational Engineering Applications Unit, RIKEN, Wako, Japan ³Sorbonne Universite, UPMC Univ Paris 06, CNRS, Institut des Systemes Intelligents et de Robotique (ISIR), Paris, France ⁴Brain Mechanism for Behaviour Unit, OIST, Okinawa, Japan ⁵INM-6, IAS-6, INM-10, Julich Research Centre, Julich, Germany ⁶INM-6, Julich Research Centre, Julich, Germany ⁷Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany ⁸Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, Aachen, Germany ⁹Mathematical Information Science Program, Department of Computer and Network Engineering, Graduate School of Informatics and Engineering, The University of Electro-Communications ¹⁰Faculty of Science and Technology, Norwegian University of Life Sciences, As, Norway The neural circuit linking the basal ganglia, the cerebellum and the cortex through the thalamus is supposed to play essential roles in motor and cognitive action selection and control. However, how such functions are realized by multiple loop circuits with neurons of multiple types is still unknown. In order to investigate the dynamic nature of the whole-brain network, we built biologically constrained spiking neural network models of the basal ganglia [1,2,3], cerebellum, thalamus, and the cortex [4,5] and ran an integrated simulation using K supercomputer. Models correspond to 1x1mm2 of cortical surface and replicate basic functions like resting state. We then increased the number of neurons and cortical surface to 5x5, 7x7, 9x9mm2. We simulated the different scales using NEST 2.16.0 [6,7,9], during 1 biological second of time, on the K supercomputer [8]. Simulations used a hybrid parallelization approach and computational efficiency was evaluated in terms of building and simulation time, and memory consumption. Moreover, we evaluated the model's biological plausibility, particularly the Basal Ganglia (BG) circuit during resting state and action selection functions, and generally the whole-brain resting state. Reproduction of basic functions in a whole-brain model will allow us to gradually study more features, like the closed loop BG-thalamo-cortical dynamics, and reproduce more complex behaviors. [1] Lienard, Jean, and Benoit Girard. A biologically constrained model of the whole basal ganglia addressing the paradoxes of connections and selection; Journal of computational neuroscience 36.3 (2014): 445-468. [2] Lienard et al., (2018). SBDM/SfN. [3] Gutierrez et al., (2018). AINI. [4] Igarashi J., Moren J., Yoshimoto J., Doya K. Selective activation of columnar neural population by lateral inhibition in a realistic model of primary motor cortex, Abstract of 44th Annual Meeting of the Society for Neuroscience, (2014) [5] Zhe Sun and Igarashi, Abstract of JNNS2018, (2018) [6] Gewaltig, Marc-Oliver, and Markus Diesmann. ;Nest (neural simulation tool).; Scholarpedia 2.4 (2007): 1430. [7] Linssen, Charl et al. (2018). NEST 2.16.0. Zenodo. 10.5281/zenodo.1400175. [8] Miyazaki, Hiroyuki, et al. Overview of the K computer system; Fujitsu Sci. Tech. J 48.3 (2012): 302-309. [9] Jordan J, et al. (2018) Extremely Scalable Spiking Neuronal Network Simulation Code: From Laptops to Exascale Computers. Front. Neuroinform. 12:2. doi: 10.3389/fninf.2018.00002