一般口演（Oral）

Oral Neural network models and artificial intelligence 一般口演神経情報学と大規模シミュレーション

7月27日（土）9:45～10:00　第7会場（朱鷺メッセ　2F　201B） 3O-07m2-1 EEG-fMRI同時計測データを用いたSPLICEフィルターによる大規模ネットワークの推定 Takeshi Ogawa（小川剛史）¹，Hiroki Moriya（守谷大樹）¹，Nobuo Hiroe（廣江総雄）²，Takashi Yamada（山田貴志）³，Motoaki Kawanabe（川鍋一晃）^1,4，Jun-ichiro Hirayama（平山淳一郎）^2,4 ¹ATR認知機構研 ²ATR脳情報解析研 ³ATR脳情報研 ⁴理研AIP Functional connectivity at rest is one of the hot topic in the neuroscience to understand cooperative processes as several networks consisting of brain regions. Network extraction methods from functional magnetic resonance imaging (fMRI) have been established not only anatomical property, but also data-driven approach such as ICA, graph theory, or clustering. On the other hand, electroencephalogram (EEG) is low spatial resolution and signal-to-noise ratio relative to fMRI, therefore an inverse source estimation to identify the cortical activity is rather costly and even error-prone due to its sill-posed nature. Thus, it is difficult to directly extract the network activity from EEG signals similar to fMRI signals. Here, we attempt to propose a practical blind source separation (BSS) approach to exploring EEG source-space network activity. In stead of conventional approach simply applying independent component analysis (ICA) or principle component analysis (PCA), we examine the use of our novel principled hierarchical extension of sensor-space ICA, referred to as stacked pooling and linear components estimation (SPLICE; Hirayama et al., 2017). In the minimal case, SPLICE learns two ICA-like linear generative layers and intermediate ""pooling"" layer, with those layers jointly and mutually optimized. SPLICE is data-drive approach to extract network signature from EEG data. We could extract large-scale network from EEG signals with the data-driven approach that timeseries of the modules estimated by SPLICE were significantly correlated with fMRI network. Considering hierarchical structure not only fMRI signals but also EEG components, we could optimize a spatiotemporal filter. Output of the modules were significantly correlated with fMRI signals not only in superficial brain regions, but also in mid brain regions containing DMN. This method may be helpful to understand network process with high-temporal resolution such as EEG/MEG signals.		7月27日（土）10:00～10:15　第7会場（朱鷺メッセ　2F　201B） 3O-07m2-2 ヒトコネクトーム上のネットワークダイナミクスを決定づける構造的要素 Makoto Fukushima（福嶋誠）^1,2，Olaf Sporns（Sporns Olaf）^3,4 ¹情報通信研究機構脳情報通信融合研究センター ²大阪大学大学院生命機能研究科 ³Department of Psychological and Brain Sciences, Indiana University, Bloomington, IN, USA ⁴Indiana University Network Science Institute, Bloomington, IN, USA Many empirical properties about co-activation patterns (functional connectivity; FC) in the resting human brain have been investigated by modeling large-scale neural dynamics on the human connectome (structural connectivity; SC). An important example that recently received attention is the dynamics of global network topology of FC, such as dynamic fluctuations between segregated and integrated network topology over the course of time-resolved FC typically estimated using sliding windows. In our previous study, we demonstrated that the magnitude of dynamic fluctuations in global network topology of empirical FC was largely reproduced by simulated functional magnetic resonance imaging (fMRI) signal generated from neural oscillator models coupled based on SC. However, while the reproducibility of dynamic fluctuations was examined using generative simulation models, the generative mechanism of dynamic fluctuations remains unclear. To approach this mechanism, we explore in this study which network properties of SC are vital to the emergence of dynamic fluctuations in global network topology of simulated FC. We tested three key ingredients of SC networks, namely, hubs, communities, and geometry. We examined their relative contributions to the dynamic fluctuations, by simulating fMRI signal with random SC surrogates preserving the hub organization, the community structure, or the weight-length relationship (geometry). We found that the magnitude of dynamic fluctuations in global network topology of simulated FC was the greatest with surrogate SC data preserving the weight-length relationship and was the second greatest with the surrogates preserving the community structure. We also found that the magnitudes of dynamic fluctuations simulated with the hub-, community-, or geometry-preserving surrogates, as well as surrogates preserving all three network properties, were smaller than the magnitude of dynamic fluctuations simulated with actual SC data. These results suggest that geometry of SC contributes the most to the dynamic fluctuations in comparison with hubs and communities of SC, but some other factors in SC may also be important for shaping the dynamic fluctuations in global FC network topology. The present study offers new insights into how SC serves to generate complex network dynamics in the resting human brain.

7月27日（土）10:15～10:30　第7会場（朱鷺メッセ　2F　201B） 3O-07m2-3 妥当かつ強固なデコーディング制度の統計検定法、iPIPIの提案 Satoshi Hirose（廣瀬智士）^1,2 ¹情報通信研脳情報通信融合研究センター ²大阪大院生命機能 In fMRI decoding studies using pattern classification, ""second-level"" group-level statistical test is performed after decoding analyses for individual participants. To test whether brain activation includes information about label, neuroscientists has often compared mean decoding accuracy across participants with chance-level by means of t-test. However, because decoding accuracy can never below chance-level, significant result of t-test can only be interpreted as ""at least one person in population has information about label in brain."" Here, we propose a novel second-level statistical test procedure, named Permutation based Information Prevalence Inference using i-th order statistic (iPIPI). In iPIPI, i-th order statistic of samples (i-th lowest decoding accuracy) is compared with null distribution, which is estimated by permutation test. Because iPIPI tests whether proportion of population with better than chance decoding accuracy (prevalence) is higher than threshold, significant result of iPIPI can interpret as ""majority of the population has information about label in brain."" We provide theoretical detail and procedures for applying iPIPI to empirical procedures. Also, by applying iPIPI to artificial data, we demonstrate that our method is robust against outlier and have high statistical power when compared with an existing method. Also, we demonstrate the use of our method by reporting a real dataset measured in fMRI decoding study. Program code for iPIPI is available in our website.		7月27日（土）10:30～10:45　第7会場（朱鷺メッセ　2F　201B） 3O-07m2-4 自由行動下における皮質脳波から筋活動のデコーディング Tatsuya Umeda（梅田達也）¹，Masashi Koizumi（小泉昌司）¹，Yuko Katakai（片貝祐子）²，Ryoichi Saito（齋藤亮一）³，Kazuhiko Seki（関和彦）¹ ¹国立精神・神経セ神経研モデル動物 ²国立精神・神経セ神経研霊長類管理室 ³予防衛生協会 Remarkable advances have recently been made in the development of Brain-Machine Interface (BMI) technologies for restoring or enhancing motor function. However, the application of these technologies may be limited to patients in static conditions, as these developments have been largely based on studies of animals in constrained movement conditions. The ultimate goal of BMI technology is to enable individuals to move their bodies naturally or control external devices without physical constraints. Here, we demonstrate accurate decoding of muscle activity from electrocorticogram (ECoG) signals in unrestrained, freely behaving marmosets. We recorded ECoG signals from the sensorimotor cortex as well as electromyogram signals from multiple muscles in the upper arm while marmosets performed two types of behaviors with no physical restraints, as follows: forced forelimb movement (lever pulling task) and natural whole-body movement (free movement within the cage). As in previous reports using restrained macaque monkeys, we confirmed that muscle activity during forced forelimb movement was accurately predicted from simultaneously recorded ECoG data. More importantly, we successfully demonstrated that accurate prediction of muscle activity from ECoG data was possible in marmosets performing natural whole-body movement. We found that high-gamma activity in the primary motor cortex primarily contributed to the prediction of muscle activity during natural whole-body movement as well as forced forelimb movement. In contrast, the contribution of high-gamma activity in the premotor and primary somatosensory cortices was significantly larger during natural whole-body movement, which suggests that activity in a larger area of the sensorimotor cortex was needed to predict muscle activity during natural whole-body movement. These results contribute to the future application of BMI systems in unrestrained individuals.