Research

研究活動

セミナー詳細

2026年02月03日

DNN-Based Multivariate Pattern Analysis of Neural Mechanisms in Top-Down Cognitive Control

日 時 2026年02月03日(火) 13:10 より 14:00 まで
講演者 Mr. Wang Siyu
講演者所属 Kyoto University
場 所 Seminar Room A/B (1F, Myodaiji)
お問い合わせ先 Keiichi Kitajo (Division of Neural Dynamics, NIPS)
要旨 Top-down cognitive control relies on a feedforward biasing mechanism coordinated across multiple cortical regions. This seminar presents two studies that investigate the functional relevance of neural substrates during goal maintenance and transitions, using visual-search and task-switching paradigms. In the first study on visual search, how feedforward bias modulates the functional relevance of cortical regions under varying endogenous attentional-control demands remains a subject of debate.
To address this, I propose a DNN-based analytic framework incorporating a Multi- Scale Short-Term Temporal Attention Network to decode residual dynamics from post-task rs- 
fMRI. Furthermore, I employ a mask-one-ROI-out analysis to quantify the informational value of specific regions. The results reveal a flexible redistribution of cognitive resources across hemispheres in response to endogenous attentional-control requirements. In the second study on task switching, feedforward bias facilitates the proactive updating of stimulus–response rules. Traditional approaches typically represent cortical signals as tensors, which fail to leverage spatial organization and preclude permutation-invariant modeling. I propose an Anatomical-Connectivity-Guided Functional Connectivity framework that embeds a group-average structural connectome as a pathway prior in recurrent graph neural networks. This approach transforms EEG data into non-Euclidean representations to identify connective features.
The findings suggest that feedforward bias is implemented by qualitatively distinct networks: switch preparation robustly recruits a frontoparietal network, hereas repeat preparation relies more stably on sensorimotor pathways. Together, these studies indicate that whole temporoparietal systems contribute to maintaining internal goals and suppressing both internal and external interference, whereas proactive reconfiguration of internal goals requires stronger frontoparietal engagement.