C01 Hiroshi Morioka

C01 Revealing high-level cyclic causal representation of predictions and actions in the brain　

This project aims to understand the high-level representation behind the prediction and actions by the brain, based on novel machine learning theory. Although some unified theories for explaining the relations between environment, cognition, and actions were proposed, such as free energy principle, they are mostly abstract and thus have difficulty to associate with the actual high-dimensional brain activities we can observe. Therefore, this project proposes a novel unsupervised data-driven framework for estimating hidden high-level representation of the brain, for connecting those theorems to the actual phenomena happening in the brain. The framework assumes cyclic causal relations between modalities (environment, brain, and action), and then estimates the latent high-level representation behind them.  We verify the proposal by applying it to high-dimensional multimodal data of the brain (fMRI, EEG, calcium imaging, spike, etc.), environment (visual, auditory, tactile stimulus, etc.), and actions (video, etc.), and then associating the obtained representation with the unified theories. Such unprecedented high-level representation of the brain is expected to give new ideas to the unified theories, which have not been known yet.