A01Nonlinear dynamical representation learning and its application to analysis and prediction of the brain
Analyzing dynamics of the brain is crucial to understand its computational mechanisms, and can give us invaluable implications for developing futuristic brain information processing systems. Although many analytical/modeling methods have been proposed so far, such as dynamical network modeling or large-scale brain models, we still have not reached its full understanding because of its complex nonlinearity, underlying unobservable hidden factors, lack of label information, and so on.
This study aims to investigate such complicated nonlinear dynamics of the brain by developing a novel nonlinear dynamical representation learning method. The method is supposed to be an extension of our recently proposed nonlinear independent component analysis (ICA) framework to a nonlinear dynamical model, and learns a low-dimensional representation (feature space) of the nonlinear dynamics and its underlying factors from the observed multidimensional time series based on a (unsupervised) deep learning framework. By applying the proposed method to brain imaging data and extracting their dynamical representation, we attempt to figure out the mechanisms of the brain dynamics and its relationship to the hidden factors in an interpretable manner. We also try to predict the brain dynamics by further extending the method to a dynamical predictive model based on the learned feature space.
In addition to the brain-scientific contributions, our possible findings could also contribute to give new insight to the artificial intelligence researches to realize more efficient systems in the future.