Sequential Bayesian Learning of Biosignal Patterns

Biosignals have non-stationary properties that change over time, so conventional static analysis methods have limitations. In this research, we are developing learning algorithms that can adapt to changing biosignal patterns by using a sequential Bayesian learning framework to recursively update the posterior distribution of classification model parameters. In particular, by constructing Bayesian classification models based on stochastic generative models of EMG signals and EEG, and connecting them to sequential Bayesian updates, we realize more reliable sequential learning that appropriately considers the uncertainty of biosignals. We have demonstrated the effectiveness through applications such as within-subject continual learning of EMG signals, transfer learning of inter-subject variance information, and adaptive classification via semi-supervised learning. This line of work is complementary to our deep learning-based personal adaptation research (in the Machine Learning / Deep Learning Recognition category): here, adaptation is addressed from the standpoint of probabilistic models and Bayesian inference.