Improving subject transfer in EEG classification with divergence estimation.

Abstract

\textit{Objective}. Classification models for electroencephalogram (EEG) data show a large decrease in performance when evaluated on unseen test subjects. We improve performance using new regularization techniques during model training. \textit{Approach}. We propose several graphical models to describe an EEG classification task. From each model, we identify statistical relationships that should hold true in an idealized training scenario (with infinite data and a globally-optimal model) but that may not hold in practice. We design regularization penalties to enforce these relationships in two stages. First, we identify suitable proxy quantities (divergences such as Mutual Information and Wasserstein-1) that can be used to measure statistical independence and dependence relationships. Second, we provide algorithms to efficiently estimate these quantities during training using secondary neural network models. \textit{Main Results}. We conduct extensive computational experiments using a large benchmark EEG dataset, comparing our proposed techniques with a baseline method that uses an adversarial classifier. We first show the performance of each method across a wide range of hyperparameters, demonstrating that each method can be easily tuned to yield significant benefits over an unregularized model. We show that, using ideal hyperparameters for all methods, our first technique gives significantly better performance than the baseline regularization technique. We also show that, across hyperparameters, our second technique gives significantly more stable performance than the baseline. The proposed methods require only a small computational cost at training time that is equivalent to the cost of the baseline. \textit{Significance}. The high variability in signal distribution between subjects means that typical approaches to EEG signal modeling often require time-intensive calibration for each user, and even re-calibration before every use. By improving the performance of population models in the most stringent case of zero-shot subject transfer, we may help reduce or eliminate the need for model calibration.