Decoding learning strategies from EEG signals provides generalizable features for decoding decision

Abstract

Recent studies have demonstrated that learning strategies can be decoded from EEG data using a computational model of model-based and model-free reinforcement learning. The results raise expectations for improving the decodability of decisions in a broader context because the decision is an inherent part of the learning strategies. In this study, we investigated this possibility using various information theory-based methods. First, we trained a simple deep neural network to decode learning strategies from EEG signals collected while human subjects perform a strategy learning task with context changes. We then evaluated the ability of the model to decode subjective decision signals from EEG signals in another decision-making scenario that was not used during training. This zero-training scheme allows us to investigate whether the learning strategy decoder gleans information generalizable to various decision-making scenarios. Notably, we found that the decoder contains a significant amount of mutual information between input, hidden, and output for the new data (decision-making task; p < 5e-2), as well as the original training data (strategy learning task; p < 1e-5). In subsequent analyses of the neural representations of the model's hidden layers, we found informative features for decoding decisions in its deep layers. The results suggest that decoding learning strategies will help design generalizable EEG decoders.