Identify Neural Areas for Bilingual by Interpreting EEG-Based Deep Learning Models

Currently, the research on bilingualism is a challenging task. This topic has captured the attention of many researchers. Inspired by the robust performance of the deep learning model, we propose to reveal the underlying characteristics of brain activity in bilingual reading by interpreting how deep learning works. To support our proposal, we offer a saliency method using mask perturbation to interpret the deep learning model. Further, we design a class-oriented loss to elaborate on the difference between the two languages. In the experiment, we collect a bilingual reading EEG dataset and build multiple representative deep-learning models. Then, the proposed saliency method is applied to identify the neural areas corresponding to the salient channels. The revealed large overlap in neural areas between the two languages is consistent with existing conclusions in cognitive neuroscience. And there are also similarities in the views of different models, suggesting the existence of underlying cognitive patterns.