CLASH: Contrastive learning through alignment shifting to extract stimulus information from EEG

Stimulus-evoked EEG data has a notoriously low signal-to-noise ratio and high inter-subject variability. We propose a novel paradigm for the self-supervised extraction of stimulus-related brain response data: a model is trained to extract similar information between two time-aligned segments of EEG in response to the same stimulus. The extracted information can subsequently be used to obtain better results in downstream tasks that utilize the response to the stimulus. We show the efficacy of our method for a downstream task of decoding the speech envelope from auditory EEG. Our method outperforms other state-of-the-art denoising techniques, improving reconstruction scores by 45\%. Additionally, we show that in contrast to the baseline denoising techniques, our method can be used with data of unseen subjects and stimuli without retraining, improving decoding performance by 19\% and 34\% over raw EEG for two holdout datasets. Finally, the last experiment reveals that the accuracies obtained in the CLASH paradigm are significantly correlated with the percentile of obtained reconstruction correlation on the null distribution. In general, we showed that the proposed paradigm is suitable to train deep learning models to extract stimulus information from EEG while being stimulus feature agnostic.