Learning Invariant Representations From EEG via Adversarial Inference

Abstract

Discovering and exploiting shared, invariant neural activity in electroencephalogram (EEG) based classification tasks is of significant interest for generalizability of decoding models across subjects or EEG recording sessions. While deep neural networks are recently emerging as generic EEG feature extractors, this transfer learning aspect usually relies on the prior assumption that deep networks naturally behave as subject-(or session-) invariant EEG feature extractors. We propose a further step towards invariance of EEG deep learning frameworks in a systemic way during model training. We introduce an adversarial inference approach to learn representations that are invariant to inter-subject variabilities within a discriminative setting. We perform experimental studies using a publicly available motor imagery EEG dataset, and state-ofthe-art convolutional neural network based EEG decoding models within the proposed adversarial learning framework. We present our results in cross-subject model transfer scenarios, demonstrate neurophysiological interpretations of the learned networks, and discuss potential insights offered by adversarial inference to the growing field of deep learning for EEG. INDEX TERMS Adversarial learning, brain-computer interface, deep neural networks, electroencephalogram, invariant representation, motor imagery.

Meta Learn on Constrained Transfer Learning for Low Resource Cross Subject EEG Classification

“…Interpretability is an important aspect of machine learning for clinical applications. A number of previous works on EEG analysis such as [5] [14] [38] [39] incorporates this in their model development and provides insights towards interpretability in the model, revealing the relevance of each input unit (either channels of the EEG device or pixels of the spatial-temporal EEG data matrix) with the classification decision. Similar interpretation analysis is performed in our work adopting the occlusion based approach [37], which reveals the different levels of reliance on the individual channels before and after the transfer process.…”

A Survey on Deep Learning-Based Short/Zero-Calibration Approaches for EEG-Based Brain–Computer Interfaces

“…The question here is why TL can be considered as an effort to reduce cost/time-consuming calibration. Most studies assumed that the subject-invariant feature space can be directly applied with zero or short-calibrations for new subjects' EEG data (Jeon et al, 2020;Özdenizci et al, 2020). Contrary to explicit TL-based methods, implicit TL-based approaches follow the hypothesis that their method can train domain-invariant feature spaces on the basis of only their internal architectures without explicitly minimizing the discrepancy.…”

“…In recent years, efforts have been made to take advantage of other real EEG samples (i.e., from a session or a subject) to train deep neural networks that decode EEG samples, thereby mitigating the data insufficiency problem (Chai et al, 2016 ; Andreotti et al, 2018 ; Fahimi et al, 2019 ; Özdenizci et al, 2020 ). These studies known as TL have focused on transferring knowledge from one dataset to another one.…”

Low-Dimensional Subject Representation-Based Transfer Learning in EEG Decoding

“…This section reviews the related works on transfer learning techniques for EEG-based BCI. Although some research aimed Transfer Learning for BCI Targeted Subject [10,19,20,21,22] Calibration Data [12,23,24,25,26,27] [ 28,11,29,30,31,32] [ 5,9], Proposed at intra-subject-cross session transfer [34], this study will mainly focus on the review of inter-subject transfer research. Fig.…”

“…Tu and Sun proposed a method that can extract both robust CSFs for all the subjects and adaptive CSFs for a single subject [10]. Recently,Özdenizci et al aimed at discovering the subjectindependent features across subjects by using convolutional neural networks and adversarial training [21]. Zhang et al and Jeon et al were developing methods for learning the invariant data representations with deep learning approach [29,31].…”