Mutual Information-Driven Subject-Invariant and Class-Relevant Deep Representation Learning in BCI

Eunjin Jeon, Wonjun Ko, Jee Seok Yoon, Heung-Il Suk

Introduction

A major hurdle in Brain-Computer Interface (BCI) technology is the high degree of inter-subject variability. Electroencephalogram (EEG) patterns for specific tasks, such as motor imagery, vary significantly between individuals. This "domain shift" typically necessitates lengthy, user-specific calibration sessions. While adversarial learning has been explored to mitigate these differences, it often leads to "negative transfer," where task-relevant information is inadvertently discarded in the pursuit of subject-invariance. This paper introduces a novel deep learning framework that leverages information theory to disentangle task-related signals from individual-specific noise, enabling robust performance in "zero-training" scenarios.

Methodology: MI-Driven Representation Learning

The proposed framework moves beyond adversarial competition, instead utilizing Mutual Information (MI) estimation to refine feature representations. The architecture operates in two primary stages:

  1. Class-relevant Feature Decomposition: The network splits intermediate features into class-relevant (intent) and class-irrelevant (subject-specific) paths. By employing a Neural Mutual Information Estimator ($M$) and a Gradient Reversal Layer (GRL), the model minimizes the MI between these paths, effectively isolating the useful signal.
  2. Enriched Feature Representation: To ensure the robustness of class-relevant features, the framework incorporates principles from Deep InfoMax (DIM). Two additional estimators ($T_l$ and $T_g$) maximize the MI between local feature map patches and the global representation, encouraging the learning of globally coherent patterns across the motor cortex.
The proposed architecture featuring class-relevant feature decomposition and enriched feature representation learning.

This "plug-and-play" module can be integrated into existing BCI backbones like EEGNet or Deep ConvNet.

Experimental Setup

The framework was evaluated using two large-scale public motor imagery datasets:

  • GIST Dataset: 50 subjects, left- vs. right-hand imagery.
  • KU Dataset: 54 subjects across two sessions. These datasets provide a rigorous testing ground for cross-subject generalization.

Results and Neurophysiological Validation

The MI-driven approach consistently outperformed baseline pooled training and state-of-the-art adversarial methods. On the GIST dataset with an EEGNet backbone, accuracy reached 76.60%, compared to a 55.38% baseline. Crucially, in zero-training scenarios for unseen subjects, the model maintained an accuracy of 70-73%.

Beyond quantitative metrics, the authors used Layer-wise Relevance Propagation (LRP) to validate the model's focus. The resulting heatmaps demonstrate that the model correctly identifies activity in the contra-lateral motor cortex, aligning with established neurophysiological principles.

Relevance heatmaps for Subject 36, comparing the proposed method with DANN-MI and PSD baselines across different scenarios.

Discussion and Future Directions

By focusing on information disentanglement rather than simple suppression, this framework offers a scalable solution for subject-invariant BCI. Unlike adversarial methods, its complexity does not scale with the number of subjects. Future work will explore the framework's performance on more complex, multi-class tasks and its robustness against real-world artifacts like ocular noise or electrode displacement.

Summary updated

Abstract

In recent years, deep learning-based feature representation methods have shown a promising impact on electroencephalography (EEG)-based brain–computer interface (BCI). Nonetheless, owing to high intra- and inter-subject variabilities, many studies on decoding EEG were designed in a subject-specific manner by using calibration samples, with no concern of its practical use, hampered by time-consuming steps and a large data requirement. To this end, recent studies adopted a transfer learning strategy, especially domain adaptation techniques. Among those, we have witnessed the potential of adversarial learning-based transfer learning in BCIs. In the meantime, it is known that adversarial learning-based domain adaptation methods are prone to negative transfer that disrupts learning generalized feature representations, applicable to diverse domains, for example, subjects or sessions in BCIs. In this article, we propose a novel framework that learns class-relevant and subject-invariant feature representations in an information-theoretic manner, without using adversarial learning. To be specific, we devise two operational components in a deep network that explicitly estimate mutual information between feature representations: 1) to decompose features in an intermediate layer into class-relevant and class-irrelevant ones and 2) to enrich class-discriminative feature representation. On two large EEG datasets, we validated the effectiveness of our proposed framework by comparing with several comparative methods in performance. Furthermore, we conducted rigorous analyses by performing an ablation study in regard to the components in our network, explaining our model’s decision on input EEG signals via layer-wise relevance propagation, and visualizing the distribution of learned features via t-SNE.
BibTeX
@article{jeon2023mutual,
  author = {Eunjin Jeon and Wonjun Ko and Jee Seok Yoon and Heung-Il Suk},
  title = {{Mutual Information-Driven Subject-Invariant and Class-Relevant Deep Representation Learning in BCI}},
  journal = {IEEE TNNLS},
  year = {2023},
  doi = {10.1109/tnnls.2021.3100583},
}
Figure