Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph-Based Learning of Spectro-Topographical EEG Representations with Gradient Alignment for Brain-Computer Interfaces

Prithila Angkan, Amin Jalali, Paul Hungler, Ali Etemad

TL;DR

This work tackles EEG-based BCI representation learning by proposing GEEGA, a graph-based framework that fuses frequency-domain topography maps and time-frequency spectrograms through a graph convolutional network. It addresses cross-domain gradient conflicts via a Pareto-based gradient alignment strategy and enhances class separability with Git-based center and pairwise losses. Evaluated on three public EEG datasets, GEEGA achieves state-of-the-art accuracy and F1 scores, supported by thorough ablation analyses that confirm the contribution of each component. The approach advances robust, multi-domain, spectro-topographical EEG representations with potential for real-time and cross-task BCI applications.

Abstract

We present a novel graph-based learning of EEG representations with gradient alignment (GEEGA) that leverages multi-domain information to learn EEG representations for brain-computer interfaces. Our model leverages graph convolutional networks to fuse embeddings from frequency-based topographical maps and time-frequency spectrograms, capturing inter-domain relationships. GEEGA addresses the challenge of achieving high inter-class separability, which arises from the temporally dynamic and subject-sensitive nature of EEG signals by incorporating the center loss and pairwise difference loss. Additionally, GEEGA incorporates a gradient alignment strategy to resolve conflicts between gradients from different domains and the fused embeddings, ensuring that discrepancies, where gradients point in conflicting directions, are aligned toward a unified optimization direction. We validate the efficacy of our method through extensive experiments on three publicly available EEG datasets: BCI-2a, CL-Drive and CLARE. Comprehensive ablation studies further highlight the impact of various components of our model.

Graph-Based Learning of Spectro-Topographical EEG Representations with Gradient Alignment for Brain-Computer Interfaces

TL;DR

This work tackles EEG-based BCI representation learning by proposing GEEGA, a graph-based framework that fuses frequency-domain topography maps and time-frequency spectrograms through a graph convolutional network. It addresses cross-domain gradient conflicts via a Pareto-based gradient alignment strategy and enhances class separability with Git-based center and pairwise losses. Evaluated on three public EEG datasets, GEEGA achieves state-of-the-art accuracy and F1 scores, supported by thorough ablation analyses that confirm the contribution of each component. The approach advances robust, multi-domain, spectro-topographical EEG representations with potential for real-time and cross-task BCI applications.

Abstract

We present a novel graph-based learning of EEG representations with gradient alignment (GEEGA) that leverages multi-domain information to learn EEG representations for brain-computer interfaces. Our model leverages graph convolutional networks to fuse embeddings from frequency-based topographical maps and time-frequency spectrograms, capturing inter-domain relationships. GEEGA addresses the challenge of achieving high inter-class separability, which arises from the temporally dynamic and subject-sensitive nature of EEG signals by incorporating the center loss and pairwise difference loss. Additionally, GEEGA incorporates a gradient alignment strategy to resolve conflicts between gradients from different domains and the fused embeddings, ensuring that discrepancies, where gradients point in conflicting directions, are aligned toward a unified optimization direction. We validate the efficacy of our method through extensive experiments on three publicly available EEG datasets: BCI-2a, CL-Drive and CLARE. Comprehensive ablation studies further highlight the impact of various components of our model.

Paper Structure

This paper contains 8 sections, 7 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Problem Statement
  5. Our Approach
  6. Experiment setup
  7. Results
  8. Conclusion

Figures (2)

  • Figure 1: (a) The overview of our proposed network is depicted. (b) The concept of the Git loss is presented where we aim to minimize intra-class distances $d_{1}$ and maximize inter-class distances $d_{2}$. (c) The concept of gradient alignment is presented.
  • Figure 2: The first row shows cosine similarities between multi-spectral topography maps and the fused domain, while the second row shows the same for spectrograms. Blue (values < 0) indicates gradient conflicts, while red (values > 0) indicates no conflict.