Spatiotemporal Convolutions on EEG signal -- A Representation Learning Perspective on Efficient and Explainable EEG Classification with Convolutional Neural Nets

Problem
This preprint addresses the limitations of existing EEG classification methodologies that predominantly utilize independent one-dimensional (1D) convolutional layers for spatial and temporal feature extraction. The authors highlight a gap in understanding the impact of using bi-dimensional (2D) spatiotemporal convolutions on learning dynamics and representational efficiency in EEG signal processing. The study aims to elucidate whether 2D convolutions can enhance training efficiency without compromising classification performance.

Method
The authors propose a novel architecture that incorporates 2D spatiotemporal convolutions, contrasting it with traditional 1D CNNs and a CNN-transformer hybrid model. The experiments are conducted on two datasets: a low-dimensional (3-channel) and a high-dimensional (22-channel) BCI motor imagery classification task. The training process and compute requirements are not explicitly detailed, but the authors emphasize the reduced training time associated with the 2D convolutional approach. The study employs representational similarity analysis to investigate the internal representations generated by the different models, focusing on the geometrical differences in feature space.

Results
The findings indicate that the 2D convolutional architecture significantly reduces training time in high-dimensional tasks while maintaining comparable classification performance to 1D models. Although specific performance metrics are not disclosed, the authors assert that the 2D model achieves similar accuracy levels as the 1D counterparts, suggesting an effective trade-off between speed and performance. The representational similarity analysis reveals that 1D and 2D models produce markedly different internal representations, which may have implications for interpretability and feature extraction in EEG classification tasks.

Limitations
The authors acknowledge that their study is limited by the scope of the datasets used, which may not fully represent the diversity of EEG signals encountered in real-world applications. Additionally, while the study demonstrates the efficiency of 2D convolutions, it does not explore the potential for overfitting or generalization issues that may arise in more complex datasets. The lack of detailed training compute specifications also limits reproducibility. Furthermore, the implications of the representational differences on downstream tasks remain unexplored.

Why it matters
This work has significant implications for the design of EEG classification systems, particularly in applications requiring real-time processing and interpretability. By demonstrating that 2D convolutions can enhance training efficiency while preserving performance, the authors provide a compelling argument for re-evaluating architectural choices in EEG signal processing. The insights into representational geometry may inform future research on model interpretability and the development of more efficient algorithms for complex multivariate signal analysis. This study encourages further exploration of architectural innovations in deep learning for biomedical applications.

Authors: Laurits Dixen, Stefan Heinrich, Paolo Burelli
Source: arXiv:2605.03874
URL: https://arxiv.org/abs/2605.03874v1