Ablation and Robustness Analysis of Domain-Adversarial EEGNet for Calibration-Free Motor Imagery Brain–Computer Interfaces
Year: 2026
Type: conference-paper
Abstract: Calibration-free motor imagery brain-computer interfaces remain difficult to deploy in practice due to strong inter-subject variability in electroencephalographic signals, which often leads to unstable cross-subject performance and requires subject-specific calibration. Domain-adversarial learning has been proposed as a promising approach to address this issue by encouraging subject-invariant feature representations. However, existing studies primarily report aggregate performance metrics and provide limited insight into how adversarial training affects the structure of learned representations. In this work, we present a representation-level ablation and robustness analysis of a domain-adversarial EEGNet model for calibration-free motor imagery classification. The study is conducted on the BCI Competition IV dataset 2a using a strict leave-one-subject-out evaluation protocol. Instead of focusing solely on classification performance, we systematically analyze how adversarial alignment influences subject-wise feature organization, class separability, robustness to alignment strength, and class-level error patterns. The results show that moderate adversarial alignment improves robustness and reduces subject-specific structure in the feature space for a substantial subset of subjects, while excessive alignment can degrade class-discriminative information, particularly for difficult cases. These findings demonstrate that domainadversarial learning is not universally beneficial under strict zerocalibration conditions. The study contributes a clearer understanding of the trade-off between domain invariance and class discrimination and highlights the importance of representationlevel analysis in the design of reliable calibration-free braincomputer interfaces. Unlike prior work that focuses primarily on aggregate performance metrics, this study provides a systematic representationlevel analysis of domain-adversarial learning under strict calibration-free conditions.
Authors: Aigerim Aitim, Zhuldyz Bagybek
Institution: International Information Technologies University
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: hybrid
OpenAlex ID: W7168248156