Mitigating Subject Dependency in EEG Decoding with Subject-Specific Low-Rank Adapters

TL;DR

The paper tackles the challenge of cross-subject variability in EEG decoding, which hampers scaling of foundation models. It introduces the Subject-Conditioned Layer, a drop-in module that decomposes layer weights into a shared $W_{ ext{general}}$ and subject-specific low-rank adapters with $W_s^T = A_s B_s$ of rank $r$, producing \bar{X} = \sigma\left(X W_{ ext{general}}^\top + \sum_s (M_s X) W_s^\top\right). Through experiments on BCIC_IV2a/2b with EEGNeX (CNN) and Patched Brain Transformer (ViT), the method outperforms a subject-agnostic baseline, the average of subject-specific models, and subject-specific LoRA baselines, with qualitative embedding analyses supporting effective disentanglement of shared and personalized features. The work proposes a scalable path toward cross-subject EEG foundation models and discusses future directions such as pre-training plus adapter fine-tuning and zero-shot adaptation via adapter-space geometry to handle unseen subjects.

Abstract

Subject-specific distribution shifts represent an important obstacle to the development of foundation models for EEG decoding. To address this, we propose Subject-Conditioned Layer,, an adaptive layer designed as a drop-in replacement for standard linear or convolutional layers in any neural network architecture. Our layer captures subject-specific variability by decomposing its weights into a shared, subject-invariant component and a lightweight, low-rank correction unique to each subject. This explicit separation of general knowledge from personalized adaptation allows existing models to become robust to subject shifts. Empirically, models equipped with our layer outperform both a shared-weight-only model (subject-agnostic model) and the average of individually trained subject-specific models. Consequently, the Subject-Conditioned Layer, offers a practical and scalable path towards building effective cross-subject foundation models for EEG.

Figures (2)

• Figure 1: Latent Representations Colored by Subject ID vs. Target Labels. The top row compares t-SNE projections of representations derived from (Left) the shared weights ($W_{\text{general}}$), (Middle) the subject-specific weights ($W_{s}$), and (Right) the fused representations from the complete Subject-Conditioned Layer. In this row, all points are colored by subject ID. The bottom row presents the same data points, but colored according to their ground-truth target labels. This comparison clearly shows that the sub-clusters formed by our full model (top right) correspond directly to the target classes (bottom right), demonstrating that the model learns discriminative, task-relevant features within each subject’s representation space. All representations are obtained from the classification token predicted by the Patched Brain Transformer, with a linear classification head used for visualization clarity.
• Figure 2: Subject-Conditioned Layer. The input $X$ is processed by a shared general weight matrix $W_{\text{general}}$. For each subject, a subject-specific low-rank adapter, parameterized by matrices $A_s$ and $B_s$ of rank $r$, computes a correction that is added to the output of the general weight matrix, yielding the layer output $\bar{X}$.