Title: EvoBrain: Enabling Continual Learning of EEG Foundation Models for Diverse BCI Applications

Abstract

Electroencephalography (EEG) serves as the fundamental technology for non-invasive brain-computer interfaces (BCIs). However, traditional decoding methods are constrained by fragmented, task-specific architectures that hinder scalability across different tasks. Although EEG foundation models, which are pre-trained on extensive datasets, offer the potential for universal brain decoding, their current post-training strategies rely on isolated fine-tuning for individual tasks. This static approach impedes knowledge transfer among heterogeneous tasks, limits model scalability, and introduces computational and storage costs that increase linearly as the number of tasks grows.

To resolve these limitations, we reframe downstream adaptation as a cross-task continual learning challenge and introduce EvoBrain, a dynamic, task-aware framework designed for unified EEG decoding. EvoBrain manages the trade-off between plasticity and stability through two synergistic mechanisms: (1) Neuro-Spectral Task Normalization (NSN), which aligns new tasks with historical statistics and recalibrates spectral responses to address distributional and neuro-spectral shifts; and (2) Response-Affinity Distillation (RAD), which, when paired with time-dependent replay, maintains the response geometry of previous tasks and facilitates selective knowledge transfer between spectrally compatible tasks, thereby significantly reducing catastrophic forgetting.

Comprehensive evaluations across six distinct BCI tasks reveal that EvoBrain consistently outperforms existing state-of-the-art methods, regardless of the foundation backbone used, while achieving an optimal balance between plasticity and stability. As far as we know, this study marks the first implementation of cross-task continual learning in the EEG field, bringing us closer to the goal of a unified, "one-for-all" brain decoding system.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC