UniNeuro: A Unified EEG Foundation Model for Heterogeneous EEG Task Learning

Abstract

Electroencephalography (EEG), as a non-invasive physiological signal, plays an increasingly important role in human-centered multimedia systems, supporting applications ranging from emotion recognition and cognitive load assessment to motor-imagery-based control. However, due to the substantial differences in temporal scales, neural patterns, and label semantics across EEG tasks, it remains challenging to unify heterogeneous downstream tasks within a single model, while reconciling shared and task-specific EEG features. To this end, we propose UniNeuro, a unified EEG foundation framework that seamlessly integrates heterogeneous-task pretraining and downstream adaptation into a single discriminative learning pipeline. Within the backbone, we introduce a task-conditional Mixture-of-Experts (MoE) with gradient reversal, enforcing shared experts to capture cross-task features while dynamically routed experts focus on task-specific features, naturally decoupling the two types of representations. At the downstream adaptation stage, we propose Hierarchical Prototype-Guided Contrastive Learning (HPGCL) to explicitly structure the latent space at both task and label levels, enhancing task-level clustering and within-task class separability. Pretrained on over 10,000 hours of EEG data and evaluated on six diverse tasks, UniNeuro shows strong unified multi-task performance and narrows the gap with single-task fine-tuning, offering a scalable "brain-modality" encoder for next-generation interactive multimedia systems.