RECTOR: Masked Region-Channel-Temporal Modeling for Cognitive Representation Learning

ICLR 2026 Conference Submission24659 Authors

Keywords: self-supervised learning, self-attention mechanism, masked modeling, contrastive learning, EEG, sEEG, affective and cognitive disorders

TL;DR: RECTOR is a novel self-supervised EEG/sEEG framework that models region, channel, and temporal dynamics via a hierarchical self-attention mechanism for cognitive representation learning.

Abstract:

Affective and cognitive disorders are characterized by complex, distributed brain network dynamics across distinct functional regions, channels, and time, posing a significant challenge to learning robust representations for clinical diagnosis. We introduce $\textbf{RECTOR}$ (Masked $\textbf{Re}$gion–$\textbf{C}$hannel–$\textbf{T}$emp$\textbf{or}$al Modeling), the first end-to-end, self-supervised brain region modeling framework that unifies region, channel, and temporal representation learning in a single architecture. At its core is $\textbf{RECTOR-SA}$, a novel hierarchical self-attention mechanism. It efficiently models region-channel-temporal interactions in a block-wise paradigm, incorporating both anatomical priors and dynamic functional gating. It further integrates $\textbf{RECTOR-Mask}$, a novel masking strategy that generates diverse region-channel-temporal
views to establish a challenging pretext task. The self-supervision is driven by $\textbf{NC}^\textbf{2}$$\textbf{-MM}$, our learning objective. It synergistically encourages the encoder to learn both predictive and consistent representations across different views. Finally, we introduce $\textbf{RCReg}$, a tailored regularization on region–channel tokens. It prevents trivial region features and enables the model to explicitly learn and disentangle both region-common and channel-specific representations. Across diverse benchmarks, RECTOR sets a new state-of-the-art in EEG emotion recognition and sEEG task-engagement classification. In addition, it achieves superior computational efficiency in spatio-temporal self-attention, demonstrates strong potential for large-scale pre-training, and provides interpretable, multi-view insights into neural representations at both brain region and channel levels.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 24659