TL;DR
Neuro-MoBRE is a novel decoding framework that explicitly manages data heterogeneity across subjects and tasks in intracranial recordings, improving generalization and zero-shot decoding capabilities.
Contribution
It introduces a brain-regional-temporal embedding and mixture-of-experts approach, along with pre-training and task-disentangled methods, to enhance neurophysiological decoding across diverse subjects and tasks.
Findings
Outperforms prior methods on intracranial data
Achieves robust zero-shot decoding on unseen subjects
Effective across multiple complex tasks
Abstract
Neurophysiological decoding, fundamental to advancing brain-computer interface (BCI) technologies, has significantly benefited from recent advances in deep learning. However, existing decoding approaches largely remain constrained to single-task scenarios and individual subjects, limiting their broader applicability and generalizability. Efforts towards creating large-scale neurophysiological foundation models have shown promise, but continue to struggle with significant challenges due to pervasive data heterogeneity across subjects and decoding tasks. Simply increasing model parameters and dataset size without explicitly addressing this heterogeneity fails to replicate the scaling successes seen in natural language processing. Here, we introduce the Neural Mixture of Brain Regional Experts (Neuro-MoBRE), a general-purpose decoding framework explicitly designed to manage the ubiquitous…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The paper presents a framework to address cross-subject and cross-task heterogeneity. By using a brain-regional mixture-of-experts mechanism, a brain-regional-temporal tokenizer, and task-disentangled aggregation, the framework separates regional, temporal, and functional variability of EEG data. 2. The authors curate and unify one of the most comprehensive intracranial EEG datasets to date, spanning 11 subjects and five heterogeneous tasks, including speech decoding, movement execution, and
1. While the paper emphasizes that Neuro-MoBRE is designed to “explicitly resolve multi-subject and multi-task heterogeneity,” the empirical evidence for this claim is qualitative and indirect. The results show performance gains across subjects and tasks, but it remains unclear how much of that improvement is attributable to reduced heterogeneity versus general over-parameterization or better representation learning. 2. The paper claims robustness to low-SNR neural recordings, but the evidence
1. Well-Motivated and Novel Methodology: The paper accurately identifies "data heterogeneity" as a fundamental challenge in neural decoding and proposes a systematic solution. The core ideas, particularly the brain-regional MoE and task-disentangled aggregation, are innovative. 2. Multi-Subject, Multi-Task Modeling: The framework successfully unifies data from multiple subjects and tasks within a single model. Its demonstrated zero-shot generalization to unseen subjects have some advantages wit
1. Limited Generalization Evidence: The model is evaluated solely on one private iEEG dataset. Its generalization capability remains unverified on any public iEEG benchmarks with varying experimental paradigms and recording parameters [1,2], raising questions about its robustness across broader data distributions. 2. Brain Region Modeling: The current approach models neural activity at the level of entire brain regions, potentially overlooking the functional complexity and finer-grained functio
1. The proposed framework achieved the best within-subject performance across all decoding tasks compared with the baselines. 2. The proposed Neuro-MoBRE framework is conceptually interesting, especially its modular mixture-of-experts (MoE) design that allocates region-specific experts for decoding across multiple brain areas.
Major concerns: 1. In the introduction, the authors identify low SNR in non-invasive neurophysiological recordings as one of the main challenges motivating this work (line 70). However, the paper states: ***“To circumvent the limitations posed by low SNRs in non-invasive neurophysiological recordings, we rigorously evaluate Neuro-MoBRE using intracranial data collected from 11 subjects across five distinct decoding tasks.”*** If the primary goal is to enhance robustness to low-SNR conditions, ev
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
