SpecMoE: Spectral Mixture-of-Experts Foundation Model for Cross-Species EEG Decoding
arXiv:2603.16739v2 Announce Type: replace-cross Abstract: Decoding the orchestration of neural activity in electroencephalography (EEG) signals is a central challenge in bridging neuroscience with artificial intelligence. Foundation models have made strides in generalized EEG decoding, yet many existing frameworks primarily relying on separate temporal and spectral masking of raw signals during self-supervised pretraining. Such strategies often tend to bias learning toward high-frequency oscillations, as low-frequency rhythmic patterns can be easily inferred from the unmasked signal. We introd — Davy Darankoum, Chlo\'e Habermacher, Julien Volle, Sergei Grudinin
View PDF HTML (experimental)
Abstract:Decoding the orchestration of neural activity in electroencephalography (EEG) signals is a central challenge in bridging neuroscience with artificial intelligence. Foundation models have made strides in generalized EEG decoding, yet many existing frameworks primarily relying on separate temporal and spectral masking of raw signals during self-supervised pretraining. Such strategies often tend to bias learning toward high-frequency oscillations, as low-frequency rhythmic patterns can be easily inferred from the unmasked signal. We introduce a foundation model that utilizes a novel Gaussian-smoothed masking scheme applied to short-time Fourier transform (STFT) maps. By jointly applying time, frequency, and time-frequency Gaussian masks, we make the reconstruction task much more challenging, forcing the model to learn intricate neural patterns across both high- and low-frequency domains. To effectively recover signals under this aggressive masking strategy, we design SpecHi-Net, a U-shaped hierarchical architecture with multiple encoding and decoding stages. To accelerate large-scale pretraining, we partition the data into three subsets, each used to train an independent expert model. We then combine these models through SpecMoE, a mixture of experts framework guided by a learned spectral gating mechanism. SpecMoE achieves state-of-the-art performance across a diverse set of EEG decoding tasks, including sleep staging, emotion recognition, motor imagery classification, abnormal signal detection, and drug effect prediction. Importantly, the model demonstrates strong cross-species and cross-subject generalization, maintaining high accuracy on both human and murine EEG datasets.
Comments: 34 pages (12 pages in the main text and 22 pages in Supplementary Information)
Subjects:
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)
Cite as: arXiv:2603.16739 [cs.LG]
(or arXiv:2603.16739v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2603.16739
arXiv-issued DOI via DataCite
Submission history
From: Sergei Grudinin [view email] [v1] Tue, 17 Mar 2026 16:20:14 UTC (8,384 KB) [v2] Mon, 30 Mar 2026 17:09:43 UTC (8,378 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
researchpaperarxiv
New memristor design uses built-in oxygen gradient to bring stability to reinforcement learning
In a recent study published in Nature Communications, researchers created a memristor that uses a built-in oxygen gradient to produce slow, stable conductance changes, enabling a reinforcement learning (RL) algorithm to learn faster and more stably than conventional approaches.
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Research Papers
New memristor design uses built-in oxygen gradient to bring stability to reinforcement learning
In a recent study published in Nature Communications, researchers created a memristor that uses a built-in oxygen gradient to produce slow, stable conductance changes, enabling a reinforcement learning (RL) algorithm to learn faster and more stably than conventional approaches.
Living brain cells enable machine learning computations
A research team at Tohoku University and Future University Hakodate has demonstrated that living biological neurons can be trained to perform a supervised temporal pattern learning task previously carried out by artificial systems. By integrating cultured neuronal networks into a machine learning framework, the team showed that these biological systems can generate complex time-series signals, marking a significant step forward in both neuroscience and bio-inspired computing.
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!