This paper formalizes the design choices for spatio-temporal masking strategies and train various models with a single pipeline under varying masking configurations across two SSL frameworks (MAE and JEPA), and systematically evaluates the resulting 58 pre-trained models under a linear probe.
Abstract
EEG foundation models hold promise for scalable brain-signal decoding across clinical and cognitive neuroscience applications, yet their pre-training pipelines remain poorly understood. Among design choices, the masking strategy is particularly critical: it determines what the network must predict and from which context. Yet it has never been ablated in isolation, as each new model bundles a new masking strategy with a new backbone and objective. In this paper, we formalize the design choices for spatio-temporal masking strategies and train various models with a single pipeline under varying masking configurations across two SSL frameworks (MAE and JEPA). We then systematically evaluate the resulting 58 pre-trained models on the 12 datasets of OpenEEGBench under a linear probe. Both frameworks agree on an optimal masking configuration and on shared failure modes. Outside these, performance is robust: 11 MAE and 9 JEPA configurations are statistically indistinguishable from the best. We further identify a novel JEPA-specific failure mode, tagged bias-inflation collapse, invisible to standard detectors. With a well-chosen mask, our pipeline reaches REVE-level downstream performance at a fraction of REVE's pre-training compute.
This perspective lays down the basic concepts needed to understand MEG foundation models and provides a didactic overview of the field's key design choices, including tokenization, sensor- versus source-space representations, sensor-geometry encoding, backbone architectures, self-supervised objectives, and pretraining...
Philipp Thölke, Hamza Abdelhedi, Yorguin-José Mantilla-Ramos et al.· 1 citation
This work examines two representative architectures, MAtt and GBWAtt, and finds that their learned attention weights remain close to uniform after training, and introduces SPD-MetaFormer, an attention-free architecture built on uniformly weighted Fr\'echet aggregation under log-Euclidean geometry.
This work proposes a modular framework for the unified evaluation of these paradigms in subject-independent learning, and conducts extensive experiments on Motor Imagery and Event-Related Potential datasets, representative of interactive BCI scenarios spanning both oscillatory (MI) and event-related (ERP) paradigms.
Lia Schmid, Jacopo Burger, Alessandro D'Amelio et al.· Proceedings of the 28th Inte...· 0 citations
Sign differences suggest that a single comparator may not provide an architecture-invariant decomposition of a pretrained-versus-supervised performance gap, and validation-fitted temperature scaling returns foundation-model calibration error to the supervised range despite substantially lower four-class accuracy.
Deep learning has become a cornerstone of EEG-based brain decoding, with a growing number of architectures proposed every day. However, how the performance of these different models scales with data volume is not clear. Although this relationship has been characterized in other fields under the name of scaling laws, it...
José Manuel Nunes de Oliveira, Bruna J. Lopes, Léo Burgund et al.· 0 citations
iBrain is introduced, a unified foundation model that jointly learns from iEEG and spiking activity that consistently outperforms single-signal pretraining baselines and achieves state-of-the-art performance on multiple benchmarks.
Ying Chen, Ti-Ou Wang, Zhi-Feng Yue· 0 citations
Related blog posts
Microsoft Research Blog· microsoft.comNov 16, 2026
AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduNov 13, 2026