Research map / Time series / Foundation models for time series
Brain-Semantoks: Learning Semantic Tokens of Brain Dynamics with a Self-Distilled Foundation Model
Sam Gijsen, Marc-Andre Schulz, Kerstin Ritter
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| self-distillation with semantic tokenizer instead of reconstruction-based pretraining (linear probe) | BrainLM / Brain-JEPA · ABIDE | balanced accuracy: 53.84 -> 65.13 | Improved |
| same | Brain-JEPA · SRPBS SZ | balanced accuracy: 57.63 -> 69.26 | Improved |
| same | BrainLM · HBN Age | balanced accuracy: 43.89 -> 39.16 | Hurt |
| add the global CLS self-distillation loss (ablation) | JEPA-style baseline · 10-task average | balanced accuracy: 47.32 -> 52.39 | Improved |
| add Teacher-guided Temporal Regularizer for first 5% of training (ablation) | Brain-Semantoks · 10-task average | balanced accuracy: 50.88 -> 52.39 | Improved |
| keep TTR for the entire pretraining instead of 5% (ablation) | Brain-Semantoks · 10-task average | balanced accuracy: 52.39 -> 49.60 | Hurt |
| use slice masking instead of random masking (ablation) | Brain-Semantoks · 10-task average | balanced accuracy: 51.03 -> 52.39 | Improved |
| use block masking instead of slice masking (ablation) | Brain-Semantoks · 10-task average | balanced accuracy: 52.39 -> 49.71 | Hurt |
| use structured dual-branch tokenizer kernel instead of single dense kernel 20 (ablation) | Brain-Semantoks · 10-task average | balanced accuracy: 48.28 -> 52.39 | Improved |
| set token loss weight lambda_Tok to 0.5 instead of 1.0 (ablation) | Brain-Semantoks · 10-task average | balanced accuracy: 51.62 -> 52.39 | Improved |
| semantic tokenizer instead of per-ROI linear projection (ablation) | Brain-Semantoks · scaling (random init) | linear probe accuracy: up to 12% gain | Improved |
| more pretraining data | Brain-Semantoks · in- and out-of-distribution tasks | linear probe accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Brain-JEPA — Brain-JEPA: Brain Dynamics Foundation Model with Gradient Positioning and Spatiotemporal Masking (NeurIPS 2024) · better on 6 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Supervised Learning of Brain Dynamics from Broad Neuroimaging Data (NeurIPS 2022 · similarity 0.89)
- BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.89)
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.88)
- Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.87)
- BrainVista: Modeling Naturalistic Brain Dynamics as Multimodal Next-Token Prediction (NeurIPS 2026 · similarity 0.87)
- BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
