| name | kast-brain-autoregressive |
| description | KAST-BAR methodology: Knowledge-Anchored Semantically-Dynamic Topology Brain Autoregressive Modeling for universal neural interpretation. Integrates Dual-Stream Hierarchical Attention (DSHA) encoder for brain topology, Knowledge-Anchored Semantic Profiler (KASP) for expert-level text profiles, and Semantic Text-Aware Refiner (STAR) with Latent Expert Queries. Pre-trained on 21 datasets, evaluated on 6 downstream tasks. Use when: building EEG foundation models, brain topology representation learning, semantic-physiological alignment, cross-task neural decoding, medical knowledge integration in BCI.
|
| category | neuroscience |
| tags | ["eeg-foundation-model","brain-topology","semantic-alignment","autoregressive","knowledge-anchored","neural-decoding","multi-modal"] |
| related_skills | ["eeg-foundation-model-adapters","eeg-foundation-lrp-interpretability","laya-eeg-foundation","reve-eeg-foundation","tta-eeg-foundation-models"] |
| activation_keywords | ["kast-bar","knowledge-anchored brain autoregressive","eeg foundation model semantic alignment","dual-stream hierarchical attention eeg","brain topology representation learning","expert-level semantic eeg profiling","universal neural interpretation","semantic text-aware eeg refiner"] |
KAST-BAR: Knowledge-Anchored Brain Autoregressive Modeling
Paper: KAST-BAR: Knowledge-Anchored Semantically-Dynamic Topology Brain Autoregressive Modeling for Universal Neural Interpretation
Authors: Haoning Wang, Wenchao Yang, Shuai Shen, Yang Li
arXiv: 2605.13133 (May 13, 2026)
Category: cs.LG, eess.SP
Code: https://github.com/KAST-BAR/KAST-BAR
Overview
EEG foundation models face two bottlenecks: (1) inadequate modeling of complex spatiotemporal brain topology, and (2) the modality gap between low-level physiological signals and high-level textual semantics. KAST-BAR addresses both by dynamically aligning multi-level brain topology representations with an expert-level semantic space through a three-stage pipeline.
Core Architecture
Stage 1: DSHA Encoder (Dual-Stream Hierarchical Attention)
Captures the brain's intrinsic non-Euclidean topology:
EEG signals (C channels, T timepoints)
-> Local Temporal Stream: models fine-grained temporal dynamics per channel
-> Global Spatial Stream: captures inter-channel topological relationships
-> Hierarchical fusion: integrates local temporal + global spatial contexts
- Models both local temporal dynamics and global spatial contexts simultaneously
- Respects the brain's non-Euclidean topological structure (unlike standard CNNs/RNNs)
- Produces physiologically-grounded representations
Stage 2: KASP (Knowledge-Anchored Semantic Profiler)
Synthesizes expert-level textual profiles from EEG representations:
DSHA representation
-> Physical grounding: links neural patterns to neurophysiological principles
-> Instance-level profiling: generates patient/session-specific text descriptions
-> Expert knowledge anchoring: incorporates medical domain expertise into semantic space
- Bridges the gap between low-level signals and high-level medical semantics
- Produces verifiable, interpretable text profiles
- Anchors semantic representations in established medical knowledge
Stage 3: STAR (Semantic Text-Aware Refiner)
Dynamically reconstructs EEG representations using semantic feedback:
KASP text profile + DSHA representation
-> Latent Expert Queries: learned query vectors that attend to semantic knowledge
-> Cross-modal attention: text profile refines neural representation
-> Refined representation: semantically-enhanced EEG embeddings
- Uses Latent Expert Queries to inject semantic knowledge back into neural representations
