| name | atoms-of-thought-eeg-microstates |
| description | Universal EEG representation learning using microstate tokenizers. Builds a discrete microstate tokenizer from large-scale EEG datasets by clustering continuous signals into sequences of quasi-stable brain activity patterns. Outperforms traditional time/frequency features across sleep staging, emotion recognition, and motor imagery. Applicable to EEG representation, BCI, sleep staging, emotion recognition, neuroinformatics. Activation: EEG microstates, universal EEG tokenizer, microstate clustering, brain state representation, EEG foundation model, microstate tokenization. |
| user-invocable | true |
Atoms of Thought: Universal EEG Representation Learning with Microstates
Source Paper: arXiv:2605.20182 - Atoms of Thought: Universal EEG Representation Learning with Microstates
Authors: Xinyang Tian, Ruitao Liu, Ziyi Ye, Siyang Xue, Xin Wang, Xuesong Chen
Published: 2026-05-19 (Accepted by MRAC 2025)
Core Methodology
1. Microstate Tokenizer
Treats EEG microstates as the "atoms" of brain activity, analogous to words in NLP or visual tokens in computer vision.
Process:
- Cluster continuous EEG signals into sequences of discrete microstates
- Learn from large medical EEG dataset
- Each microstate represents a quasi-stable brain activity pattern at microscopic time scales
2. Universal Representation Learning
The microstate tokenizer is universally applicable across diverse downstream tasks:
Tasks evaluated:
- Sleep staging
- Emotion recognition
- Motor imagery classification
Key finding: Microstate-based representations consistently outperform:
- Traditional time-domain features (power bands, Hjorth parameters)
- Frequency-domain features (PSD, spectral entropy)
- Raw signal inputs under the same model architectures
3. Interpretability and Scalability
Interpretability:
- Microstates map to known functional brain networks
- Each state corresponds to a distinct topographical pattern
- Clinically meaningful transitions between states
Scalability:
- Tokenizer can be pre-trained once and applied universally
- Efficient discrete representation reduces computational cost
- Compatible with transformer and sequence models
Key Contributions
- Novel representation: First systematic demonstration that microstates are effective as a universal EEG representation for deep learning
- Cross-task generality: Single tokenizer works across sleep, emotion, and motor tasks
- Interpretability: Microstates offer inherent neurophysiological interpretability compared to continuous features
- Efficiency: Discrete token sequences are more computationally efficient than continuous time-series
Implementation Notes
Building a Microstate Tokenizer
1. Preprocess EEG: bandpass filter (1-40 Hz), artifact removal
2. Compute GFP (Global Field Power) peaks
3. Extract topographies at GFP peak timepoints
4. Cluster topographies (k-means, typically 4-8 microstates)
5. Assign each timepoint to nearest microstate
6. Use resulting sequence of microstate labels as tokens
Model Architecture Options
- Transformer encoder on microstate token sequences
- LSTM/GRU on microstate embeddings
- CNN with microstate embedding + positional encoding
- Compatible with any sequence model architecture
Recommended Microstate Configurations
- Number of microstates: 4-8 (standard), can be task-dependent
- Time window: 60-120 ms per microstate
- Embedding dimension: 16-64 for downstream tasks
Related Skills
- [[eeg-brain-connectivity-bci]] - EEG connectivity analysis for BCI
- [[eeg-hopfield-emotion-energy]] - EEG emotion recognition via energy landscapes
- [[laya-eeg-foundation]] - LeJEPA approach to EEG representation learning
- [[dance-eeg-event-detection]] - End-to-end EEG event detection
