| name | rgflow-transformer-eeg |
| description | RG-Flow Transformer for encoding scale-free dynamics in scarce EEG data - uses renormalization-group inductive bias to improve interpretability and spectral exponent recovery from limited neural recordings |
| category | ai_collection/neuroscience |
RG-Flow Transformer: Encoding Scale-Free Dynamics in Scarce EEG
Core Methodology
The RG-Flow Transformer integrates renormalization-group (RG) principles into transformer architecture to better capture scale-free neural dynamics, particularly in scarce EEG data scenarios.
-
Architecture Components
- Standard self-attention mechanism for local temporal dependencies
- Scale-aware stream with learnable anomalous dimension γ
- Block-spin coarse-graining for multi-scale feature extraction
- Entropy-gated synchronization bridge for cross-scale communication
-
Theoretical Foundation
- Models brain field potentials as scale-free systems with 1/f^β power spectra
- Uses renormalization group theory to understand how neural dynamics scale across spatiotemporal scales
- The anomalous dimension γ quantifies deviation from classical scaling behavior
- Entropy-gated synchronization enables information flow across scales based on criticality
-
Training Procedure
- Train on PhysioNet Sleep-EDF corpus with leave-one-subject-out cross-validation
- Compare against parameter-matched vanilla transformer and hierarchy-only ablation
- Sweep per-subject data budget to identify inductive bias crossover point
- Evaluate sleep staging classification and out-of-sample spectral exponent (β) recovery
Implementation Steps
-
Model Architecture
- Implement standard transformer encoder layers
- Add parallel scale-aware processing stream with learnable γ parameter
- Implement block-spin coarse-graining layers (spatial/temporal pooling)
- Add entropy-gated synchronization mechanism between streams
- Combine outputs via learned gating mechanism
-
Training Protocol
- Use leave-one-subject-out cross-validation for subject generalization
- Train with standard cross-entropy loss for sleep stage classification
- Add auxiliary loss for β-recovery: L_β = ||γ_predicted - β_measured||²
- Vary training data availability per subject from 5% to 100%
- Use Adam optimizer with learning rate scheduling
-
Evaluation Metrics
- Sleep staging accuracy (5-class AASM)
- Spectral exponent recovery R² (out-of-sample β prediction)
- Comparison with vanilla transformer and hierarchy-only ablations
- Statistical significance testing (paired t-tests)
Key Findings from Paper
- RG-Flow and vanilla transformer show comparable sleep staging performance (77.3% vs 77.0% accuracy)
- No clear scarce-data crossover observed - vanilla model performs better at all data levels
- Key advantage: RG-Flow recovers continuous spectral exponent out-of-sample (β-recovery R² = 0.416)
- Vanilla architecture lacks any mechanism for spectral exponent estimation
- Provides interpretability link between model parameters and biophysical properties
Pitfalls and Limitations
- Computational overhead from dual-stream architecture
- Requires careful tuning of entropy-gating threshold
- Block-spin coarse-graining may lose fine-grained temporal information
- Performance advantage primarily in interpretability rather than prediction accuracy
- Limited evaluation on only 5 subjects from Sleep-EDF dataset
- May require adaptation for other neural signal types (ECoG, LFP, MEG)
Activation Keywords
- rgflow transformer
- renormalization group
- scale-free dynamics
- eeg spectral exponent
- scarce neural data
- brain criticality
- multi-scale neural processing
- sleep staging interpretation