RG-Flow Transformer matches vanilla but reveals brain's scale-free dynamics
77% accuracy on sleep staging, but recovers brain's spectral exponent with R²=0.416
Brain field potentials follow scale-free power spectra (1/f^β) whose aperiodic exponent β tracks cortical state, notably sleep depth. The RG-Flow Transformer, introduced by Dibakar Sigdel, embeds this physical prior via a learnable anomalous dimension γ, block-spin coarse-graining, and an entropy-gated synchronization bridge. It was tested on the PhysioNet Sleep-EDF corpus with strict leave-one-subject-out cross-validation on 5 subjects (5 seeds each).
Against a parameter-matched vanilla transformer, RG-Flow achieved 77.3% vs 77.0% 5-class AASM sleep staging accuracy (paired p=0.294). The anticipated scarce-data crossover was absent—vanilla led at every data budget. However, RG-Flow recovers the continuous spectral exponent β out-of-sample (R²=0.416), a capability the vanilla model lacks entirely, making it a powerful interpretability tool for neuroscience.
- RG-Flow matches vanilla transformer on sleep staging (77.3% vs 77.0%, p=0.294)
- No scarce-data crossover advantage; vanilla leads at all data budgets
- RG-Flow uniquely recovers the continuous spectral exponent β out-of-sample (R²=0.416)
Why It Matters
Interpretable AI that reveals brain dynamics from scarce EEG could transform sleep medicine and neuroscience.