RG-Flow Transformer Shows Interpretability for Scarce EEG Data

Brain field potentials exhibit scale-free properties, with their power spectra following a 1/f^beta law where the exponent beta indicates cortical state, particularly sleep depth. Researchers investigated whether a specialized RG-Flow Transformer, which incorporates a renormalization-group inductive bias, could outperform a standard transformer on real, limited EEG data for sleep staging. The RG-Flow Transformer couples self-attention with a scale-aware stream, block-spin coarse-graining, and an entropy-gated synchronization bridge. Benchmarking on the PhysioNet Sleep-EDF corpus for 5-class AASM sleep staging showed that both the RG-Flow and vanilla transformers achieved statistically indistinguishable accuracy (around 77%). Surprisingly, the predicted advantage of the RG-Flow in scarce data scenarios did not materialize, with the vanilla model slightly ahead at every data-limited budget. However, a key differentiator emerged in interpretability: the RG-Flow model successfully recovered the continuous spectral exponent (beta) out-of-sample with an R^2 of 0.416, a capability the vanilla architecture lacks. This suggests that while not superior in raw classification accuracy for this task, the RG-Flow offers valuable insights into underlying physiological processes.