Spectral Renormalization-Group Theory Explains Neural Network Layer Convergence

This research links deep neural network layer convergence to Wilsonian Renormalization-Group ideas, identifying an optimal power-law spectral exponent of alpha=2. This value serves as a boundary between stable learning and overfitting by satisfying a scale-invariant marginal condition.