Statistical Mechanics Model Describes Learning Dynamics in Function Space

A new statistical-mechanical description of deep learning defines error dynamics through a learning operator and a statistical operator. The framework shows that minimizing error requires pairing large eigenvalues of the learning operator with small eigenvalues of the local curvature.