Unlock: Continuous-Time Gradient Flow (SLT View)

Gradient flow as the step-size-to-zero limit of gradient descent: an ODE on weight space. On least squares it converges to the minimum-norm OLS solution; early stopping is implicit ridge regularization; on overparameterized two-layer networks the mean-field limit yields the global-optimum convergence theorems of Mei-Montanari and Chizat-Bach.