Logarithmic landscape and power-law escape rate of SGD

Stochastic gradient descent (SGD) undergoes complicated multiplicative noise for the mean-square loss. We use this property of the SGD noise to derive a stochastic differential equation (SDE) with simpler additive noise by performing a non-uniform transformation of the time variable. In the SDE, the gradient of the loss is replaced by that of the logarithmized loss. Consequently, we show that, near a local or global minimum, the stationary distribution $P_\mathrm{ss}(\theta)$ of the network parameters $\theta$ follows a power-law with respect to the loss function $L(\theta)$, i.e. $P_\mathrm{ss}(\theta)\propto L(\theta)^{-\phi}$ with the exponent $\phi$ specified by the mini-batch size, the learning rate, and the Hessian at the minimum. We obtain the escape rate formula from a local minimum, which is determined not by the loss barrier height $\Delta L=L(\theta^s)-L(\theta^*)$ between a minimum $\theta^*$ and a saddle $\theta^s$ but by the logarithmized loss barrier height $\Delta\log L=\log[L(\theta^s)/L(\theta^*)]$. Our escape-rate formula explains an empirical fact that SGD prefers flat minima with low effective dimensions.