Stochastic Gradient Descent as Approximate Bayesian Inference

Stochastic Gradient Descent with a constant learning rate (constant SGD) simulates a Markov chain with a stationary distribution. With this perspective, we derive several new results. (1) We show t...