Classical worst-case optimization theory neither explains the success of optimization in machine learning, nor does it help with step size selection. We establish a connection between Bayesian Optimization (i.e. average case optimization theory) and classical optimization using a 'stochastic Taylor approximation' to rediscover gradient descent. This rediscovery yields a step size schedule we call Random Function Descent (RFD), which, in contrast to classical derivations, is scale invariant. Furthermore, our analysis of RFD step sizes yields a theoretical foundation for common step size heuristics such as gradient clipping and gradual learning rate warmup. We finally propose a statistical procedure for estimating the RFD step size schedule and validate this theory with a case study on the MNIST dataset.
View on arXiv