Speed learning on the fly

TL;DR

This paper introduces an online method to adapt step sizes in stochastic gradient algorithms by performing gradient descent on the step size itself, improving practical performance without costly computations.

Contribution

It proposes a novel online step size adaptation technique for stochastic gradient methods that is computationally efficient and does not require backward passes over data.

Findings

Effective online step size adaptation demonstrated.

Reduces need for manual tuning of step sizes.

Applicable to various stochastic gradient algorithms.

Abstract

The practical performance of online stochastic gradient descent algorithms is highly dependent on the chosen step size, which must be tediously hand-tuned in many applications. The same is true for more advanced variants of stochastic gradients, such as SAGA, SVRG, or AdaGrad. Here we propose to adapt the step size by performing a gradient descent on the step size itself, viewing the whole performance of the learning trajectory as a function of step size. Importantly, this adaptation can be computed online at little cost, without having to iterate backward passes over the full data.