Speed learning on the fly
This addresses a practical bottleneck for machine learning practitioners using stochastic gradient methods, offering an incremental improvement to automate step size tuning.
The paper tackles the problem of manually tuning step sizes in online stochastic gradient descent algorithms, which is tedious in practice, by proposing an online adaptation method that optimizes the step size as a function of the learning trajectory, resulting in a low-cost solution without requiring backward passes over full data.
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.