Understanding Generalization and Stochastic Gradient Descent.
This paper tackles two related questions at the heart of machine learning; how can we predict if a minimum will generalize to the test set, and why does stochastic gradient descent find minima that generalize well? Our work is inspired by Zhang et al. (2017), who showed deep networks can easily memorize randomly labeled training data, despite generalizing well when shown real labels of the same inputs. We show here that the same phenomenon occurs in small linear models. These observations are explained by evaluating the Bayesian evidence in favor of each model, which penalizes sharp minima. Next, we explore the "generalization gap" between small and large batch training, identifying an optimum batch size which maximizes the test set accuracy. Noise in the gradient updates is beneficial, driving the dynamics towards robust minima for which the evidence is large. Interpreting stochastic gradient descent as a stochastic differential equation, we predict the optimum batch size is proportional to both the learning rate and the size of the training set, and verify these predictions empirically.
Publisher URL: http://arxiv.org/abs/1710.06451
Keeping up-to-date with research can feel impossible, with papers being published faster than you'll ever be able to read them. That's where Researcher comes in: we're simplifying discovery and making important discussions happen. With over 19,000 sources, including peer-reviewed journals, preprints, blogs, universities, podcasts and Live events across 10 research areas, you'll never miss what's important to you. It's like social media, but better. Oh, and we should mention - it's free.
Researcher displays publicly available abstracts and doesn’t host any full article content. If the content is open access, we will direct clicks from the abstracts to the publisher website and display the PDF copy on our platform. Clicks to view the full text will be directed to the publisher website, where only users with subscriptions or access through their institution are able to view the full article.