Self-Supervised Task Discovery

Abstract

While large datasets, such as ImageNet, have propelledthe performance of deep neural networks, they are expen-sive and time-consuming to curate. This has motivated theinterest in low-cost self-supervised tasks that can generategeneral-purpose features, e.g. that support ImageNet clas-sification, for the case of vision, without requiring manuallylabeled data. We propose a framework for automaticallydiscovering a self-supervised task that results in powerfulvisual features. Our approach uses a model, the proposer,to generate the parameters that define the task. Traininga network, the solver, to solve this task results in metricsabout how good the task is. Using this metric as a rewardsignal, we can then update the proposer to generate betterand better tasks. We apply our method to rotation predic-tion, where the task is to recognize which rotation angle,from a set, was applied to an image, and demonstrate thatthe algorithm is able to find a set of angles that achievecomparable performance to state-of-the-art self-supervisedmethods.