查看文章

Nested optimization, whereby an optimization problem is constrained by the solutions of other optimization problems, has recently seen a surge in its application to Deep Learning. While the study of such problems started nearly a century ago in the context of market theory, many of the algorithms developed since do not scale to modern Deep Learning applications. In this thesis, I push the understanding and applicability of nested optimization to three machine learning domains: 1) adversarial games, 2) metalearning and 3) deep equilibrium models. For each domain, I tackle a particular goal. In 1) I adversarially learn model compression, in the case where training data isn’t available, in 2) I meta-learn hyperparameters for long optimization processes without introducing greediness, and in 3) I use deep equilibrium models to improve temporal coherence in video landmark detection.

The first part of my thesis deals with casting model compression as an adversarial game. Performing knowledge transfer from a large teacher network to a smaller student is a popular task in deep learning. However, due to growing dataset sizes and stricter privacy regulations, it is increasingly common not to have access to the data that was used to train the teacher. I propose a novel method which trains a student to match the predictions of its teacher without using any data or metadata. This is achieved by nesting the training optimization of the student with that of an adversarial generator, which searches for images on which the student poorly matches the teacher. These images are used to train the student in an online fashion. The student closely approximates its …