Title: Learning to Distill Datasets by Matching Expert Training Trajectories

Project Page: https://georgecazenavette.github.io/mtt-distillation/

Abstract:
Dataset distillation is the task of synthesizing a small dataset such that a model trained on the synthetic set will match the test accuracy of the model trained on the full dataset. In this talk, we review 3 several of our recent papers in this field, starting with our introduction of a new state-of-the-art method: Matching Training Trajectories. Given a network, we train it for several iterations on our distilled data and optimize the distilled data with respect to the distance between the synthetically trained parameters and the parameters trained on real data. To efficiently obtain the initial and target network parameters for large-scale datasets, we pre-compute and store training trajectories of expert networks trained on the real dataset. Our method handily outperforms existing methods and also allows us to distill higher-resolution visual data. We also review another of our papers that shows how our new method can be used to synthesize class-based tileable textures and a third paper that augments our method with a GAN to increase the generality of our distilled images.

Committee:
Jun-Yan Zhu (co-advisor)
Simon Lucey (co-advisor)
Kris Kitani
Nathaniel Chodosh