Learning to Imitate Datasets by Long-Range Matching of Expert Training Trajectories

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 thesis, we present 2 methods. First, our base method of 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. This method can also be augmented to produce infinitely tileable class-based textures. Our second method introduces a Deep Generative Prior to the distillation process. Rather than directly optimizing pixels, this method now distills into the latent space of a pre-trained generative model. This greatly reduces overfitting to the backbone model and allows our distilled data to generalize much better to unseen architectures.