Dataset Distillation by Matching Training Trajectories

TL;DR

This paper introduces a novel dataset distillation method that matches training trajectories of networks to create small, effective synthetic datasets, outperforming existing approaches and enabling higher-resolution data distillation.

Contribution

The paper proposes a trajectory matching approach for dataset distillation that improves performance and scalability, allowing higher-resolution data synthesis.

Findings

Outperforms existing dataset distillation methods.

Enables distillation of higher-resolution visual data.

Efficiently matches training trajectories for better synthetic data quality.

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 paper, we propose a new formulation that optimizes our distilled data to guide networks to a similar state as those trained on real data across many training steps. 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.