Dataset condensation reflecting the training dynamics

Abstract

In recent advancements in deep learning, there is a continuous generation of large datasets, and mitigating the costs associated with handling these datasets is a significant concern. Dataset condensation is a research area generating small size of synthetic dataset preserving information of original datasets. Among the various approaches to creating synthetic data, we discuss issues with methodologies that incorporate information from original datasets during the network training process. As an enhancement strategy, we propose an algorithm that generates synthetic data reflecting the training dynamics of the network and evaluate its performance. We showed that a methodology reflecting the training dynamics has the potential to generate synthetic data that more accurately captures the information of original datasets.