Dataset Distillation by Influence Matching

Influence Matching is a new dataset distillation method that learns synthetic data by matching the final effect of the full dataset on model parameters.

Instead of aligning intermediate optimization steps, it directly matches training outcomes, which improves performance over prior process-matching methods.

The approach reports stronger results on image classification benchmarks such as Tiny-ImageNet and on Flickr30K image-text retrieval.

It uses a linear-time influence estimator to make the distillation process more efficient with fewer synthetic samples.