Dataset Distillation by Influence Matching

Influence Matching (Inf-Match) reframes dataset distillation around the final effect of data on converged model parameters instead of matching per-step gradients or training trajectories. It introduces a differentiable sample-level influence estimator based on unrolled optimization and a first-order Taylor approximation, avoiding inverse-Hessian products and convexity assumptions. The authors report 31.5% accuracy on Tiny-ImageNet at IPC=10, 4.7 percentage points above NCFM, and extend the method to Flickr30K vision-language distillation, where the stated average retrieval result exceeds NCFM by 2.5 percentage points.