Dataset Distillation by Influence Matching
Inf-Match trains a small synthetic dataset to match the full dataset’s effect on the final trained model parameters.
The paper argues this is different from matching gradients or training trajectories during optimization. Its influence estimator is described as fully differentiable, sample-level, and linear-time, avoiding inverse-Hessian products and convexity assumptions. On Tiny-ImageNet at IPC=10, it reports 31.5% accuracy, 4.7 points above NCFM. It also reports stronger Flickr30K vision-language distillation results than process-matching baselines. HF Daily Papers' note
The paper argues this is different from matching gradients or training trajectories during optimization. Its influence estimator is described as fully differentiable, sample-level, and linear-time, avoiding inverse-Hessian products and convexity assumptions. On Tiny-ImageNet at IPC=10, it reports 31.5% accuracy, 4.7 points above NCFM. It also reports stronger Flickr30K vision-language distillation results than process-matching baselines. HF Daily Papers' note
score 5