Spectral Gradient Surgery for Domain-Generalizable Dataset Distillation

In three linesSpectral Gradient Surgery (SGS) improves dataset distillation for out-of-distribution generalization. The method disentangles class-discriminative from domain-specific information in compressed synthetic data via spectral analysis of cross-domain gradients. SGS integrates as a plug-and-play module with existing Distribution Matching methods.

Summary generated by Claude — human-verified