Applying sharpness-aware minimization to the outer loop of trajectory-matching dataset condensation yields a few percentage points of accuracy improvement on image benchmarks.
For a fair comparison, the hyperparameters of each method are properly tuned for the adaption to all the tasks including Cifar100 with 1 IPC and Tiny ImageNet with 3 IPC
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Enhancing Generalization via Sharpness-Aware Trajectory Matching for Dataset Condensation
Applying sharpness-aware minimization to the outer loop of trajectory-matching dataset condensation yields a few percentage points of accuracy improvement on image benchmarks.