FAST: Topology-Aware Frequency-Domain Distribution Matching for Coreset Selection

Researchers propose FAST, a new DNN-free framework for coreset selection that compresses large datasets into representative subsets for training deep neural networks. The method uses frequency-domain distribution matching and achieves 9.12% average accuracy improvement while reducing power consumption by 96.57% compared to existing methods.

• →FAST introduces the first DNN-free distribution-matching framework for coreset selection, avoiding model-specific architectural bias.
• →The method employs Characteristic Function Distance (CFD) in frequency domain to capture full distributional information more accurately than traditional metrics.
• →A new Attenuated Phase-Decoupled CFD addresses the "vanishing phase gradient" issue in medium and high-frequency regions.
• →Progressive Discrepancy-Aware Sampling strategy schedules frequency selection from low to high for better convergence and global structure preservation.
• →FAST achieves 9.12% average accuracy gain, 96.57% power consumption reduction, and 2.2x speedup compared to baseline coreset methods.