FreqCore: A Frequency Domain Perspective on Coreset Selection

Coreset selection aims to select an informative subset that achieves performance comparable to the whole dataset. However, existing methods are often susceptible to spurious correlations present in high-frequency domain, consequently leading to inaccurate sample importance estimation and suboptimal coresets. To address the problem, we propose FreqCore, a novel dynamic coreset selection method that focuses on the robust Low-Frequency Property (LFP) of hidden representations. Building upon stable low-frequency components, FreqCore inherently mitigates spurious high-frequency correlations, enabling a more accurate estimation of sample importance. Our FreqCore is directly inspired by the key discovery that the Global Commonality Subspace (GCS) projection of LFP effectively reveals Information Redundancy (IR). Based on the observation, we mathematically formulate a criterion that prunes high-IR samples, as they make minimal contribution to dataset diversity, and retains low-IR samples that capture unique information. Without introducing additional constraints, FreqCore significantly reduces computational costs. Extensive experiments demonstrate the superior generalization capability of FreqCore. For example, with ResNet-18 on CIFAR-100 at a 70% pruning rate, FreqCore achieves 79.08% accuracy, surpassing the full dataset by 0.88%. The excellent performance also extends to Transformer architectures. With ViT-Ti-16 on CIFAR-100 with a 50% pruning ratio, FreqCore achieves 81.28% accuracy, outperforming the full dataset by 0.44%.