Focal-SAM: Focal Sharpness-Aware Minimization for Long-Tailed Classification

About

Real-world datasets often follow a long-tailed distribution, making generalization to tail classes difficult. Recent methods resorted to long-tail variants of Sharpness-Aware Minimization (SAM), such as ImbSAM and CC-SAM, to improve generalization by flattening the loss landscape. However, these attempts face a trade-off between computational efficiency and control over the loss landscape. On the one hand, ImbSAM is efficient but offers only coarse control as it excludes head classes from the SAM process. On the other hand, CC-SAM provides fine-grained control through class-dependent perturbations but at the cost of efficiency due to multiple backpropagations. Seeing this dilemma, we introduce Focal-SAM, which assigns different penalties to class-wise sharpness, achieving fine-grained control without extra backpropagations, thus maintaining efficiency. Furthermore, we theoretically analyze Focal-SAM's generalization ability and derive a sharper generalization bound. Extensive experiments on both traditional and foundation models validate the effectiveness of Focal-SAM.

Sicong Li, Qianqian Xu, Zhiyong Yang, Zitai Wang, Linchao Zhang, Xiaochun Cao, Qingming Huang · 2025

Related benchmarks

Task	Dataset	Result
Long-Tailed Image Classification	ImageNet-LT (test)	Top-1 Acc (Overall)54.3	246
Image Classification	CIFAR-10-LT (IF 50)	Top-1 Accuracy82	88
Image Classification	CIFAR-100 LT (IF=50)	Top-1 Acc48.1	82
Image Classification	CIFAR-10-LT IF 100	Top-1 Accuracy77.2	78
Image Classification	CIFAR-100-LT Imbalance Factor 100 (test)	Top-1 Accuracy44	56
Image Classification	CIFAR-100 LT (IF=100)	Top-1 Acc44	55
Image Classification	CIFAR-LT-100 Imbalance Factor 50 (test)	Top-1 Accuracy48.1	54
Long-tailed recognition	iNaturalist 2018	Top-1 Accuracy71.8	31
Image Classification	CIFAR-10 LT (IF=200)	Top-1 Acc71.7	28
Image Classification	CIFAR-100-LT (IF=200)	Top-1 Acc39.6	27