Research map / Label-efficient and robust learning / Recognition and long-tail
SSE-SAM: Balancing Head and Tail Classes Gradually Through Stage-Wise SAM
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SSE-SAM, a two-stage sharpness-aware training method that first applies SAM-style perturbation to head classes and then switches to ImbSAM (tail-only perturbation) to balance head and tail saddle escaping on long-tailed data.
Problem
Long-tailed classification where SAM is dominated by head-class gradients and under-serves tail classes, while ImbSAM improves tail classes but under-trains head classes; the paper seeks a balance between the two.
Approach
Combine SAM and ImbSAM in a stage-wise schedule: the first phase uses SAM (head and tail perturbations) to help head classes escape saddles, and the second phase removes the head perturbation (ImbSAM) to help tail classes escape, with a transition hyperparameter gamma.
Key results
- On CIFAR-100-LT (IF=100), SSE-SAM with CE reaches 49.3% overall accuracy vs. 46.2% for SAM and 47.9% for ImbSAM.
- With LA loss on CIFAR-10-LT, SSE-SAM reaches 85.1% overall accuracy vs. 83.1% for LA+SAM and 82.8% for LA+ImbSAM.
- On ImageNet-LT, CE+SSE-SAM achieves 47.5% overall accuracy vs. 46.0% for CE+SAM and 46.1% for CE+ImbSAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SSE-SAM rather than ImbSAM | CE · CIFAR-100-LT (IF=100) | Avg accuracy: 47.6 -> 48.3 | Helped |
| Apply SSE-SAM rather than SAM | CE · CIFAR-100-LT (IF=50) | Avg accuracy: 52.1 -> 54.5 | Helped |
| Apply SSE-SAM rather than DRO-LT | CE · ImageNet-LT | Avg accuracy: 53.5 -> 47.5 | Hurt |
| Use the two-stage schedule instead of SAM (ext.) (ablation) | CE · CIFAR-100-LT (IF=100, ρtail=0.10) | Avg accuracy: 47.8 -> 49.3 | Helped |
| Use the two-stage schedule instead of ImbSAM (ablation) | CE · CIFAR-100-LT (IF=100, ρtail=0.10) | Avg accuracy: 47.9 -> 49.3 | Helped |
| Use the two-stage schedule instead of SAM (ext.) (ablation) | CE · CIFAR-100-LT (IF=50, ρtail=0.10) | Avg accuracy: 53.4 -> 55.6 | Helped |
| Use the two-stage schedule instead of ImbSAM (ablation) | CE · CIFAR-100-LT (IF=50, ρtail=0.10) | Avg accuracy: 54.0 -> 55.6 | Helped |
| Use SSE-SAM instead of SAM (ablation) | CE · CIFAR-100-LT (IF=100) | Many accuracy: 75.3 -> 73.1 | Hurt |
| Use SSE-SAM instead of SAM (ablation) | CE · CIFAR-100-LT (IF=100) | Few accuracy: 12.9 -> 15.8 | Helped |
| Apply SSE-SAM rather than SAM to VS | VS · CIFAR-100-LT (IF=100) | Avg accuracy: 51.8 -> 53.5 | Helped |
| Apply SSE-SAM rather than SAM to LA | LA · CIFAR-10-LT (IF=100) | Avg accuracy: 83.1 -> 85.1 | Helped |
Surprising findings
- SSE-SAM outperforms ImbSAM on medium and few classes even though its second stage is ImbSAM-like, which the authors say exceeds their expectations and lacks theoretical explanation.
- ImbSAM underperforms SAM on head classes, indicating that removing head-class perturbation hurts head-class generalization.
Limitations
- The unexpectedly good tail-class performance of SSE-SAM over ImbSAM lacks theoretical explanation.
- Only SAM and ImbSAM are considered as smoothness regularizers; other smoothness methods are not explored.
- Hyperparameters (rho, gamma) are selected per setting based on overall accuracy, with limited sweeps.
Experimental setup
Datasets: CIFAR-100-LT CIFAR-10-LT ImageNet-LT
Benchmarks: CIFAR-100-LT CIFAR-10-LT ImageNet-LT
Metrics: overall accuracy many-shot accuracy medium-shot accuracy few-shot accuracy lambda ratio (non-convexity measure)
Baselines: SAM ImbSAM CE LDAM LA VS CB tau-norm cRT LWS DisAlign DRO-LT
Builds on: Sharpness-Aware Minimization Imbalanced SAM LDAM logit adjustment Vector Scaling loss
Closest papers by meaning
- Focal-SAM: Focal Sharpness-Aware Minimization for Long-Tailed Classification (ICML 2025 · similarity 0.91)
- LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups (NeurIPS 2025 · similarity 0.89)
- Implicit Regularization of Sharpness-Aware Minimization for Scale-Invariant Problems (NeurIPS 2024 · similarity 0.88)
- Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning (ICLR 2024 · similarity 0.88)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.87)
- Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.