Appau Gideon Kofi Amo

Appau Gideon Kofi Amo

M.S. student at USTC working on optimizer-level methods for long-tailed visual recognition and on world models for decision-making.

Class-Adaptive Focal Sharpness-Aware Minimization for Long-Tailed Visual Recognition

Published in Manuscript in preparation (target: CVPR), 2026

Standard Sharpness-Aware Minimization applies a uniform perturbation to all classes and ignores how training-data scarcity changes optimization difficulty. CA-Focal-SAM instead adapts how class contributions enter the nominal and perturbed gradient components of SAM:

  • GammaAdaptor: just 2 learnable scalars map class frequency to per-class focal parameters, so rare classes receive a larger effective SAM weight.
  • Quadratic perturbation-radius curriculum: a smooth \(\rho(t)\) schedule that yields flatter minima, verified by Hessian spectral analysis (\(\lambda_{\max}\) 114.7 vs. Focal-SAM 121.4 via stochastic Lanczos quadrature).

Evaluated with a pre-registered multi-seed protocol (3 seeds, mean ± std) on CIFAR-10/100-LT, ImageNet-LT, and CUB-200-LT, with 20+ qualitative analyses (t-SNE embeddings, 2D/3D loss landscapes, confusion matrices). Code and experiment logs: github.com/Amo-Gideon/sfca-ltr.

Recommended citation: Appau, Gideon Kofi Amo, and Wang, Pengkun. (2026). "Class-Adaptive Focal Sharpness-Aware Minimization for Long-Tailed Visual Recognition." Manuscript in preparation.
Download Paper