AES: Curing Optimizer Blindness in Long-Tailed Recognition via State-Aware Correction

Long-tailed recognition fundamentally suffers from optimizer blindness where the optimization process mistakenly conflates the magnitude of gradient accumulation with the scarcity of semantic information. Existing strategies relying on static frequency-based priors fail to correct this bias and result in state blindness regarding supervision and micro-level blindness regarding parameter updates. To address these limitations, we propose the AES framework to establish a dynamic and state-aware correction system across the entire learning lifecycle. We specifically introduce Adaptive Residual Supervision loss to act as a real-time reality check for supervision completeness via precision shielding. We also propose Entropy-aware PCGrad to resolve parameter-level conflicts by quantifying task specificity through gradient entropy. Additionally, we devise Sample-level Conflict Arbitrated Fusion to serve as a dynamic inference arbiter that routes predictions based on instance difficulty. Extensive experiments on CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 demonstrate that our method consistently achieves state-of-the-art performance by effectively balancing head-class stability and tail-class discrimination. Code is available at here

Recognition tasks with long-tailed data face a common training defect. The optimization process cannot accurately distinguish sparse semantic information, bringing incorrect updating tendencies. Conventional frequency-based methods fail to fix this flaw, and cause incomplete supervision and improper parameter adjustment. We put forward a novel framework to dynamically revise biases in the whole training procedure. It verifies valid supervision information, eliminates internal conflicting signals during parameter learning, and makes adaptive judgment according to varying sample complexity. Evaluated on standard long-tailed datasets, our approach surpasses existing solutions. It maintains stable performance on frequent categories and enhances identification capability for scarce samples, realizing balanced and accurate recognition results.