Class Confidence Aware Reweighting for Long Tailed Learning

TL;DR

The paper tackles long-tailed recognition by addressing the imbalance not only at the data or decision level but directly within the optimization objective. It introduces Class-Confidence Aware Reweighting (CCAR), a loss-level reweighting term with Ω(p_t, f_c) derived from a maximum-entropy margin framework, and a dual-phase coupling that adapts to prediction confidence. The approach yields an exponential weighting Ω(p_t, f_c) ∝ (e − f_c′(p_t))^{ω − p_t} that amplifies gradients for low-confidence tail samples and suppresses gradients for high-confidence head samples, while maintaining continuity and stable gradient dynamics. Extensive experiments on CIFAR-100-LT, ImageNet-LT, and iNaturalist2018 demonstrate improved tail accuracy and compatibility with logit-adjustment methods, confirming its value as a lightweight, single-stage complement to existing long-tailed learning strategies.

Abstract

Deep neural network models degrade significantly in the long-tailed data distribution, with the overall training data dominated by a small set of classes in the head, and the tail classes obtaining less training examples. Addressing the imbalance in the classes, attention in the related literature was given mainly to the adjustments carried out in the decision space in terms of either corrections performed at the logit level in order to compensate class-prior bias, with the least attention to the optimization process resulting from the adjustments introduced through the differences in the confidences among the samples. In the current study, we present the design of a class and confidence-aware re-weighting scheme for long-tailed learning. This scheme is purely based upon the loss level and has a complementary nature to the existing methods performing the adjustment of the logits. In the practical implementation stage of the proposed scheme, we use an Ω(p_t, f_c) function. This function enables the modulation of the contribution towards the training task based upon the confidence value of the prediction, as well as the relative frequency of the corresponding class. Our observations in the experiments are corroborated by significant experimental results performed on the CIFAR-100-LT, ImageNet-LT, and iNaturalist2018 datasets under various values of imbalance factors that clearly authenticate the theoretical discussions above.

Figures (3)

• Figure 1: Gradient Modulation Analysis. Comparison of effective gradient magnitudes. The proposed method (red) amplifies gradients in the low-confidence region ($p_t < \omega$) to accelerate learning for hard samples, while suppressing gradients in the high-confidence region ($p_t > \omega$) to prevent overfitting to easy head-class samples.
• Figure 2: Three-dimensional wireframe plot of weighting function $\Omega(p_t,f_c)$. The above surface shows how the weight varies with prediction confidence $p_t$ for varying class frequencies $f_c$. The pivot point $\omega$ helps distinguish between the low and high confidence regions, indicating asymmetric pattern for tail and head classes with smooth transitions.
• Figure 3: Performance by Frequency Group.

Theorems & Definitions (2)

• Lemma 1: Gradient Derivation
• proof