Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition

TL;DR

This work tackles long-tailed semi-supervised recognition by introducing a probabilistic framework that unifies existing long-tail learning methods through Gaussian kernel density estimation. It then extends this framework to LTSSL with continuous contrastive learning (CCL), combining reliable pseudo-labels and smoothed pseudo-labels propagated from nearby samples to mitigate confirmation bias. The method employs a dual-branch network, logit adjustment with estimated class priors, energy-based data selection for calibrated pseudo-labels, and a three-term objective that couples classification with both reliable and smoothed contrastive losses. Across CIFAR-LT, STL10-LT, and ImageNet-127, CCL achieves state-of-the-art results, including around 4% gains on ImageNet-127, demonstrating robust representation learning under unknown unlabeled data distributions.

Abstract

Long-tailed semi-supervised learning poses a significant challenge in training models with limited labeled data exhibiting a long-tailed label distribution. Current state-of-the-art LTSSL approaches heavily rely on high-quality pseudo-labels for large-scale unlabeled data. However, these methods often neglect the impact of representations learned by the neural network and struggle with real-world unlabeled data, which typically follows a different distribution than labeled data. This paper introduces a novel probabilistic framework that unifies various recent proposals in long-tail learning. Our framework derives the class-balanced contrastive loss through Gaussian kernel density estimation. We introduce a continuous contrastive learning method, CCL, extending our framework to unlabeled data using reliable and smoothed pseudo-labels. By progressively estimating the underlying label distribution and optimizing its alignment with model predictions, we tackle the diverse distribution of unlabeled data in real-world scenarios. Extensive experiments across multiple datasets with varying unlabeled data distributions demonstrate that CCL consistently outperforms prior state-of-the-art methods, achieving over 4% improvement on the ImageNet-127 dataset. Our source code is available at https://github.com/zhouzihao11/CCL

Figures (9)

• Figure 1: Comparison of class prior estimation error and ECE on CIFAR100-LT.
• Figure 2: Generalize to more realistic LTSSL settings for ACR and CCL on CIFAR10/100-LT dataset in fixed $\gamma_l$ and various $\gamma_u$ settings.
• Figure 3: Illustration of the proposed framework.
• Figure 4: Illustration of reliable pseudo-labels and smoothed pseudo-labels in CCL. To generalize the framework in Section 2 to LTSSL, the main challenge is unknown $\mathbb{P}_{u}(Y = y \mid \boldsymbol{x}^u)$, where $\boldsymbol{x}^u$ denotes a sample in the unlabeled dataset. We first approximate $\mathbb{P}_{u}(Y = y \mid \boldsymbol{x}^u)$ using the output of the calibrated and integrated classifier and use energy score to filter out reliable unlabeled data, ensuring the model's calibration, which constitutes the reliable pseudo-labels subset. Furthermore, we can also estimate the unknown $\mathbb{P}_{u}(Y = y \mid \boldsymbol{x}^u)$ by leveraging the smoothness assumption. Specifically, we construct smoothed pseudo-labels by propagating labels from nearby samples using the Gaussian kernel density estimation.
• Figure 5: Sensitive analysis of hyperparameters under consistent setting of CIFAR100-LT.