Towards Realistic Long-tailed Semi-supervised Learning in an Open World

TL;DR

This work tackles realism in open-world long-tailed semi-supervised learning by introducing Realistic Open-world Long-tailed SSL (ROLSSL), where labeled data for known classes is scarce and unlabeled data—including novel classes—may have a different distribution, with no assumed relationship between known and novel categories. It proposes Dual-Stage Post-hoc Logit Adjustment (DPLA), which first adjusts logits for labeled known classes by incorporating factors such as the number of classes and dataset size, and then adjusts unlabeled-class logits based on predicted class frequencies to suppress majority classes and emphasize minority ones. The method integrates a refined pseudo-label strategy and masking within an overall objective $ ext{L}_{rolssl} = ext{L}_{pair} + ext{λ}_1 ext{L}_{ce} + ext{λ}_2 ext{L}_{b ext{_ce}} + ext{L}_{reg}$, achieving advanced performance on CIFAR-10/100, SVHN, ImageNet-100, Tiny ImageNet, and Oxford-IIIT Pet, with improvements up to 50.1% in certain settings. The results establish DPLA as a strong, simple baseline for ROLSSL and highlight the importance of distribution-aware logit adjustments in open-world, long-tailed SSL scenarios; code is publicly available.

Abstract

Open-world long-tailed semi-supervised learning (OLSSL) has increasingly attracted attention. However, existing OLSSL algorithms generally assume that the distributions between known and novel categories are nearly identical. Against this backdrop, we construct a more \emph{Realistic Open-world Long-tailed Semi-supervised Learning} (\textbf{ROLSSL}) setting where there is no premise on the distribution relationships between known and novel categories. Furthermore, even within the known categories, the number of labeled samples is significantly smaller than that of the unlabeled samples, as acquiring valid annotations is often prohibitively costly in the real world. Under the proposed ROLSSL setting, we propose a simple yet potentially effective solution called dual-stage post-hoc logit adjustments. The proposed approach revisits the logit adjustment strategy by considering the relationships among the frequency of samples, the total number of categories, and the overall size of data. Then, it estimates the distribution of unlabeled data for both known and novel categories to dynamically readjust the corresponding predictive probabilities, effectively mitigating category bias during the learning of known and novel classes with more selective utilization of imbalanced unlabeled data. Extensive experiments on datasets such as CIFAR100 and ImageNet100 have demonstrated performance improvements of up to 50.1\%, validating the superiority of our proposed method and establishing a strong baseline for this task. For further researches, the anonymous link to the experimental code is at \href{https://github.com/heyuanpengpku/ROLSSL}{\textcolor{brightpink}{https://github.com/heyuanpengpku/ROLSSL}}

Figures (5)

• Figure 1: The overview of the ROLSSL setting and the Dual-stage Post-hoc Logit Adjustment method. On the left, the dataset composition within the ROLSSL framework is illustrated. On the right, the overall process of the Dual-stage Post-hoc Logit Adjustment is shown. In the first stage of logit adjustment, factors such as the number of classes, sample frequency, and overall dataset size are considered to encourage a larger relative margin between the logits of rare and dominant labels. In the second stage, the predicted class frequencies are used to adjust the logits for the unlabeled data, further guiding the model to focus on learning from predicted minority class samples and reducing the attention given to samples from the predicted majority classes.
• Figure 2: Figures (a) and (c) show the t-SNE visualizations of OpenLDN on the CIFAR-10 and SVHN datasets, respectively. Figures (b) and (d) present the t-SNE visualizations of the proposed method on the CIFAR-10 and SVHN datasets. It is evident that DPLA demonstrates better recognition performance compared to OpenLDN.
• Figure 3: Ablation study on the performance of CIFAR-100 for the Scaling Factor.
• Figure 4: Ablation study on the performance of ImageNet-100 for the Scaling Factor.
• Figure 5: OSSL method performance in ROLSSL settings.