RICASSO: Reinforced Imbalance Learning with Class-Aware Self-Supervised Outliers Exposure

TL;DR

This research shows that data mixing, a potent augmentation technique for long-tailed recognition, can generate pseudo-OOD data that exhibit the features of both in-distribution (ID) data and OOD data, so that by using mixed data instead of real OOD data, this framework can address long-tailed recognition and OOD detection holistically.

Abstract

In real-world scenarios, deep learning models often face challenges from both imbalanced (long-tailed) and out-of-distribution (OOD) data. However, existing joint methods rely on real OOD data, which leads to unnecessary trade-offs. In contrast, our research shows that data mixing, a potent augmentation technique for long-tailed recognition, can generate pseudo-OOD data that exhibit the features of both in-distribution (ID) data and OOD data. Therefore, by using mixed data instead of real OOD data, we can address long-tailed recognition and OOD detection holistically. We propose a unified framework called Reinforced Imbalance Learning with Class-Aware Self-Supervised Outliers Exposure (RICASSO), where "self-supervised" denotes that we only use ID data for outlier exposure. RICASSO includes three main strategies: Norm-Odd-Duality-Based Outlier Exposure: Uses mixed data as pseudo-OOD data, enabling simultaneous ID data rebalancing and outlier exposure through a single loss function. Ambiguity-Aware Logits Adjustment: Utilizes the ambiguity of ID data to adaptively recalibrate logits. Contrastive Boundary-Center Learning: Combines Virtual Boundary Learning and Dual-Entropy Center Learning to use mixed data for better feature separation and clustering, with Representation Consistency Learning for robustness. Extensive experiments demonstrate that RICASSO achieves state-of-the-art performance in long-tailed recognition and significantly improves OOD detection compared to our baseline (27% improvement in AUROC and 61% reduction in FPR on the iNaturalist2018 dataset). On iNaturalist2018, we even outperforms methods using real OOD data. The code will be made public soon.

Figures (2)

• Figure 1: Ours vs. other long-tailed recognition (LTR) and long-tailed OOD (LTOOD) detection methods on CIFAR-10, IR100. Without the need for real OOD data, our RICASSO outperforms all of the other methods on the iNaturalist2018 iNature2018 dataset.
• Figure 2: The overview of RICASSO. (\ref{['fig:pipeline']}) The pipeline of RICASSO. In the data mixing phase, both MixupMixup2018 and Cutmix CutMix2019 is performed to generate the mixed data $\mathcal{D}_{\text{mix}}$. Both the ID data and mixed data are used as training data. The backbone is a Mixture of Experts (MoE) network. (\ref{['fig:modules']}) The elaboration of each module. AALA (Green) takes OOD data as the extreme under-represented tailed classes data and thus uses the ambiguity of features to re-calibrate the ID priors for each expert. The calibrated logits are then fed to NOD to conduct a pseudo outlier exposure (Pink) without the need for real-OOD data. Finally, to further boost the feature clustering, the mixed data is also used for contrastive learning (Blue) and representation learning (Yellow).