[논문 리뷰] Long-Tailed Partial Label Learning via Dynamic Rebalancing

Hong Feng, Jiangchao Yao외 3명|arXiv (Cornell University)|2023. 02. 10.

Text and Document Classification Technologies인용 수 8

한 줄 요약

RECORDS는 학습 중 레이블 모호화에 적응하는 긴 꼬리 부분 레이블 학습을 위한 동적 재균형 메커니즘을 도입하여, 사전 클래스 분포를 필요로 하지 않고도 성능을 향상시킨다.

ABSTRACT

Real-world data usually couples the label ambiguity and heavy imbalance, challenging the algorithmic robustness of partial label learning (PLL) and long-tailed learning (LT). The straightforward combination of LT and PLL, i.e., LT-PLL, suffers from a fundamental dilemma: LT methods build upon a given class distribution that is unavailable in PLL, and the performance of PLL is severely influenced in long-tailed context. We show that even with the auxiliary of an oracle class prior, the state-of-the-art methods underperform due to an adverse fact that the constant rebalancing in LT is harsh to the label disambiguation in PLL. To overcome this challenge, we thus propose a dynamic rebalancing method, termed as RECORDS, without assuming any prior knowledge about the class distribution. Based on a parametric decomposition of the biased output, our method constructs a dynamic adjustment that is benign to the label disambiguation process and theoretically converges to the oracle class prior. Extensive experiments on three benchmark datasets demonstrate the significant gain of RECORDS compared with a range of baselines. The code is publicly available.

연구 동기 및 목표

• Motivate LT-PLL as a practical scenario with both label ambiguity and class imbalance.
• Identify limitations of constant rebalancing in LT-PLL even with oracle priors.
• Propose RECORDS, a dynamic, prior-free rebalancing method to stabilize label disambiguation.
• Theoretically analyze convergence and relation to oracle priors under small ambiguity.
• Demonstrate empirical gains across multiple LT-PLL benchmarks and baselines.

제안 방법

• Formulate LT-PLL with biased model outputs and candidate label sets.
• Introduce RECORDS that debiases outputs via a dynamic, parametric class distribution.
• Estimate dynamic class distribution P_train(y|Θ) by a momentum-updated prototype feature F (Eq. 5).
• Compute P_uni(y|x;Θ) by adjusting logits with -log P_train(y|Θ) (Eq. 6).
• Embed RECORDS as a lightweight module that can plug into existing PLL methods in an end-to-end manner.
• Provide theoretical insight showing convergence of dynamic rebalancing toward the oracle prior under small ambiguity (Proposition 4.1).

실험 결과

연구 질문

• RQ1 Can dynamic rebalancing improve LT-PLL without access to true class distributions?
• RQ2 How does dynamic rebalancing interact with label disambiguation during training?
• RQ3 Does RECORDS converge toward the oracle prior as training progresses under realistic ambiguity?
• RQ4 What are the empirical gains of RECORDS across standard LT-PLL benchmarks and PLL baselines?
• RQ5 Is RECORDS orthogonal and easily pluggable into existing PLL methods?

주요 결과

• RECORDS consistently improves over strong PLL baselines (e.g., CORR, PRODEN, LW, CAVL) on CIFAR-10-LT, CIFAR-100-LT, and PASCAL VOC under LT-PLL settings.
• RECORDS yields substantial gains over the best baseline, with up to 32.03% relative improvement on Pascal VOC and notable gains on CIFAR-10-LT/CIFAR-100-LT.
• The dynamic rebalancing converges toward the oracle class prior during training, and final estimates closely match the prior (L2 distance reduction).
• The method is lightweight, end-to-end, and orthogonal to existing PLL losses, and remains robust under non-uniform candidate generation.
• Linear probing experiments show RECORDS yields superior representations compared to baselines.
• RECORDS outperforms alternative dynamic strategies and maintains improved tail-class performance without over-adjusting to head classes.