Research map / Continual learning / Online and replay-based
Towards Macro-AUC Oriented Imbalanced Multi-Label Continual Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RLDAM loss (reweighted univariate loss with label-distribution-aware margins) plus a Weight Retain Updating memory strategy for Macro-AUC-oriented imbalanced multi-label class-incremental learning, with generalization bounds for batch MLL and MLCL.
Problem
Multi-label continual learning suffers from label imbalance (within and between labels), and cross-entropy-style losses yield low Macro-AUC; no prior work directly optimizes Macro-AUC in MLCL.
Approach
Combine the reweighted univariate (RU) loss, which surrogates Macro-AUC, with label-distribution-aware margins (LDAM) to form the RLDAM loss, and pair it with a memory update (WRU) that stores per-class positive/negative counts so reweighting factors match the full dataset.
Key results
- Ours achieves Macro-AUC of 88.69 on C-VOC, 77.93 on C-COCO and 79.77 on C-NUS, versus best baseline OCDM 79.98, OCDM 74.79, KRT 75.23.
- Ablation: adding RLDAM components and WRU raises C-VOC Macro-AUC from 79.17 (ER) to 88.69.
- Ours outperforms ER across memory sizes 200 to 2000 and is less sensitive to memory size.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RLDAM with WRU instead of the strongest listed baseline | Replay-based MLCL · C-VOC | Macro-AUC: 79.98 -> 88.69 | Helped |
| Use RLDAM with WRU instead of the strongest listed baseline | Replay-based MLCL · C-COCO | Macro-AUC: 74.79 -> 77.93 | Helped |
| Use RLDAM with WRU instead of the strongest listed baseline | Replay-based MLCL · C-NUS | Macro-AUC: 75.23 -> 79.77 | Helped |
| Add reweighted loss to plain ER with BCE (ablation) | ER with BCE · C-VOC, C-COCO, and C-NUS | Macro-AUC: 79.17 -> 83.04; 72.68 -> 74.86; 69.31 -> 76.81 | Helped |
| Add margin loss to plain ER with BCE (ablation) | ER with BCE · C-VOC, C-COCO, and C-NUS | Macro-AUC: 79.17 -> 78.80; 72.68 -> 73.95; 69.31 -> 76.03 | Mixed |
| Add margin loss to reweighted loss (ablation) | Reweighted-loss method · C-VOC, C-COCO, and C-NUS | Macro-AUC: 83.04 -> 86.29; 74.86 -> 77.61; 76.81 -> 78.59 | Helped |
| Add WRU to reweighted loss and margin loss (ablation) | RLDAM without WRU · C-VOC, C-COCO, and C-NUS | Macro-AUC: 86.29 -> 88.69; 77.61 -> 77.93; 78.59 -> 79.77 | Helped |
| Add WRU to reweighted loss without margin loss (ablation) | Reweighted-loss method · C-VOC, C-COCO, and C-NUS | Macro-AUC: 83.04 -> 84.92; 74.86 -> 75.36; 76.81 -> 77.60 | Helped |
Surprising findings
- Finetune with RLDAM loss alone performs comparably to ER on C-COCO and C-NUS-WIDE.
- Ours has higher forgetting than some baselines (e.g., OCDM -0.19 and PRS 4.93 on C-COCO) despite best Macro-AUC.
Limitations
- Forgetting is not a primary focus and the method is not best on forgetting for all benchmarks.
- Theoretical analysis uses kernel/NTK-based hypothesis class while experiments use deep ResNets.
Experimental setup
Datasets: PASCAL VOC MSCOCO NUS-WIDE
Benchmarks: C-PASCAL-VOC C-MSCOCO C-NUS-WIDE
Metrics: Macro-AUC Forgetting
Baselines: FT FT-RLDAM ER ER-RS PRS OCDM KRT
Builds on: reweighted univariate loss label-distribution-aware margin loss experience replay reservoir sampling
Closest papers by meaning
- Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.90)
- Cut out and Replay: A Simple yet Versatile Strategy for Multi-Label Online Continual Learning (ICML 2025 · similarity 0.87)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.86)
- Convergence and Implicit Bias of Gradient Descent on Continual Linear Classification (ICLR 2025 · similarity 0.86)
- L3A: Label-Augmented Analytic Adaptation for Multi-Label Class Incremental Learning (ICML 2025 · similarity 0.85)
- Continual Learning on a Diet: Learning from Sparsely Labeled Streams Under Constrained Computation (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.