Research map / Continual learning / Class-incremental
L3A: Label-Augmented Analytic Adaptation for Multi-Label Class Incremental 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 L3A, an exemplar-free multi-label class-incremental learning method that combines confidence-thresholded pseudo-labels for missing historical labels with a frequency-weighted recursive analytic (ridge regression) classifier on a frozen TResNet-M backbone.
Problem
Multi-label class-incremental learning (MLCIL), where each phase only annotates current classes, causing label absence and class imbalance, under exemplar-free (privacy) constraints.
Approach
Generate pseudo-labels for previously learned classes using the old classifier to complete label sets, and solve a recursive closed-form ridge regression classifier with sample-specific weights derived from inverse-square-root class frequencies to counter imbalance.
Key results
- On MS-COCO B0-C10, L3A reaches 77.6% last mAP and 81.5% average mAP, exceeding the best exemplar-free method CSC by 4.8 points last mAP.
- On PASCAL VOC B0-C4, L3A achieves 94.1% last mAP, surpassing MULTI-LANE by 5.3 points.
- Ablation on MS-COCO B0-C10 shows the baseline at 25.6 last mAP, w/o WAC at 56.7, w/o PL at 77.3, and full L3A at 77.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add WAC (ablation) | L3A · COCO B0-C10 | Last mAP: 56.7 -> 77.6 | Helped |
| add WAC (ablation) | L3A · COCO B40-C10 | Last mAP: 72.1 -> 78.8 | Helped |
| add PL (ablation) | L3A · COCO B0-C10 | Last mAP: 77.3 -> 77.6 | Helped |
| add PL (ablation) | L3A · COCO B40-C10 | Last mAP: 78.4 -> 78.8 | Helped |
| increase gamma from 0.1 to 1000 (ablation) | L3A · COCO B0-C10 | Last mAP: 76.83 -> 77.56 | Helped |
| increase buffer layer size from 512 to 8192 (ablation) | L3A · COCO B0-C10 | Last mAP: 75.24 -> 77.56 | Helped |
| use dynamic rather than fixed confidence threshold (ablation) | L3A · COCO B0-C10 | Last mAP: 77.57 -> 76.56 | Hurt |
| replace 1/sqrt(frequency) with 1/frequency weighting (ablation) | L3A WAC · COCO B0-C10 | Last mAP: 77.57 -> 77.30 | Hurt |
| replace 1/sqrt(frequency) with 1/(log(frequency)+1) weighting (ablation) | L3A WAC · COCO B0-C10 | Last mAP: 77.57 -> 77.41 | Hurt |
| use L3A instead of CSC-R | MLCIL · MS-COCO B0-C10 | Last mAP: 74.8 -> 77.6 | Helped |
| use L3A instead of CSC-R | MLCIL · MS-COCO B40-C10 | Last mAP: 76.7 -> 78.8 | Helped |
| use L3A instead of MULTI-LANE | MLCIL · VOC B0-C4 | Last mAP: 88.8 -> 94.1 | Helped |
Surprising findings
- Dynamic confidence thresholding for pseudo-labels gave no notable gain over a fixed threshold of 0.7.
Limitations
- Relies on a fixed confidence threshold (0.7) for pseudo-labels, with dynamic thresholding not improving results.
- Buffer layer size and regularization term require hyperparameter selection (gamma=1000, buffer 8192).
Experimental setup
Datasets: MS-COCO 2014 PASCAL VOC 2007
Benchmarks: MS-COCO B0-C10 MS-COCO B40-C10 PASCAL VOC B0-C4 PASCAL VOC B10-C2
Metrics: mAP average mAP last mAP CF1 OF1
Baselines: FT iCaRL BiC ER TPCIL PODNet DER++ KRT-R CSC-R PRS OCDM KRT
Builds on: Analytic Continual Learning (ACIL) recursive least squares MULTI-LANE CSC
Closest papers by meaning
- Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.89)
- Specifying What You Know or Not for Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.88)
- Generative Multi-modal Models are Good Class Incremental Learners (CVPR 2024 · similarity 0.86)
- Towards Macro-AUC Oriented Imbalanced Multi-Label Continual Learning (AAAI 2025 · similarity 0.85)
- AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection (NeurIPS 2025 · similarity 0.84)
- Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning (CVPR 2025 · similarity 0.84)
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.