Research map / Continual learning / Online and replay-based
Dual-Enhanced Coreset Selection with Class-wise Collaboration for Online Blurry 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 DECO, a class-wise balanced memory plus class-wise gradient-similarity coreset selection with diverse score guidance and adaptive score constraint, for online blurry class incremental learning.
Problem
Online blurry class incremental learning (OBCIL), where tasks share classes and major classes shift across tasks, causing class/data imbalance that complicates coreset selection for replay.
Approach
Establishes a real-time class-wise balanced memory (CBM) and a class-wise gradient-based scoring system (CW_S, CW_V, CW_A) whose score guidance differs by class group (past major, current major, future major), with an adaptive split-gradient affinity constraint that relaxes cross-class constraints to boost plasticity.
Key results
- DECO reaches 60.03% FAA on CIFAR10 (|M|=500) vs 58.14% for RM, the strongest competitor, and 90.89% on MNIST.
- On ImageNet (|M|=20000) DECO achieves 42.58% FAA vs 41.29% for RM, with FF 1.03 vs 1.41.
- DECO attains negative final forgetting on CIFAR10 (-1.96) and CIFAR100 is 33.93% FAA vs 32.89% for RM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add class-wise balanced memory (CBM) (ablation) | ER · CIFAR10 | FAA: 52.60 ±2.43 -> 56.19 ±1.21 | Helped |
| add diverse score guidance (DSG) (ablation) | CBM · CIFAR10 | FAA: 56.19 ±1.21 -> 57.36 ±0.92 | Helped |
| add adaptive score constraint (ASC) (ablation) | CBM + DSG · CIFAR10 | FAA: 57.36 ±0.92 -> 60.03 ±1.37 | Helped |
| use DECO instead of RM | online blurry class incremental learning · CIFAR10 | FAA: 58.14 ±1.96 -> 60.03 ±1.37 | Helped |
| increase buffer size from 200 to 500 (ablation) | DECO · CIFAR10 | FAA: 45.61 -> 60.03 | Helped |
| increase buffer size from 500 to 1000 (ablation) | DECO · CIFAR10 | FAA: 60.03 -> 69.07 | Helped |
| increase blurry extent from Blurry10 to Blurry20 (ablation) | DECO · CIFAR10 | FAA: 60.03 -> 60.13 | Helped |
| increase blurry extent from Blurry20 to Blurry30 (ablation) | DECO · CIFAR10 | FAA: 60.13 -> 60.66 | Helped |
Surprising findings
- DECO achieves negative final forgetting on CIFAR10, indicating later accuracy on major classes exceeds their earlier best.
Limitations
- RM is included as a comparison although it reuses all stream data and is impractical in the online setting.
- Results rely on the two-stage training paradigm following prior work, not evaluated with alternative training schemes.
Experimental setup
Datasets: MNIST CIFAR10 CIFAR100 ImageNet
Benchmarks: MNIST CIFAR10 CIFAR100 ImageNet
Metrics: Final Average Accuracy Final Forgetting
Baselines: ICARL ER Gdumb OCS RM MIR SCR DVC BIC ER-ACE
Builds on: OCS RM (online blurry CL) experience replay reservoir sampling
Closest papers by meaning
- OrCo: Towards Better Generalization via Orthogonality and Contrast for Few-Shot Class-Incremental Learning (CVPR 2024 · similarity 0.81)
- Attraction Diminishing and Distributing for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.80)
- DS-AL: A Dual-Stream Analytic Learning for Exemplar-Free Class-Incremental Learning (AAAI 2024 · similarity 0.80)
- Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.79)
- CoMBO: Conflict Mitigation via Branched Optimization for Class Incremental Segmentation (CVPR 2025 · similarity 0.79)
- Tensor Decomposition Based Memory-Efficient Incremental Learning (ICML 2025 · similarity 0.79)
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.