Research map / Continual learning / Class-incremental
Adaptive Decision Boundary for Few-Shot 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 ADBS, a plug-and-play per-class adaptive decision boundary plus inter-class constraint loss that improves existing few-shot class-incremental learning methods on CIFAR100, miniImageNet and CUB200.
Problem
Few-shot class-incremental learning suffers from decision-space conflicts between old and novel classes when the backbone is frozen, since fixed or unified decision boundaries do not reserve adequate space for new classes.
Approach
Assign each class its own learnable decision boundary multiplier that is adapted during base and incremental sessions (new-class boundaries fine-tuned while old ones stay fixed), plus an inter-class constraint loss derived from a probability-based proposition that pushes apart class prototypes and boundaries.
Key results
- SAVC+ADBS reaches 71.03% average accuracy on CIFAR100 versus 63.07% for reproduced SAVC (+7.96).
- On miniImageNet, SAVC+ADBS achieves 72.36% average accuracy, +6.07 over reproduced SAVC.
- Baseline+ADBS improves average accuracy on CIFAR100 from 58.71% to 63.77% (+5.06) and on CUB200 from 62.34% to 67.12% (+4.78).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ADBS to the FSCIL baseline | FSCIL baseline · CIFAR100 | average accuracy: 58.71 -> 63.77 (+5.06) | Helped |
| Add ADBS to SAVC | SAVC · CIFAR100 | average accuracy: 63.07 -> 71.03 (+7.96) | Helped |
| Add ADBS to the FSCIL baseline | FSCIL baseline · miniImageNet | average accuracy: 60.25 -> 66.02 (+5.77) | Helped |
| Add ADBS to SAVC | SAVC · miniImageNet | average accuracy: 66.29 -> 72.36 (+6.07) | Helped |
| Add ADBS to SAVC | SAVC · CUB200 | average accuracy: 68.95 -> 69.87 (+0.93) | Helped |
| Add ADBS to ALICE | ALICE · CIFAR100 | average accuracy: 63.32 -> 65.63 (+2.31); session 0: 80.37 -> 80.12 (-0.25) | Mixed |
| Add ADBS to OrCo | OrCo · CIFAR100 | average accuracy: 60.19 -> 60.32 (+0.13); session 2: 62.39 -> 61.89 (-0.50) | Mixed |
| Add adaptive decision boundary without inter-class constraint (ablation) | FSCIL baseline · miniImageNet | session 8 accuracy: 50.81 -> 54.67 | Helped |
| Add inter-class constraint to adaptive decision boundary (ablation) | FSCIL baseline with ADB · miniImageNet | session 8 accuracy: 54.67 -> 55.38 | Helped |
| Add adaptive decision boundary without inter-class constraint (ablation) | FSCIL baseline · CIFAR100 | session 8 accuracy: 48.60 -> 51.09 | Helped |
| Add inter-class constraint to adaptive decision boundary (ablation) | FSCIL baseline with ADB · CIFAR100 | session 8 accuracy: 51.09 -> 51.15 | Helped |
| Add adaptive decision boundary without inter-class constraint (ablation) | FSCIL baseline · CUB200 | session 10 accuracy: 56.26 -> 58.94 | Helped |
Surprising findings
- Integrating ADBS with OrCo yields only +0.13 average gain on CIFAR100, and ALICE+ADBS gains are marginal on CUB200 (+0.13).
- ADBS sometimes slightly lowers base-session accuracy for OrCo on miniImageNet (-0.32 for ALICE) rather than improving it uniformly.
Limitations
- Gains are marginal for methods that already optimize boundaries (ALICE, OrCo).
- Proposition relies on the assumption f(x) approximately equals the class prototype, which is a heuristic.
- Requires fine-tuning boundaries for up to 10 epochs per incremental session, adding some training cost.
Experimental setup
Datasets: CIFAR100 miniImageNet CUB200
Benchmarks: CIFAR100 miniImageNet CUB200
Metrics: top-1 accuracy per session average accuracy relative improvement on last session class separation degree
Baselines: TOPIC CEC FACT C-FSCIL CLOM DBONet MCNet SoftNet WaRP TEEN NC-FSCIL ALFSCIL
Builds on: TOPIC SAVC ALICE OrCo BPNF
Closest papers by meaning
- Few-Shot Incremental Learning With Continually Evolved Classifiers (CVPR 2021 · similarity 0.93)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.90)
- Diffusion-Classifier Synergy: Reward-Aligned Learning via Mutual Boosting Loop for FSCIL (NeurIPS 2025 · similarity 0.87)
- Pseudo Informative Episode Construction for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.87)
- M2SD:Multiple Mixing Self-Distillation for Few-Shot Class-Incremental Learning (AAAI 2024 · similarity 0.86)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.86)
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.