Research map / Continual learning / Class-incremental
Adaptive Discovering and Merging for Incremental Novel Class Discovery
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 Adaptive Discovering and Merging (ADM) for class-incremental novel class discovery, combining decoupled triplet-comparison/probability-regularized novel discovery with Adaptive Model Merging that folds the novel branch into the base model without parameter growth.
Problem
Class-incremental novel class discovery must discover novel classes from unlabelled data across sequential tasks while avoiding catastrophic forgetting of previously learned classes, without access to old data.
Approach
Decouples representation learning from classifier-based novel discovery, uses triplet comparison and probability regularization for adaptive category assignment, and merges a gated novel branch into the base model via base BN scaling factors (AMM), giving lossless fusion with no parameter growth.
Key results
- ADM achieves average All accuracy 61.3% across CIFAR-10, CIFAR-100 and Tiny-ImageNet in the one-step class-iNCD setting, versus 58.3% for FRoST.
- FRoST+AMM reaches 60.2% average All accuracy, with AMM matching or exceeding AFF overall without adding parameters.
- Applying AMM to DyTox+ improves CIFAR-100 accuracy from 64.01 to 65.67 (25 stages) and 69.28 to 70.52 (10 stages) with no added parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AMM to FRoST | FRoST · CIFAR-10 | all accuracy: 63.4 -> 64.2 | Helped |
| Add AMM to FRoST | FRoST · CIFAR-100 | all accuracy: 59.2 -> 61.1 | Helped |
| Add AMM to FRoST | FRoST · Tiny-ImageNet | all accuracy: 52.3 -> 54.7 | Helped |
| Combine adaptive novel class discovery with AMM | FRoST+AMM · CIFAR-10 | all accuracy: 64.2 -> 65.2 | Helped |
| Combine adaptive novel class discovery with AMM | FRoST+AMM · CIFAR-100 | all accuracy: 61.1 -> 63.2 | Helped |
| Combine adaptive novel class discovery with AMM | FRoST+AMM · Tiny-ImageNet | all accuracy: 54.7 -> 55.6 | Helped |
| Add AMM | iCaRL · CIFAR-100, 25 stages | performance: 50.60 -> 60.59 | Helped |
| Add AMM | iCaRL · CIFAR-100, 10 stages | performance: 53.78 -> 62.20 | Helped |
| Add AMM | PODNet · CIFAR-100, 25 stages | performance: 60.72 -> 63.45 | Helped |
| Add AMM | PODNet · CIFAR-100, 10 stages | performance: 63.19 -> 64.32 | Helped |
| Add AMM | NECIL · CIFAR-100, 25 stages | performance: 61.01 -> 61.49 | Helped |
| Add AMM | NECIL · CIFAR-100, 10 stages | performance: 61.07 -> 61.83 | Helped |
Surprising findings
- Adding AMM to the class-IL methods improves performance more as the number of incremental stages grows.
- Some NCD baselines such as NCL and ResTune retain high old-class accuracy but nearly zero new-class accuracy, indicating they fail to balance old and new classes.
Limitations
- Gains on new-class accuracy are modest; FRoST+AMM and ADM sometimes trail baselines on specific New/Old columns.
- Experiments are limited to ResNet-18 and small image benchmarks (CIFAR, Tiny-ImageNet).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet
Metrics: Old accuracy New accuracy All accuracy (with Hungarian re-assignment)
Baselines: AutoNovel ResTune NCL DTC FRoST iCaRL PODNet NECIL AFC DER FOSTER DyTox
Builds on: FRoST Incremental Model Merging AutoNovel DyTox
Closest papers by meaning
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.88)
- Continual Generalized Category Discovery: Learning and Forgetting from a Bayesian Perspective (ICML 2025 · similarity 0.88)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.87)
- Happy: A Debiased Learning Framework for Continual Generalized Category Discovery (NeurIPS 2024 · similarity 0.87)
- Task Confusion and Catastrophic Forgetting in Class-Incremental Learning: A Mathematical Framework for Discriminative and Generative Modelings (NeurIPS 2024 · similarity 0.87)
- Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning (CVPR 2025 · 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.