Research map / Continual learning / Class-incremental
Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement
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 NO-CL, a non-exemplar online class-incremental setting, and a dual-prototype method (vanilla and hyperdimensional prototypes with online refinement) that preserves base-class discernibility without storing data examples.
Problem
Online class-incremental learning where base-class knowledge must be retained without any stored exemplars, while novel classes are learned from a single-pass data stream.
Approach
Pre-train on base classes with dual projection heads producing vanilla and high-dimensional (hyperdimensional) prototypes, then during online sessions sample prototypes from calibrated Gaussians and refine them while finetuning only the projection module, avoiding exemplar buffers.
Key results
- On CIFAR-100 with 40%+6%x10 partition, the method reaches 38.6 overall accuracy and 39.0 HM, versus 37.2 HM for SCR (MS 1000 setting) as reported in the appendix tables.
- On the 80%+2%x10 partition, the method achieves 55.2 accuracy on CORE-50 and 54.6 HM, outperforming the best compared method (SCR_ft at 48.0 accuracy with HM 43.9 for DVC_ft).
- Online training time on CIFAR-100 is 35 s versus 512 s for ALICE and 561 s for OCM, with 1.4 GB memory versus 12.8 GB for OCM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed dual-prototype self-augmentation and refinement method instead of the strongest reported baseline | NO-CL · CORE-50, 40% + 6% × 10 | HM: 39.7 -> 45.3 | Helped |
| use the proposed method instead of the strongest reported baseline | NO-CL · CIFAR100, 40% + 6% × 10 | HM: 36.8 -> 39.0 | Helped |
| use the proposed method instead of the strongest reported baseline | NO-CL · Mini-ImageNet, 40% + 6% × 10 | HM: 36.6 -> 39.3 | Helped |
| use the proposed method instead of the strongest reported baseline | NO-CL · CIFAR100, 80% + 2% × 10 | HM: 44.9 -> 51.1 | Helped |
| use the proposed method instead of the strongest reported baseline | NO-CL · Mini-ImageNet, 80% + 2% × 10 | HM: 45.5 -> 51.7 | Helped |
| add self-supervised learning loss to base-session training (ablation) | Ours with supervised contrastive training · CIFAR100 | accuracy: 48.6 -> 50.4 | Helped |
| add self-supervised learning loss to base-session training (ablation) | Ours with supervised contrastive training · Mini-ImageNet | accuracy: 50.7 -> 52.2 | Helped |
| add data augmentation to base-session training (ablation) | Ours with supervised contrastive training · CIFAR100 | accuracy: 48.6 -> 51.2 | Helped |
| add data augmentation to base-session training (ablation) | Ours with supervised contrastive training · Mini-ImageNet | accuracy: 50.7 -> 53.2 | Helped |
| increase embedding dimension from 256 to 2048 (ablation) | Ours · CIFAR100 | HM: 42.9 -> 47.2 | Helped |
| increase embedding dimension from 2048 to 3074 (ablation) | Ours · CIFAR100 | HM: 47.2 -> 47.0 | No clear effect |
| reduce the number of base classes to 40% | NO-CL · CORE-50, 40% + 6% × 10 | accuracy | Hurt |
Surprising findings
- Example-based OCL methods with finetuning (e.g., SCR_ft) often do not dominate the non-exemplar method, and PASS retains almost no novel-class accuracy (below 1%) in the reported partitions.
Limitations
- Performance depends strongly on the quality of the base-session pretrained backbone; fewer base classes lead to poor performance on both base and novel classes.
Experimental setup
Datasets: CORE-50 CIFAR-100 Mini-ImageNet
Benchmarks: CORE-50 CIFAR-100 Mini-ImageNet
Metrics: accuracy (base/novel) harmonic accuracy (HM) training time memory footprint
Baselines: SCR OCM DVC PASS ALICE MIR GDumb ASER SSRE
Builds on: SCR PASS ALICE OCM
Closest papers by meaning
- Not Just Selection, but Exploration: Online Class-Incremental Continual Learning via Dual View Consistency (CVPR 2022 · similarity 0.91)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.88)
- Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization (CVPR 2025 · similarity 0.88)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · similarity 0.87)
- Autoencoder-Based Hybrid Replay for Class-Incremental Learning (ICML 2025 · similarity 0.87)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.87)
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.