Research map / Continual learning / Class-incremental
Resurrecting Old Classes with New Data for Exemplar-Free Continual 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 Adversarial Drift Compensation (ADC), which uses targeted adversarial perturbations of new-task images to act as pseudo-exemplars for estimating and compensating drift of old class prototypes in exemplar-free class-incremental learning under small-start settings.
Problem
Exemplar-free class-incremental learning degrades sharply when the first task is small, because feature drift of stored old-class prototypes is hard to estimate without exemplars.
Approach
Adversarially perturb current-task samples so that, in the old feature extractor, they embed close to old class prototypes; use these pseudo-exemplars to measure the old-to-new embedding drift and shift the prototypes accordingly, exploiting the transferability of such adversarial samples to the new model.
Key results
- On CIFAR-100 with 5 tasks, ADC reaches 59.14% last-task accuracy, 4.2% above the second-best method SDC.
- On CUB-200 5-task, ADC achieves 64.46% last-task accuracy, 5.78% above the runner-up.
- ADC adds about 2.5-3.1% computation overhead over normal training on CIFAR-100 and ImageNet-Subset settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ADC rather than SDC for prototype drift compensation | ResNet18 exemplar-free continual learning · CIFAR-100, T=5 | Alast: 54.94 -> 59.14 | Helped |
| Use ADC rather than SDC for prototype drift compensation | ResNet18 exemplar-free continual learning · CIFAR-100, T=10 | Alast: 41.36 -> 46.48 | Helped |
| Use ADC rather than SDC for prototype drift compensation | ResNet18 exemplar-free continual learning · TinyImageNet, T=5 | Alast: 40.05 -> 41.0 | Helped |
| Use ADC rather than SDC for prototype drift compensation | ResNet18 exemplar-free continual learning · TinyImageNet, T=10 | Alast: 27.42 -> 32.32 | Helped |
| Use ADC rather than SDC for prototype drift compensation | ResNet18 exemplar-free continual learning · ImageNet-Subset, T=5 | Alast: 59.82 -> 62.40 | Helped |
| Use ADC rather than SDC for prototype drift compensation | ResNet18 exemplar-free continual learning · ImageNet-Subset, T=10 | Alast: 43.72 -> 47.58 | Helped |
| Use ADC rather than LwF for prototype drift compensation | ResNet18 exemplar-free continual learning · CUB-200, T=5 | Alast: 58.68 -> 64.46 | Helped |
| Use ADC rather than FeCAM for prototype drift compensation | ResNet18 exemplar-free continual learning · CUB-200, T=10 | Alast: 51.78 -> 57.97 | Helped |
| Use ADC rather than LwF for prototype drift compensation | ResNet18 exemplar-free continual learning · Stanford Cars, T=7 | Alast: 45.18 -> 54.86 | Helped |
| Use ADC rather than FeCAM for prototype drift compensation | ResNet18 exemplar-free continual learning · Stanford Cars, T=14 | Alast: 37.50 -> 45.07 | Helped |
| Increase perturbation iterations from 1 to 3 (ablation) | ADC · CIFAR-100, T=10 | Alast: 45.96 -> 46.48 | Helped |
| Increase perturbation iterations from 3 to 10 (ablation) | ADC · CIFAR-100, T=10 | Alast: 46.48 -> 45.28 | Hurt |
Surprising findings
- A simple LwF-plus-NCM baseline outperforms several dedicated exemplar-free methods (SSRE, PASS, FeTrIL, FeCAM) in small-start settings.
- Adversarial samples crafted on the old feature extractor remain effective (misclassified as target class) on the new extractor, i.e., continual adversarial transferability.
Limitations
- Requires task boundaries to trigger drift compensation and access to enough current-task data, so it is not directly applicable to online CL.
- Stores all class prototypes, as most existing methods do.
- Not evaluated in continual few-shot or online settings.
Experimental setup
Datasets: CIFAR-100 TinyImageNet ImageNet-Subset CUB-200 Stanford Cars
Benchmarks: CIFAR-100 TinyImageNet ImageNet-Subset CUB-200 Stanford Cars
Metrics: last task accuracy average incremental accuracy
Baselines: LwF NCM SDC PASS SSRE FeTrIL FeCAM NME
Builds on: SDC LwF PyCIL targeted adversarial attacks
Closest papers by meaning
- LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning (CVPR 2025 · similarity 0.89)
- Model Inversion with Layer-Specific Modeling and Alignment for Data-Free Continual Learning (NeurIPS 2025 · similarity 0.87)
- Adaptive Retention & Correction: Test-Time Training for Continual Learning (ICLR 2025 · similarity 0.87)
- Doubly Perturbed Task Free Continual Learning (AAAI 2024 · similarity 0.87)
- Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning (ICLR 2024 · similarity 0.87)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 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.