Research map / Continual learning / Class-incremental
Exemplar-free Continual Representation Learning via Learnable Drift Compensation
Alex Gomez-Villa, Dipam Goswami, Kai Wang, Andy Bagdanov, Bartlomiej Twardowski, Joost van de Weijer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add learnable drift compensation (LDC) to LwF+NCM prototypes | LwF with NCM classifier · CIFAR-100 (10 tasks, cold start) | last-task accuracy: 40.5 -> 45.4 | Improved |
| replace SDC with LDC for prototype correction | LwF · ImageNet100 (10 tasks, cold start) | last-task accuracy: 42.6 -> 51.4 | Improved |
| add SDC prototype correction to LwF+NCM in cold start | LwF with NCM classifier · CIFAR-100 (10 tasks, cold start) | last-task accuracy: 40.5 -> 40.6 | No clear effect |
| add LDC to fine-tuning with pretrained ViT-B/16 | FT+NCM · Stanford Cars (10 tasks) | accuracy after task 10: 53.4 -> 62.9 | Improved |
| add LDC to self-supervised CL methods for semi-supervised exemplar-free learning | CaSSLe · CIFAR-100 semi-supervised (25% labels) | last-task accuracy: 23.6 -> 35.0 | Improved |
| add LDC to PFR in semi-supervised setting with very few labels | PFR · CIFAR-100 semi-supervised (0.8% labels) | last-task accuracy: 20.5 (naive) -> 18.6 | Hurt |
| use a single linear projector instead of ReLU after the linear layer (ablation) | LwF+LDC · CIFAR-100 (10 tasks) | last-task accuracy: 41.2 -> 45.4 | Improved |
| use a single linear projector instead of a two-layer MLP (ablation) | CaSSLe+LDC · CIFAR-100 semi-supervised (25% labels) | last-task accuracy: 34.9 -> 35.0 | No clear effect |
| add bias to the linear projector (ablation) | LwF+LDC · CIFAR-100 (10 tasks) | last-task accuracy: 45.4 -> 45.2 | No clear effect |
| project stored per-class features instead of class means (ablation) | CaSSLe+LDC · CIFAR-100 semi-supervised (25% labels) | last-task accuracy: 35.0 -> 35.05 (N=500) | No clear effect |
| give NNCSL exemplars versus LDC without exemplars | exemplar-free LDC vs NNCSL · ImageNet100 semi-supervised (25% labels) | last-task accuracy: NNCSL (5120) 47.1 vs CaSSLe+LDC 35.8 | Hurt |
| replace exemplar-based NME with LDC for prototype correction (ablation) | LwF+NCM · CIFAR-100, Tiny-ImageNet, ImageNet100 | last-task accuracy | No clear effect |
Closest papers by meaning
- LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning (CVPR 2025 · similarity 0.91)
- Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (CVPR 2024 · similarity 0.90)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.88)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 2025 · similarity 0.87)
- AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection (NeurIPS 2025 · similarity 0.87)
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning (ICLR 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.
