Research map / Continual learning / Online and replay-based
Anytime Continual Learning for Open Vocabulary Classification
Zhen Zhu, Yiming Gong, Derek Hoiem
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 CLS-attention weighting to per-instance PCA (ablation) | feature compression · CIFAR100 | FT accuracy: 74.6 -> 77.9 | Improved |
| use per-instance PCA instead of dataset-wide PCA (ablation) | feature compression · CIFAR100 | FT accuracy and KB/example: 76.4 at 40.2 KB -> 77.9 at 19.4 KB (with CLS-weight) | Improved |
| compress features with VQ instead of PCA (ablation) | feature compression · CIFAR100 | FT accuracy: 77.8 -> 64.8 | Hurt |
| add int-quantization to PCA compression (ablation) | feature compression · CIFAR100 | FT accuracy: 77.9 -> 77.5 and 19.4 -> 5.3 KB/example | No clear effect |
| use compressed features instead of full features for training (ablation) | partial fine-tuning · CIFAR100 | ms/batch: 25.6 -> 13.9 ms/batch; 153.6 -> 5.3 KB | Improved |
| replace AIM weighting with Online Class-wise Weighting (ablation) | AnytimeCL · data and class incremental, flexible inference | accuracy | Improved |
| tune only the last transformer block with fixed label embedding (ablation) | AnytimeCL · flexible inference | accuracy on novel labels | Improved |
| class-balanced vs uniform sampling of stored samples (ablation) | AnytimeCL · union data incremental | accuracy | No clear effect |
| add the 'other' logit regularization (ablation) | AnytimeCL · union data incremental | accuracy | Improved |
| use DINOv2 as the tuned model instead of CLIP (ablation) | AnytimeCL · data incremental | accuracy | Improved |
| increase online batch size above 32 (ablation) | AnytimeCL · union data incremental | accuracy | Mixed |
| use AnytimeCL instead of LinProbe (AIM) on MTIL | task incremental · MTIL Order I | Avg.: 77.1 -> 77.0 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- iCaRL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- OVS Meets Continual Learning: Towards Sustainable Open-Vocabulary Segmentation (NeurIPS 2025 · similarity 0.89)
- RATT: Recurrent Attention to Transient Tasks for Continual Image Captioning (NeurIPS 2020 · similarity 0.88)
- Continual Learning on a Diet: Learning from Sparsely Labeled Streams Under Constrained Computation (ICLR 2024 · similarity 0.88)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.88)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.87)
- Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters (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.
