Research map / Continual learning / Online and replay-based
The CLEAR Benchmark: Continual LEArning on Real-World Imagery
Zhiqiu Lin, Jia Shi, Deepak Pathak, Deva Ramanan
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 |
|---|---|---|---|
| Favor newer samples in the one-bucket replay buffer (dynamic alpha 0.75 * i/k rather than uniform alpha 1.0) (ablation) | Linear classifier trained from scratch on MoCo-YFCC-B0 features · CLEAR iid | In-Domain Accuracy: 88.4\% -> 89.0\% | Improved |
| Favor newer samples in the one-bucket replay buffer (dynamic alpha 0.75 * i/k rather than uniform alpha 1.0) (ablation) | Linear classifier trained from scratch on MoCo-YFCC-B0 features · CLEAR iid | Backward Transfer: 88.9\% -> 88.3\% | Hurt |
| Use finetuning rather than training the linear classifier from scratch (ablation) | Linear classifier on MoCo-YFCC-B0 features with one-bucket memory · CLEAR iid | Accuracy: 88.8\% -> 90.4\% | Improved |
| Increase dynamic alpha from 0.25 * i/k to 1.00 * i/k, retaining more recent samples (ablation) | Finetuned linear classifier on MoCo-YFCC-B0 features with one-bucket memory · CLEAR streaming | Next-Domain Accuracy: 89.7\% -> 91.1\% | Improved |
| Use cumulative memory rather than a one-bucket buffer (ablation) | Finetuned linear classifier on MoCo-YFCC-B0 features · CLEAR iid | Accuracy: 91.7\% -> 93.1\% | Improved |
| Use Pretrained-ImageNet rather than CLIP features (ablation) | Finetuned linear classifier with one-bucket memory and alpha 1.00 * i/k · CLEAR iid | In-Domain Accuracy: 82.9\% -> 94.8\% | Improved |
| Use CLIP features rather than Pretrained-ImageNet features (ablation) | Finetuned linear classifier with one-bucket memory and alpha 1.00 * i/k · CLEAR iid | In-Domain Accuracy: 94.8\% -> 82.9\% | Hurt |
| Train on CLEAR's temporally evolving real-world imagery rather than assuming an artificial benchmark captures real-world variation | Continual image classification benchmark design · CLEAR | test accuracy: null | Mixed |
Closest papers by meaning
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.90)
- Real-Time Evaluation in Online Continual Learning: A New Hope (CVPR 2023 · similarity 0.90)
- CTP:Towards Vision-Language Continual Pretraining via Compatible Momentum Contrast and Topology Preservation (ICCV 2023 · similarity 0.90)
- A Practitioner's Guide to Real-World Continual Multimodal Pretraining (NeurIPS 2024 · similarity 0.89)
- TiC-CLIP: Continual Training of CLIP Models (ICLR 2024 · similarity 0.89)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.88)
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.
