Research map / Continual learning / Online and replay-based
CRAM: Large Scale Video Continual Learning with Bootstrapped Compression
Shivani Mall, Joao F. Henriques
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 |
|---|---|---|---|
| replace code refresh with drift-free (no refresh) baseline, i.e. DF vs CRAM uses 13779 Mb vs 1404 Mb memory (ablation) | CRAM · Kinetics-700 (incremental setting) | eval accuracy (%): 37.5 vs 38.80 | No clear effect |
| add code refreshing on top of compressed vision (CV) (ablation) | CRAM · Kinetics-700 (incremental setting) | eval accuracy (%): 26.5 -> 38.80 | Improved |
| add code refreshing on top of compressed vision (CV) (ablation) | CRAM · Kinetics-700 (incremental setting) | average forgetting (%): 60.4 -> 15.20 | Improved |
| add rehearsal buffer of compressed codes vs naive SGD (ablation) | CRAM · Kinetics-700 (incremental setting) | eval accuracy (%): 10.1 -> 38.80 | Improved |
| use 50x compression instead of 256x (ablation) | CRAM · Kinetics-700 (incremental setting) | eval accuracy (%): 38.80 -> 40.1 | Improved |
| use 50x compression instead of 256x (ablation) | CRAM · Kinetics-700 (pretraining setting) | eval accuracy (%): 46.50 -> 47.4 | Improved |
| train 40 epochs per task instead of 30 (ablation) | CRAM · Kinetics-700 (incremental setting) | eval accuracy (%): 38.80 -> 41.9 | Improved |
| train 40 epochs per task instead of 30 (ablation) | CRAM · EpicKitchens-100 (incremental setting) | eval accuracy (%): 28.10 -> 31.1 | Improved |
| use 15 classes per task for 45 tasks instead of 20 per task (ablation) | CRAM · Kinetics-700 (incremental setting) | eval accuracy (%): 38.80 -> 36.1 | Hurt |
| use 20 classes per task for 15 tasks instead of 30 per task (ablation) | CRAM · Kinetics-700 (pretraining setting) | eval accuracy (%): 46.50 -> 47.0 | No clear effect |
| replace REMIND with CRAM | REMIND · Kinetics-700 (pretraining setting) | eval accuracy (%): 38.20 -> 46.50 | Improved |
| replace SMILE with CRAM | SMILE · Kinetics-700 (incremental setting) | eval accuracy (%): 29.20 -> 38.80 | Improved |
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.
- GDumb — GDumb: A Simple Approach that Questions Our Progress in Continual Learning (ECCV 2020) · better on 3 of 3 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
- Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 2025 · similarity 0.88)
- Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling (ICLR 2025 · similarity 0.88)
- Towards Backward-Compatible Continual Learning of Image Compression (CVPR 2024 · similarity 0.87)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.87)
- FreshMem: Brain-Inspired Frequency-Space Hybrid Memory for Streaming Video Understanding (NeurIPS 2026 · similarity 0.87)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (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.
