Research map / Continual learning / Online and replay-based
Instance and Category Supervision are Alternate Learners for Continual Learning
Xudong Tian, Zhizhong Zhang, Xin Tan, Jun Liu, Chengjie Wang, Yanyun Qu, Guannan Jiang, Yuan Xie
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 a supervised branch to the self-supervised learner without alternation (ablation) | SSL continual learner · TinyImageNet 20-split (task-incremental) | avg. forgetting: 3.54 -> 5.29 | Hurt |
| add alternate learner inner loop to SSL+SL (ablation) | SSL+SL · CIFAR-100 10-split (task-incremental) | avg. accuracy: 69.1 -> 75.8 | Improved |
| apply outer-loop momentum update only to SSL (ablation) | SSL · CIFAR-100 10-split (task-incremental) | avg. accuracy: 68.4 -> 68.9 | No clear effect |
| combine SSL and SL with DualNet-style rigid combination (ablation) | SSL+SL · TinyImageNet 20-split (task-incremental) | avg. accuracy: 77.52 -> 78.41 | No clear effect |
| use full alternate learner (inner + outer loop) instead of ALin (ablation) | SSL+SL+AL · CIFAR-100 10-split (task-incremental) | avg. accuracy: 75.8 -> 77.5 | Improved |
| add the proposed disentangled SSL losses (Eq. 4 and Eq. 5) (ablation) | alternate learner · CIFAR-100 10-split (task-incremental) | avg. accuracy: 77.5 -> 80.0 | Improved |
| add the proposed SSL losses to SL on TinyImageNet (ablation) | alternate learner · TinyImageNet 20-split (task-incremental) | avg. accuracy: 81.43 -> 82.81 | Improved |
| use too few SSL iterations (e_SSL = 5) (ablation) | alternate learner · CIFAR-100 10-split | avg. accuracy: below the 68.4 baseline | Hurt |
| use too many SSL iterations (ablation) | alternate learner · CIFAR-100 10-split | avg. accuracy | Hurt |
| use larger embedding dimension without the proposed SSL losses (ablation) | baseline SSL · CIFAR-100 10-split | avg. accuracy | Improved |
| use proposed SSL losses with compact 128-d embedding rather than 2048-d (ablation) | alternate learner · CIFAR-100 10-split | avg. accuracy | 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.
- FOSTER — FOSTER: Feature Boosting and Compression for Class-Incremental Learning (ECCV 2022) · better on 1 of 1 shared result
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
- Revisiting Supervision for Continual Representation Learning (ECCV 2024 · similarity 0.90)
- Self-Supervised Models Are Continual Learners (CVPR 2022 · similarity 0.90)
- Probing Representation Forgetting in Supervised and Unsupervised Continual Learning (CVPR 2022 · similarity 0.90)
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.88)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.88)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 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.
