Research map / Label-efficient and robust learning / Self-supervised and contrastive
Revisiting Supervision for Continual Representation Learning
Daniel Marczak, Sebastian Cygert, Tomasz Trzcinski, Bartlomiej Twardowski
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 MLP projector to supervised training (SL+MLP) (ablation) | supervised continual finetuning · CIFAR100/5 | k-NN accuracy: 45.3 -> 61.9 | Improved |
| add MLP projector to SL on ImageNet100/5 (ablation) | SL · ImageNet100/5 | k-NN accuracy: 40.4 -> 62.4 | Improved |
| SL+MLP versus SSL on CIFAR10/5 | SL+MLP · CIFAR10/5 | k-NN accuracy: 65.9 vs BarlowTwins 76.2 | Hurt |
| add MLP projector to SL for transfer learning (ablation) | SL · 8 downstream datasets avg | k-NN accuracy: 23.3 -> 41.9 | Improved |
| add PFR distillation to SL+MLP | SL+MLP · CIFAR10/5 | k-NN accuracy: 65.9 -> 76.3 | Improved |
| add LwF distillation to SL+MLP | SL+MLP · CIFAR100/5 | k-NN accuracy: 61.9 -> 58.7 | Hurt |
| add MLP projector to SL (ablation) | SL · C10 to C100 (CIFAR10 forgetting) | forgetting F: 12.4 -> 4.5 | Improved |
| reset the MLP projector at each new task instead of carrying it over (ablation) | SL+MLP · CIFAR100/5 | k-NN accuracy: 60.1 -> 61.9 | Improved |
| carry over the projector instead of resetting it for SSL (ablation) | SimCLR · CIFAR100/5 | k-NN accuracy: 47.4 -> 48.9 | Improved |
| increase projector depth and width beyond one block (ablation) | SL+MLP projector · CIFAR100/5 | k-NN accuracy | No clear effect |
| train with only 30% of data and 40% label noise (ablation) | SL+MLP · CIFAR100/5 | k-NN accuracy | No clear effect |
| pretrain SL on semantically distant SVHN before CIFAR10 | SL · SVHN to C10 | forward transfer FT: -0.4 (SL) vs -0.1 (SL+MLP) | Hurt |
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.
- BarlowTwins — Barlow Twins: Self-Supervised Learning via Redundancy Reduction (ICML 2021) · better on 3 of 4 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
- Self-Supervised Models Are Continual Learners (CVPR 2022 · similarity 0.91)
- Probing Representation Forgetting in Supervised and Unsupervised Continual Learning (CVPR 2022 · similarity 0.91)
- Representational Continuity for Unsupervised Continual Learning (ICLR 2022 · similarity 0.91)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.89)
- Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective (ICML 2025 · similarity 0.88)
- Continual Learning with Global Alignment (NeurIPS 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.
