Research map / Continual learning / Class-incremental
PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning
Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, Eduardo Valle
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 POD-spatial on intermediate features (on top of POD-flat) (ablation) | PODNet LSC · CIFAR100 50 steps (NME) | average incremental accuracy: 53.29 -> 61.40 | Improved |
| add POD-flat on the final embedding (ablation) | LSC classifier · CIFAR100 50 steps (NME) | average incremental accuracy: 41.56 -> 53.29 | Improved |
| replace cosine classifier with LSC (ablation) | PODNet with POD-flat and POD-spatial · CIFAR100 50 steps (NME) | average incremental accuracy: 56.69 -> 61.40 | Improved |
| replace cosine classifier with LSC (ablation) | PODNet with POD-flat and POD-spatial · CIFAR100 50 steps (CNN) | average incremental accuracy: 55.72 -> 57.98 | Improved |
| use NCA-based loss instead of cross-entropy for LSC (ablation) | PODNet · CIFAR100 50 steps (NME) | average incremental accuracy: 59.86 -> 61.40 | Improved |
| use POD-pixels instead of no intermediate distillation (ablation) | LSC with POD-flat · CIFAR100 50 steps (NME) | average incremental accuracy: 53.29 -> 49.74 | Hurt |
| use POD-spatial instead of POD-width alone (ablation) | LSC with POD-flat · CIFAR100 50 steps (NME) | average incremental accuracy: 60.92 -> 61.40 | Improved |
| use Perceptual Style distillation instead of none (ablation) | LSC with POD-flat · CIFAR100 50 steps (NME) | average incremental accuracy: 53.29 -> 51.01 | Hurt |
| use GradCam distillation instead of none (ablation) | LSC with POD-flat · CIFAR100 50 steps (CNN) | average incremental accuracy: 52.98 -> 52.48 | No clear effect |
| add POD-flat when POD-spatial is already used (ablation) | LSC with POD-spatial · CIFAR100 50 steps (NME) | average incremental accuracy: 61.42 -> 61.40 | No clear effect |
| use 10 proxies per class instead of 1 (ablation) | LSC · CIFAR100 50 steps (CNN) | average incremental accuracy: 56.80 -> 57.98 | Improved |
| add POD-spatial on top of POD-flat (ablation) | PODNet · CIFAR100 old vs new classes | accuracy: +7 p.p. old classes; -0.2 p.p. new classes | Mixed |
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 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
- A continual learning survey: Defying forgetting in classification tasks (TPAMI 2021 · similarity 0.89)
- MIND: Multi-Task Incremental Network Distillation (AAAI 2024 · similarity 0.88)
- Class Incremental Learning with Multi-Teacher Distillation (CVPR 2024 · similarity 0.87)
- Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective (ICML 2025 · similarity 0.87)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.86)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (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.
