Research map / Continual learning / Class-incremental
First Session Adaptation: A Strong Replay-Free Baseline for Class-Incremental Learning
Aristeidis Panos, Yuriko Kobe, Daniel Olmeda Reino, Rahaf Aljundi, Richard E. Turner
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 FiLM adaptation in the first session (ablation) | FSA with EfficientNet-B0 · VTAB+ average, 50 shots (offline) | accuracy: 49.9 -> 64.5 | Improved |
| add FiLM adaptation to ResNet18 backbone (ablation) | ResNet18 · VTAB+ average, 50 shots (offline) | accuracy: 49.6 -> 59.5 | Improved |
| meta-learning based body adaptation instead of FiLM (ablation) | FSA · VTAB+ average, 50 shots (offline) | accuracy: 64.5 -> 51.7 | Hurt |
| full-body adaptation instead of FiLM (ablation) | FSA · VTAB+ average, 50 shots (offline) | accuracy: 64.5 -> 64.9 | No clear effect |
| FiLM adaptation versus no adaptation on 50-shot CL (ablation) | FSA-FiLM · CIFAR100 (50-shot CL) | accuracy: 57.4 -> 70.9 | Improved |
| FSA-FiLM versus no adaptation on SVHN 50-shot CL (ablation) | FSA-FiLM · SVHN (50-shot CL) | accuracy: 28.3 -> 51.3 | Improved |
| FSA-FiLM versus a ConvNext pretrained on ImageNet-22K without adaptation | FSA-FiLM · 50-shot CL, CIFAR100 and DomainNet | accuracy: CIFAR100 87.1 vs 70.9; SVHN 43.6 vs 51.3; dSprites-loc 13.6 vs 35.7 | Mixed |
| use 50-shot/2-way sessions instead of 10-shot/10-way (ablation) | FSA-FiLM · VTAB+ average CL | accuracy: 52.6 -> 55.8 | Improved |
| FSA full-body adaptation versus FiLM in high-shot first session with many classes (ablation) | FSA with EfficientNet-B0 · CIFAR100 (large first session) | accuracy: 61.8 -> 66.0 | Improved |
| use FSA with full body adaptation on a small first session of 10 classes (ablation) | FSA · CIFAR100 (full shots, 10-class first session) | accuracy: FSA 62.8 vs FSA-FiLM 73.8 vs NA 68.2 | Hurt |
| FSA-FiLM versus FACT on CIFAR100 full-shot sessions | FSA-FiLM · CIFAR100 (full-shot, 10 classes per session) | accuracy: 10.2 -> 73.8 | Improved |
| replay-free FSA-FiLM versus GDumb replay on CORE50 | FSA-FiLM · CORE50 | accuracy: 90.0 vs 85.4 | 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.
- GDumb — GDumb: A Simple Approach that Questions Our Progress in Continual Learning (ECCV 2020) · better on 2 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
- Attraction Diminishing and Distributing for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.85)
- Autoencoder-Based Hybrid Replay for Class-Incremental Learning (ICML 2025 · similarity 0.84)
- SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025 · similarity 0.84)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.84)
- CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · similarity 0.83)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.83)
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.
