Research map / Continual learning / Class-incremental
Adaptive Aggregation Networks for Class-Incremental Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 AANets to PODNet-CNN | PODNet-CNN · CIFAR-100, ImageNet-Subset, and ImageNet; N=5, 10, 25 | average incremental accuracy (%): 64.83 -> 66.31; 63.19 -> 64.31; 60.72 -> 62.31; 75.54 -> 76.96; 74.33 -> 75.58; 68.31 -> 71.78; 66.95 -> 67.73; 64.13 -> | Helped |
| Use separate plastic and stable blocks with adaptive aggregation rather than a single LUCIR branch | LUCIR · CIFAR-100 and ImageNet-Subset, N=5, 10, 25 | accuracy (%): 63.17 -> 66.74; 60.14 -> 65.29; 57.54 -> 63.50; 70.84 -> 72.55; 68.32 -> 69.22; 61.44 -> 67.60 | Helped |
| Add AANets to each of iCaRL, LUCIR, Mnemonics, and PODNet-CNN | Four CIL baselines · CIFAR-100, ImageNet-Subset, and ImageNet | average incremental accuracy (%) | Helped |
| Use all-learnable blocks in both branches (ablation) | AANets · CIFAR-100 and ImageNet-Subset, N=25 | accuracy (%): 63.50 -> 58.87; 67.60 -> 63.29 | Hurt |
| Use scaling weights in the plastic branch and freeze the stable branch (ablation) | AANets · CIFAR-100 and ImageNet-Subset, N=25 | accuracy (%): 63.50 -> 62.89; 67.60 -> 70.30 | Mixed |
| Train aggregation weights on imbalanced data instead of a balanced exemplar subset (ablation) | AANets · CIFAR-100 and ImageNet-Subset, N=25 | accuracy (%): 63.50 -> 63.08; 67.60 -> 66.89 | Hurt |
| Fix aggregation weights at 0.5 rather than adapting them (ablation) | AANets · CIFAR-100 and ImageNet-Subset, N=25 | accuracy (%): 63.50 -> 62.89; 67.60 -> 66.34 | Hurt |
| Reduce exemplar count to meet the strict memory budget (ablation) | AANets with LUCIR · CIFAR-100 and ImageNet-Subset, N=5 | accuracy (%): 66.74 -> 66.46; 72.55 -> 72.21 | Hurt |
| Use four all-learnable blocks at each residual level (ablation) | AANets · CIFAR-100 and ImageNet-Subset, N=25 | accuracy (%): 63.50 -> 59.40; 67.60 -> 64.00 | Hurt |
| Use four scaling blocks at each residual level (ablation) | AANets · CIFAR-100 and ImageNet-Subset, N=25 | accuracy (%): 63.50 -> 63.16; 67.60 -> 67.12 | Hurt |
| Update the frozen base model after each phase (ablation) | AANets with LUCIR · CIFAR-100, same number of classes across phases, N=5, 10, 25 | last-phase accuracy (%): 52.6 -> 54.3; 46.1 -> 47.4; 41.9 -> 42.4 | Helped |
| Update the frozen base model after each phase (ablation) | AANets with LUCIR · ImageNet-Subset, same number of classes across phases, N=5, 10, 25 | last-phase accuracy (%): 68.8 -> 61.4; 60.8 -> 52.5; 56.8 -> 48.2 | Mixed |
Closest papers by meaning
- Class-Incremental Learning With Strong Pre-Trained Models (CVPR 2022 · similarity 0.90)
- On the Stability-Plasticity Dilemma of Class-Incremental Learning (CVPR 2023 · similarity 0.89)
- Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning (ICML 2024 · similarity 0.86)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.86)
- Order-Robust Class Incremental Learning: Graph-Driven Dynamic Similarity Grouping (CVPR 2025 · similarity 0.86)
- AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection (NeurIPS 2025 · 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.