Research map / Architectures and efficiency / Pruning and efficient tuning
Incremental Learning Through Deep Adaptation
Amir Rosenfeld, John K. Tsotsos
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 |
|---|---|---|---|
| use ImageNet-pretrained base network instead of Sketch-trained base (ablation) | DAN · 8 datasets | mean top-1 accuracy: 83.7 -> 86.5 | Improved |
| use selected best base per dataset (ImageNet or Sketch) instead of single base (ablation) | DAN · 8 datasets | mean top-1 accuracy: 86.46 -> 87.76 | Improved |
| use two base networks concatenated (Caltech-256 and Sketch) (ablation) | DAN · 8 datasets | mean top-1 accuracy: same as DAN_sketch | No clear effect |
| use randomly initialized base network (ablation) | DAN · 8 datasets | mean top-1 accuracy: 79.9 -> 76.3 | Hurt |
| use random base instead of Daimler-trained base (ablation) | DAN · 8 datasets | mean top-1 accuracy: 75 -> 76.3 | Improved |
| use DAN controller from Sketch base instead of full fine-tuning | DAN sketch · 8 datasets | mean top-1 accuracy: 84.2 -> 83.7 | No clear effect |
| fine-tune all layers instead of last layer only (ablation) | Sketch base · 8 datasets | mean transfer accuracy: 50.3 -> 84.2 | Improved |
| freeze batch-norm during full fine-tuning (ft-full-bn-off) (ablation) | Sketch base · 8 datasets | mean transfer accuracy: 84.2 -> 82.5 | Hurt |
| use DAN instead of Residual Adapters | DAN · Visual Decathlon | decathlon score: 2118 -> 2851 | Improved |
| use DAN instead of Residual Adapters (joint training) | DAN · Visual Decathlon | mean accuracy: 77.17 -> 77.01 | Mixed |
| use DAN instead of vanilla fine-tuning | DAN · Visual Decathlon | decathlon score: 2500 -> 2851 | Improved |
| quantize DAN weights to 8 bits (ablation) | DAN · 8 datasets | accuracy | No clear effect |
Closest papers by meaning
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.89)
- Adaptive Neural Networks for Efficient Inference (ICML 2017 · similarity 0.88)
- Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (TPAMI 2022 · similarity 0.88)
- Incremental Learning Using Conditional Adversarial Networks (ICCV 2019 · similarity 0.88)
- RePr: Improved Training of Convolutional Filters (CVPR 2019 · similarity 0.87)
- Incremental Learning via Rate Reduction (CVPR 2021 · 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.
