Research map / Continual learning / Domain and prompt-based
When Prompt-based Incremental Learning Does Not Meet Strong Pretraining
Yu-Ming Tang, Yi-Xing Peng, Wei-Shi Zheng
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 L_conC knowledge-pool constraint on the classifier (ablation) | APG · ImageNet-Sub B50-T10 | Avg. Acc.: 19.37 -> 71.84 | Improved |
| add L_conA constraint on the APG (ablation) | APG · ImageNet-Sub B50-T10 | Avg. Acc.: 71.84 -> 73.46 | Improved |
| add class-specific attention loss L_attn (ablation) | APG · ImageNet-Sub B50-T10 | Avg. Acc.: 73.46 -> 74.70 | Improved |
| add triplet loss L_tri (ablation) | APG · ImageNet-Sub B50-T10 | Avg. Acc.: 74.70 -> 75.94 | Improved |
| place APG at layer 10 instead of layers 6-9 (ablation) | APG · ImageNet-Sub B50-T10 | Avg. Acc. | Hurt |
| generate 1 prompt instead of none (N_P=0) (ablation) | APG · ImageNet-Sub B50-T10 | Avg. Acc.: +4.1% | Improved |
| increase number of generated prompts beyond 1 (ablation) | APG · ImageNet-Sub B50-T10 | Avg. Acc. | No clear effect |
| replace prompt-pool retrieval with learnable adaptive prompt generator (non-pretrained) | L2P / DualPrompt · ImageNet-Subset B50-T10 | Avg. Acc.: 35.82 -> 75.52 (vs DualPrompt) | Improved |
| use APG with Tiny-ImageNet pretraining instead of prompt pool | L2P / DualPrompt · CIFAR100 B0-20 (Tiny-ImageNet pretrained) | Avg. Acc.: 79.13 -> 87.50 (vs DualPrompt) | Improved |
| use APG with strong ImageNet-21k pretraining, 5 tasks | DualPrompt · CIFAR100 B0-5 (ImageNet-21K pretrained) | Avg. Acc.: 89.56 -> 88.48 | No clear effect |
| use APG instead of DualPrompt with 20 tasks under ImageNet-21k pretraining | DualPrompt · ImageNet-R B0-T20 | Avg. Acc.: 70.48 -> 71.22 | Improved |
| exemplar-free APG vs rehearsal methods with 20 exemplars/class, 50 tasks | AFC · ImageNet-Subset B50-T50 | Avg. Acc.: 71.97 -> 75.70 | Improved |
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.
- DualPrompt — DualPrompt: Complementary Prompting for Rehearsal-Free Continual Learning (ECCV 2022) · better on 6 of 8 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
- Evolving Parameterized Prompt Memory for Continual Learning (AAAI 2024 · similarity 0.88)
- Mixture of Experts Meets Prompt-Based Continual Learning (NeurIPS 2024 · similarity 0.88)
- Vector Quantization Prompting for Continual Learning (NeurIPS 2024 · similarity 0.88)
- Convolutional Prompting meets Language Models for Continual Learning (CVPR 2024 · similarity 0.88)
- REP: Resource-Efficient Prompting for Rehearsal-Free Continual Learning (NeurIPS 2025 · similarity 0.87)
- Training Consistent Mixture-of-Experts-Based Prompt Generator for Continual Learning (AAAI 2025 · 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.
