Research map / Continual learning / Domain and prompt-based
Componential Prompt-Knowledge Alignment for Domain Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KA-Prompt, a prompt-based domain incremental learning method that aligns component-wise prompt structure via greedy reusable-prompt mining for initialization and an online alignment loss against old prompts.
Problem
Domain incremental learning with prompt-based methods suffers from component-wise misalignment between independently learned domain prompts, causing conflicting knowledge when prompts are fused at inference.
Approach
Initialize new prompts from a greedily searched set of old prompts that cover new-domain knowledge (reusable knowledge mining), then during training apply a weighted alignment loss that pulls the fused prompt toward similar old prompts (online alignment preservation), building on C-Prompt.
Key results
- KA-Prompt reaches 71.30% average Avg-ACC across four DIL benchmarks versus 66.57% for C-Prompt (+4.73%).
- On DomainNet KA-Prompt gets 62.91% vs 58.66% for C-Prompt; on ImageNet-R 66.51% vs 62.43%.
- Reusable knowledge mining gives 1.81% over using the previous-stage prompt for initialization; online alignment gives 2.28% over the baseline in ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KA-Prompt instead of C-Prompt | DIL models · DomainNet | Avg-ACC: 4.25% | Helped |
| Use KA-Prompt instead of C-Prompt | DIL models · ImageNet-R | Avg-ACC: 4.08% | Helped |
| Use KA-Prompt instead of C-Prompt | DIL models · ImageNet-C | Avg-ACC: 5.59% | Helped |
| Use KA-Prompt instead of C-Prompt | DIL models · ImageNet-Mix | Avg-ACC: 5.00% | Helped |
| Use KA-Prompt instead of C-Prompt | DIL models · Average across four DIL benchmarks | Avg-ACC: 4.73% | Helped |
| Add reusable knowledge mining (prompt-data relation extraction and greedy prompt search) to new prompt training (ablation) | KA-Prompt · DIL benchmarks | Avg-ACC: 1.81% | Helped |
| Add online historical prompt alignment (ablation) | KA-Prompt · DIL benchmarks | Avg-ACC: 2.28% | Helped |
| Use prompt-data relation extraction or greedy prompt search alone (ablation) | KA-Prompt · DIL benchmarks | Avg-ACC | No clear effect |
| Add all proposed modules together (ablation) | KA-Prompt · DIL benchmarks | Avg-ACC | Helped |
| Shuffle prompt components (ablation) | KA-Prompt · DomainNet last domain | accuracy: 6.4% | Hurt |
| Shuffle prompt components under Shuffle-B/C/D/E (ablation) | C-Prompt · DomainNet last domain | accuracy | Helped |
| Shuffle prompt components (ablation) | KA-Prompt · DomainNet last domain | accuracy | Hurt |
Surprising findings
- Randomly shuffling prompt components of C-Prompt before fusion sometimes improves its accuracy, indicating its learned component order is suboptimal.
Limitations
- Training overhead is higher than C-Prompt due to the online alignment module (f_A).
- Greedy search is approximate; the exhaustive search is NP-hard, and the method relies on heuristic selection.
Experimental setup
Datasets: DomainNet ImageNet-R ImageNet-C ImageNet-Mix
Benchmarks: DomainNet ImageNet-R ImageNet-C ImageNet-Mix
Metrics: Avg-ACC
Baselines: EWC LwF L2P S-Prompts DualPrompt ESN CODA-Prompt InfLoRA CPrompt C-Prompt
Builds on: C-Prompt L2P DualPrompt CODA-Prompt
Closest papers by meaning
- Prompt-Based Distribution Alignment for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.88)
- Enhancing Domain Adaptation through Prompt Gradient Alignment (NeurIPS 2024 · similarity 0.87)
- BayesPrompt: Prompting Large-Scale Pre-Trained Language Models on Few-shot Inference via Debiased Domain Abstraction (ICLR 2024 · similarity 0.86)
- CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class Incremental Learning (AAAI 2025 · similarity 0.86)
- Mix-CPT: A Domain Adaptation Framework via Decoupling Knowledge Learning and Format Alignment (ICLR 2025 · similarity 0.86)
- Adapting to Distribution Shift by Visual Domain Prompt Generation (ICLR 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.