Research map / Continual learning / Domain and prompt-based
Learning to Prompt Knowledge Transfer for Open-World Continual 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 Pro-KT, a rehearsal-free prompt-bank method with task-aware adaptive thresholds for open-world class-incremental continual learning that detects unknowns and classifies knowns.
Problem
Continual learning in an open world, where each new task may contain novel classes; models must avoid forgetting known classes while detecting unknowns and later learning them as knowns.
Approach
A prompt bank with learnable keys stores task-specific and task-generic prompts, selected per sample by key matching for knowledge transfer, combined with an open-set boundary given by adaptive thresholds on softmax entropy (task-ID or latest-task threshold selection).
Key results
- On Split CIFAR100 (ViT), Pro-KT reaches 84.07% average final accuracy with 5.43% forgetting, versus 83.18% for DualPrompt.
- On 5-datasets, Pro-KT achieves 88.60 AUC_N and 45.70 FPR_N for unknown detection, versus 66.78 AUC_N for the best MaxLogits baseline.
- Removing the prompt bank drops Split CIFAR100 known accuracy from 84.07% to 40.49%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Pro-KT instead of the strongest reported baseline | Unknown detection · Split CIFAR100 (10 tasks) | AUC_N: 52.63 -> 92.69 | Helped |
| use Pro-KT instead of the strongest reported baseline | Unknown detection · 5-datasets (5 tasks) | AUC_N: 66.78 -> 88.60 | Helped |
| use Pro-KT instead of the strongest reported baseline | Known classification with ViT · Split CIFAR100 (10 tasks) | A_N: 83.18 -> 84.07 | Helped |
| use Pro-KT instead of the strongest reported baseline | Known classification with ViT · 5-datasets (5 tasks) | A_N: 70.14 -> 71.70 | Helped |
| add task IDs for threshold selection (ablation) | Pro-KT · Split CIFAR100 (10 tasks) | AUC_N: 91.01 -> 92.69 | Helped |
| add task IDs for threshold selection (ablation) | Pro-KT · 5-datasets (5 tasks) | AUC_N: 82.76 -> 88.60 | Helped |
| add the task-aware open-set boundary (ablation) | Pro-KT · Split CIFAR100 (10 tasks) | AUC_N: 50.02 -> 92.69 | Helped |
| add the task-aware open-set boundary (ablation) | Pro-KT · 5-datasets (5 tasks) | AUC_N: 46.78 -> 88.60 | Helped |
| add the prompt bank instead of fine-tuning the backbone and classifier (ablation) | Pro-KT · Split CIFAR100 (10 tasks) | A_N: 40.49 -> 84.07 | Helped |
| add the prompt bank instead of fine-tuning the backbone and classifier (ablation) | Pro-KT · 5-datasets (5 tasks) | A_N: 15.42 -> 71.70 | Helped |
| increase prompt selection size K (ablation) | Pro-KT · Split CIFAR100 and 5-datasets | A_N | Hurt |
| vary prompt length L_P (ablation) | Pro-KT · Split CIFAR100 and 5-datasets | A_N | No clear effect |
Surprising findings
- Pro-KT without task IDs still outperforms all unknown-detection baselines, and the paper notes that the 5-datasets setting is more sensitive to the number of prompts M.
Limitations
- Unknown samples are only grouped as unknown and not fine-grained clustered; future work targets this.
- Task-ID-free threshold selection performs worse on dissimilar tasks (5-datasets AUC_N drops from 88.60 to 82.76).
Experimental setup
Datasets: Split CIFAR100 5-datasets
Benchmarks: Split CIFAR100 5-datasets
Metrics: AUC_N FPR_N A_N F_N
Baselines: OpenMax MSP MCD EnergyBased Entropy MaxLogits ODIN EWC MAS DER++ GEM LwF
Builds on: L2P DualPrompt MAS
Closest papers by meaning
- Convolutional Prompting meets Language Models for Continual Learning (CVPR 2024 · similarity 0.87)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.86)
- Rethinking Momentum Knowledge Distillation in Online Continual Learning (ICML 2024 · similarity 0.86)
- C$^2$Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning (NeurIPS 2025 · similarity 0.86)
- Advancing Prompt-Based Methods for Replay-Independent General Continual Learning (ICLR 2025 · similarity 0.86)
- Scalable Language Model with Generalized Continual Learning (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.