Research map / Continual learning / Class-incremental
OVOR: OnePrompt with Virtual Outlier Regularization for Rehearsal-Free Class-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 OVOR, which adds a NPOS-based virtual outlier hinge-loss regularizer to a single-prompt method (OnePrompt) to reduce inter-task confusion in rehearsal-free class-incremental learning.
Problem
Rehearsal-free class-incremental learning suffers from inter-task confusion because classes from different tasks are never trained together, and prompt-pool methods add query and composition cost.
Approach
Generate virtual outliers in feature space with NPOS and add an energy-based hinge regularizer on the classifier head, which is applied to a single shared prompt (OnePrompt) to tighten decision boundaries without prompt pools or separate outlier detectors.
Key results
- OVOR on 10-task ImageNet-R reaches 75.61% average accuracy versus 73.34% for OnePrompt without regularization.
- VOR improves L2P, DualPrompt, CODA-P and OnePrompt across all four datasets; e.g. CUB-200 CODA-P rises from 72.74% to 78.72%.
- OVOR uses 0.26M learnable parameters and 17.60 GFLOPs versus 3.99M and 35.17 GFLOPs for CODA-P on ImageNet-R 10-task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add virtual outlier regularization (VOR) | L2P · 10-task ImageNet-R | A_T: +1.83 | Helped |
| add virtual outlier regularization (VOR) | DualPrompt · 10-task ImageNet-R | A_T: +1.80 | Helped |
| add virtual outlier regularization (VOR) | CODA-P · 10-task ImageNet-R | A_T: +1.95 | Helped |
| add virtual outlier regularization (VOR) | OnePrompt · 10-task ImageNet-R | A_T: +2.27 | Helped |
| add VOR to prompt-based methods | L2P, DualPrompt, CODA-P, OnePrompt · CIFAR-100, CUB-200, ImageNet-A (10-task) | A_T: gains on all four methods across datasets, largest on CUB-200 (e.g. OnePrompt 72.19 to 78.12) | Helped |
| single shared prompt (OnePrompt) instead of prompt pool | prompt-based CIL · 10-task ImageNet-R | A_T: OnePrompt 73.34 vs L2P 69.31 and DualPrompt 71.47 | Helped |
| OnePrompt with VOR (OVOR) vs CODA-P | OnePrompt · 10-task ImageNet-R | A_T: OVOR 75.61 vs CODA-P 74.87 (CODA-P-VOR 76.82) | Helped |
| use deeper prompt insertion (OnePrompt-Deep, length 5 in all layers) | OnePrompt · 20-task ImageNet-R | A_T: 71.35 vs 70.42 | Helped |
| regularized training over last 10% of epochs instead of 20% (ablation) | OVOR · 10-task ImageNet-R | A_T: 74.42 vs 75.61 | Hurt |
| regularized training over last 30% of epochs instead of 20% (ablation) | OVOR · 10-task ImageNet-R | A_T: 73.99 vs 75.61 | Hurt |
| MSE loss instead of Huber loss in regularization (ablation) | OVOR · 10-task ImageNet-R | A_T: 73.75 vs 75.61 | Hurt |
| regularization weight lambda varied from 0.1 to 0.05, 0.5, 1.0 (ablation) | OVOR · 10-task ImageNet-R | A_T: 75.23 to 75.61 (0.5 and 0.1 equal) | No clear effect |
Surprising findings
- OnePrompt with a single shared prompt outperforms L2P and DualPrompt, which use prompt pools.
- VOR increases average forgetting for several methods (e.g. CODA-P on ImageNet-R rises from 6.73 to 8.05) even while raising accuracy.
Limitations
- Only applicable when task input spaces are non-overlapping; not directly applicable to domain-incremental or blurry CIL.
- Outlier generation requires storing and searching feature vectors, raising scalability concerns as datasets grow.
- Forgetting metric is not consistently improved by VOR.
Experimental setup
Datasets: ImageNet-R CIFAR-100 CUB-200 ImageNet-A
Benchmarks: ImageNet-R CIFAR-100 CUB-200 ImageNet-A
Metrics: average accuracy average forgetting
Baselines: L2P DualPrompt CODA-P L2 SimpleCIL ADAM A-Prompt
Builds on: L2P DualPrompt CODA-Prompt NPOS prefix tuning
Closest papers by meaning
- F-OAL: Forward-only Online Analytic Learning with Fast Training and Low Memory Footprint in Class Incremental Learning (NeurIPS 2024 · similarity 0.86)
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.86)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.85)
- Multi-layer Rehearsal Feature Augmentation for Class-Incremental Learning (ICML 2024 · similarity 0.85)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.85)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.85)
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.