Research map / Continual learning / Online and replay-based
Adaptive Retention & Correction: Test-Time Training for 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 ARC, a test-time plug-in for memory-free class-incremental learning that detects past-task samples at inference, re-balances the classifier with one entropy-regularized gradient step (ARR), and corrects misclassified predictions via task-based softmax scores (ARC-C), boosting existing CL methods by about 2.7% on CIFAR-100 and 2.6% on Imagenet-R.
Problem
Classifier bias toward the most recent task in class-incremental continual learning, especially in memory-free settings where past-task training data cannot be replayed.
Approach
Exploit past-task test samples available at inference: an Out-of-Task detector based on softmax confidence and predicted class routes samples to either a one-step online classifier re-balancing (cross-entropy on pseudo-labels plus entropy minimization) or a training-free task-based softmax score correction, with no change to training.
Key results
- Plugged into Finetune, iCaRL, Memo, DER, L2P, DualPrompt, CodaPrompt and SLCA, ARC raises average accuracy by about 2.7% on Split CIFAR-100 and 2.6% on Split Imagenet-R across settings.
- On Split CIFAR-100 Inc5, DER with ARC rises from 70.7 to 74.0 average accuracy, and Finetune with ARC rises from 64.6 to 66.0.
- Average accuracy gains are larger on long task sequences (3.5% and 3.0% on Inc5/Inc20 settings as stated) than on short ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ARC to continual learning methods | Existing CL methods · Split CIFAR-100 | Average Accuracy: 2.7% | Helped |
| Add ARC to continual learning methods | Existing CL methods · Split Imagenet-R | Average Accuracy: 2.6% | Helped |
| Add ARC to continual learning methods | Existing CL methods · Split CIFAR-100 and Split Imagenet-R | Average Accuracy: 2.7% on both | Helped |
| Add ARC to continual learning methods | Existing CL methods · Split CIFAR-100 Inc10 | Average Accuracy: +6.8 | Helped |
| Add ARC to continual learning methods | Existing CL methods · Split Imagenet-R Inc20 | Average Accuracy: +6.6 | Helped |
| Add ARC-R (ablation) | ARC · Split CIFAR-100 and Split Imagenet-R | Average Accuracy: 1.8% | Helped |
| Add ARC-C (ablation) | ARC · Split CIFAR-100 | Average Accuracy: 1.1% | Helped |
| Add ARC-C (ablation) | ARC · Split Imagenet-R | Average Accuracy: 1.2% | Helped |
| Remove the TSS scaling temperature (ablation) | ARC-C · Split CIFAR-100 | Average Accuracy: 0.6% | Hurt |
| Replace the confidence ratio w with raw confidence c (ablation) | ARC-C · Split CIFAR-100 | Average Accuracy: 0.5% | Hurt |
| Vary the beta and gamma thresholds (ablation) | ARC · Split CIFAR-100 | Average Accuracy | No clear effect |
| Use cross-entropy loss in ARC-R (ablation) | ARC-R with L2P · Split CIFAR-100 | Average Accuracy | Hurt |
Surprising findings
- Linear probing on frozen pretrained representations after training on each new task achieves performance comparable to joint training, suggesting the representation layer is not the main source of forgetting.
- Applying CoTTA to prompt-based methods decreases their performance.
Limitations
- Gains from the correction component are small for prompt-based methods because few test samples satisfy the misclassification assumption.
- Adds inference-time cost, and the thresholds beta and gamma must be set by hand.
Experimental setup
Datasets: Split CIFAR-100 Split Imagenet-R 5-dataset
Benchmarks: Split CIFAR-100 Split Imagenet-R 5-dataset
Metrics: average accuracy forgetting
Baselines: Finetune iCaRL Memo DER L2P DualPrompt CodaPrompt SLCA TENT CoTTA
Builds on: SLCA TENT SHOT CoTTA
Closest papers by meaning
- Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · similarity 0.88)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 2025 · similarity 0.88)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.88)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.87)
- An Effective Dynamic Gradient Calibration Method for Continual Learning (ICML 2024 · similarity 0.87)
- BECAME: Bayesian Continual Learning with Adaptive Model Merging (ICML 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.