Adaptive Retention & Correction: Test-Time Training for Continual Learning
Authors: Haoran Chen, Micah Goldblum, Zuxuan Wu, Yu-Gang Jiang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our proposed method can be plugged in to virtually any existing continual learning approach without requiring any modifications to its training procedure. Specifically, when integrated with state-of-the-art approaches, ARC achieves an average performance increase of 2.7% and 2.6% on the CIFAR-100 and Imagenet-R datasets, respectively. |
| Researcher Affiliation | Academia | 1Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University 2Shanghai Collaborative Innovation Center of Intelligent Visual Computing 3New York University. All authors are affiliated with universities or public research centers, indicating an academic affiliation. |
| Pseudocode | Yes | Algorithm 1 Adaptive Retention & Correction (ARC) |
| Open Source Code | Yes | Code is available at Github Link. |
| Open Datasets | Yes | We conduct our main experiments on two popular benchmark datasets for continual learning: Split CIFAR-100 and Split Imagenet-R. CIFAR-100 is a relatively simple dataset comprising 100 classes. Imagenet-R, on the other hand, includes 200 classes... We also conduct experiments on another challenging benchmark, 5-dataset, where unlike the first two, it is a composite of five distinct image classification datasets: CIFAR-10, MNIST, Fashion MNIST, SVHN, and not MNIST. |
| Dataset Splits | Yes | We use Inc-n to denote the data split setting, where n denotes the number of classes for each incremental stage. For example, Split CIFAR-100 Inc5 means that there are 5 classes for each training step, resulting in a total of 20 tasks. For a fair comparison, we split the classes following the order of Sun et al. (2023). |
| Hardware Specification | No | No specific hardware details (like GPU/CPU models, processor types, or memory amounts) were mentioned in the paper text. |
| Software Dependencies | No | The paper mentions using "the repository provided by Sun et al. (2023)" as a toolbox, but does not provide specific software names with version numbers. |
| Experiment Setup | Yes | Implementation details. We use the repository provided by Sun et al. (2023) to test the effectiveness of our method... Finetune, i Car L, Memo, and Der are equipped with a memory buffer of 2000 and 4000 for Split CIFAR-100 and Split Imagenet-R, respectively, whereas L2P, Dual Prompt, Coda Prompt and SLCA are memory-free. For all tested methods, we utilize a Vi T-b-16 backbone pretrained on Imagenet-21K. We set the training configuration to be the same as provided. |