近年來，深度學習模型已在視覺辨識任務中取得顯著成果。然而，在許多實際 應用中，資料會隨時間依序到達，模型必須在不重新從頭訓練的情況下持續整合 新知識。持續學習(continual learning, CL)旨在使模型能夠從循序資料中學習， 同時避免遺忘先前取得的知識。在持續學習的不同情境中，類別增量學習(class- incremental learning, CIL)要求模型隨時間學習新類別，並在推論階段無法取得任務 識別資訊的情況下辨識所有已學習的類別。在此情境下，模型適應新類別時，容易 遺忘先前學習的類別，此現象稱為災難性遺忘(catastrophic forgetting)。為了減輕 此問題，回放式方法(replay-based methods)會在記憶緩衝區中儲存少量先前類別的 樣本，並使用回放樣本與新類別資料共同訓練模型。此策略不需要儲存所有先前資 料，仍能協助模型保留既有知識，因此已廣泛應用於類別增量學習。儘管回放式類 別增量學習方法能有效減少遺忘，多數方法仍使用乾淨影像進行訓練與評估，並通 常假設訓練資料與測試資料來自相同或相似的分布。在實際應用中，視覺觀測可能 受到光照變化、感測器雜訊、影像模糊與壓縮失真等視覺擾動影響。這些擾動通常 不會改變輸入的語意類別，但會改變其視覺外觀，由此產生的分布偏移可能降低模 型的預測準確率。因此，如何在維持類別增量學習效能的同時，提升模型對常見視 覺擾動的穩健性，是一項重要的實務挑戰。
為解決此問題，本研究提出應用於回放式類別增量學習的領域一致性正則化 (Domain Consistency Regularization, DCR)方法。DCR 鼓勵模型對可靠回放樣本的 乾淨版本與受擾動版本產生一致的預測。此方法不修改模型架構、記憶更新機制或 回放策略，而是在回放訓練過程中加入一致性正則化目標。在訓練期間，DCR 根據 乾淨回放樣本產生保留語意的受擾動版本，並以乾淨版本的預測作為一致性正則化 的依據。為了減少不可靠的一致性監督，DCR 根據預測正確性、預測信心及決策 邊界間隔，選擇適合作為一致性參考的可靠回放樣本。本研究在三個類別增量學習 基準上評估 DCR，包括 Seq-CIFAR-10、Seq-CIFAR-100 與 Seq-TinyImageNet，並使 用 CIFAR-10-C、CIFAR-100-C 與 TinyImageNet-C 評估模型面對常見視覺擾動時的穩 健性。實驗結果顯示，在不同資料集、記憶緩衝區大小及回放式基準方法下，DCR 均能提高受擾動輸入的準確率，並縮小乾淨輸入與受擾動輸入之間的穩健性差距。 同時，DCR 在乾淨資料上仍能維持具競爭力的類別增量學習效能。進一步分析顯 示，DCR 所學習的預測穩定性也適用於不同類型的視覺擾動。本研究也將 DCR 調整後應用於以 UR5 機器人操作任務為基礎的離線行為複製情境，以檢驗此方法在影像 分類以外任務中的適用性。實驗結果顯示，調整後的 DCR 方法可降低模型面對受 擾動視覺觀測時的連續動作預測誤差，同時在乾淨輸入上維持相近的效能。整體而 言，研究結果顯示 DCR 是一種輕量且有效的正則化方法，可提升回放式類別增量學 習模型在視覺擾動下的穩健性，其核心一致性原則也可應用於連續動作預測任務。

Deep learning models have achieved remarkable success in visual recognition tasks. However, in many real-world applications, data arrive sequentially over time, and models must continu- ously integrate new knowledge without retraining from scratch. Continual learning (CL) aims to enable models to learn from sequential data while avoiding forgetting previously acquired knowledge. Among CL scenarios, class-incremental learning (CIL) requires models to learn new classes over time and recognize all learned classes at inference time without access to task identity. In this setting, models tend to forget previously learned classes when adapting to new ones, a phenomenon known as catastrophic forgetting. To mitigate this issue, replay-based methods store a limited number of samples from previously learned classes in a memory buffer and train the model using both replay samples and new-class data. This strategy helps preserve previous knowledge without storing all previous data and has therefore been widely adopted in CIL. Despite their effectiveness in reducing forgetting, most replay-based CIL methods are trained and evaluated on clean images. These methods commonly assume that the training and test data follow the same or similar distributions. In practical applications, visual observations may be affected by illumination changes, sensor noise, image blur, and compression artifacts. These corruptions often preserve the semantic identity of an input but alter its visual appearance. The resulting distribution shifts can reduce prediction accuracy. Improving robustness to common visual corruptions while maintaining CIL performance is therefore an important practical challenge.
To address this problem, we develop a Domain Consistency Regularization (DCR) formulation for replay-based CIL. DCR encourages consistent predictions between the clean and corrupted versions of reliable replay samples. It does not modify the model architecture, memory update mechanism, or replay strategy. Instead, it adds a consistency regularization objective to the replay training process. During training, DCR generates semantics-preserving corrupted versions from clean replay samples and uses the clean predictions as references for consistency regularization. To reduce unreliable consistency supervision, DCR selects reliable replay samples based on prediction correctness, prediction confidence, and decision- boundary margin. We evaluate DCR on three CIL benchmarks, including Seq-CIFAR- 10, Seq-CIFAR-100, and Seq-TinyImageNet. Robustness to common visual corruptions is evaluated using CIFAR-10-C, CIFAR-100-C, and TinyImageNet-C. The experimental results show that DCR improves corrupted-input accuracy and reduces the robustness gap between clean and corrupted inputs across multiple datasets, buffer sizes, and replay-based baselines. At the same time, it maintains competitive CIL performance on clean data. Further analysis indicates that the predictive stability learned by DCR applies across corruption types. We further adapt DCR to an offline sequential behavior cloning setting with UR5 robotic manipulation tasks to examine its use beyond image classification. The results show that the adapted DCR formulation reduces continuous action prediction errors under corrupted visual observations while maintaining comparable performance on clean inputs. Overall, the results suggest that DCR is a lightweight and effective regularization approach for improving the corruption robustness of replay-based CIL. The UR5 results further indicate that the underlying consistency principle can be applied to continuous action prediction tasks.

[1]  German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54– 71, 2019.
[2]  Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist networks. Psychology of Learning and Motivation, 24:109–165, 1989.
[3]  David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Greg Wayne. Experience replay for continual learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019.
[4]  Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020.
[5]  Elaheh Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. In Interna- tional Conference on Learning Representations (ICLR), 2022.
[6]  Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR), 2017.
[7]  Antti Tarvainen and Harri Valpola. Mean teachers are better role models. In Advances in Neural Information Processing Systems (NeurIPS), 2017.
[8]  Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi- supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems (NeurIPS), 2020.
[9]  Dan Hendrycks, Norman Mu, Ekin D. Cubuk, et al. Augmix: A simple data processing method to improve robustness and uncertainty. In International Conference on Learning Representations (ICLR), 2020.
[10]  James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences (PNAS), 114(13):3521–3526, 2017.
[11] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International Conference on Machine Learning (ICML), 2017.
[12]  Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016.
[13]  Zhizhong Li and Derek Hoiem. Learning without forgetting. In European Conference on Computer Vision (ECCV), 2016.
[14]  Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, et al. Learning to prompt for continual learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 139–149, 2022.
[15]  Zifeng Wang, Chen-Yu Lee, et al. Dualprompt: Complementary prompting for rehearsal- free continual learning. In European Conference on Computer Vision (ECCV), 2022.
[16]  James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda- prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11909–11919, 2023.
[17]  Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lam- pert. icarl: Incremental classifier and representation learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
[18]  Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Alesˇ Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366–3385, 2022.
[19]  Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations (ICLR), 2019.
[20]  Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, et al. Towards deep learn- ing models resistant to adversarial attacks. In International Conference on Learning Representations (ICLR), 2018.
[21]  Dequan Wang, Evan Shelhamer, Shaoteng Liu, et al. Continual test-time domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: A regularization method for supervised and semi-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(8):1979–1993, 2019.
[23]  ChongkaiGao,HaichuanGao,ShangqiGuo,TianrenZhang,andFengChen.CRIL:Con- tinual robot imitation learning via generative and prediction model. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6747–6754. IEEE, 2021.
[24]  Weikang Wan, Yifeng Zhu, Rutav Shah, and Yuke Zhu. LOTUS: Continual imitation learning for robot manipulation through unsupervised skill discovery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 537–544. IEEE, 2024.
[25]  Jogendra Nath Kundu, Prasanna Uppala, Anoop Pahuja, and R. Venkatesh Babu. Adadepth: Unsupervised content congruent adaptation for depth estimation. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2656–2665, 2018.
[26]  Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009.
[27]  Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. Technical report, Stanford University, 2015.
[28]  Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.