由於人工智慧的崛起，深度學習在近年來非常熱門，成為實現 AI 的主流方法。大多的 AI 應用仰賴足夠的訓練資料與監督式學習方法。然而，若考量到永續的應用模式，傳統監督式學習模式存在的巨大的隱患，模型在訓練完資料以後，如果再讓模型訓練新的資料，此時會發生模型遺忘過去訓練過的資料，這種現象被稱作「災難性遺忘」。該現象也成為持續性學習 (Continual Learning) 的核心問題，因此持續性學習的研究大多在減緩或避免該現象的發生。持續性學習毫無疑問是未來人工智慧的重要技術，讓機器能夠更像人類，擁有終身學習的能力。
　　本研究為應用偽標籤 (Pseudo-labeling)、知識蒸餾 (Knowledge distillation)、與生成重放 (Generative replay) 三種持續性學習技巧於物件偵測的議題。我們提出兩種情境，分別為類別上的持續性學習情境與標籤上的持續性學習，並依序在VOC2007與BreCaHAD兩個資料集上進行驗證。在VOC2007實驗中，我們將任務中每一類的GAN模型訓練好，之後先僅使用偽標籤與知識蒸餾兩種技巧進行訓練，再對mAP@.5低於0.5的各個類別分別加入40~60張由GAN生成出的生成影像與其偽標籤一同進行訓練，實現生成重放。同樣地，在BreCaHAD實驗中，我們先使用全體訓練影像訓練出一個GAN模型，並生成約120張混合影像，再使用前一個任務的模型標註偽標籤，之後加入這些影像到各個任務的資料集中進行訓練。這種利用 GAN 生成的偽影像來提升訓練品質屬於生成重放的技巧。實驗結果表明，在50個epochs的VOC2007任務四取得58.2% mAP@.5的表現，在2000個epochs的BreCaHAD任務三取得52.1% mAP@.5的表現，證實了結合生成重放與偽標籤的組合在持續性學習物件偵測上的巨大優勢。

With the rise of artificial intelligence (AI), deep learning has become extremely popular in recent years, and become the primary approach to achieve AI. Most AI applications rely on an adequate amount of training data and supervised learning methods. However, considering sustainable applications, the traditional supervised learning paradigm would lead to a significant issue known as "catastrophic forgetting." This phenomenon occurs when a model that was trained on certain data forgets the previously learned data after training with new data. This issue is the core problem in Continual Learning (CL) research. Much CL research aims to mitigate or avoid this phenomenon. The CL is undoubtedly an important technology for the future of artificial intelligence, enabling machines to become more human-like with the ability for lifelong learning.
　　This study applies three CL techniques, namely Pseudo-labeling, Knowledge Distillation, and Generative Replay, to the problem of object detection. We propose two CL scenarios: class-updating scenario and label-updating scenario, and validate them on the VOC2007 and BreCaHAD datasets, respectively. In the VOC2007 experiment, we first train GAN models for all classes of the task. Then, we initially use only Pseudo-labeling and Knowledge distillation techniques for training. Subsequently, for each class with mAP@.5 lower than 0.5, we incorporate 40 to 60 GAN-generated fake images with their pseudo-labels for further training, to fulfill generative replay. Similarly, in the BreCaHAD experiment, we first train a GAN model on all training images and generate around 120 mixed images for generative replay purpose, and generate their pseudo-labels by using the previous model. Later, these images are incorporated with the datasets of respective tasks for training. Experimental results demonstrate that in the Task4 of VOC2007, after 50 epochs, a performance of 58.2% mAP@.5 is achieved, and in the Task3 of BreCaHAD, after 2000 epochs, a performance of 52.1% mAP@.5 is attained. This confirms the substantial advantages and potentials of using combining Generative replay and Pseudo-labeling simultaneously in CL for object detection.

論文審定書 i
摘要 ii
Abstract iii
目錄 v
圖次 vii
表次 ix
第一章 緒論 1
1.1　研究背景 1
1.2　研究動機與目的 2
1.3　本文貢獻 3
1.4　論文大綱 3
第二章 文獻探討 4
2.1　持續性學習 4
2.1.1　災難性遺忘 6
2.1.2　知識蒸餾 8
2.1.3　偽標籤 9
2.1.4　生成重放 10
2.2　Yolo系列 11
2.3　GAN 15
2.4　醫學影像 17
第三章 實驗方法與分析 19
3.1　持續性學習實驗情境-1 19
3.1.1　VOC2007資料集 20
3.1.2　實驗流程圖-1 22
3.2　持續性學習實驗情境-2 23
3.2.1　BreCaHad資料集 24
3.2.2　實驗流程圖-2 25
3.3　YOLOv7模型 26
3.4　StyleGAN2模型 30
3.5　實驗設定 33
3.6　評估指標 34
第四章 實驗結果 36
4.1　實驗環境 36
4.2　GAN生成影像展示 37
4.3　BreCaHAD生成混合影像展示 40
4.4　實驗結果分析-VOC2007 41
4.5　實驗結果分析-BreCaHAD 48
4.6　ablation study 52
第五章 結論與未來展望 56
參考文獻 58

參考文獻
[1] B. Liu, "Lifelong machine learning: a paradigm for continuous learning," Frontiers of Computer Science, vol. 11, pp. 359-361, 2017.
[2] S. Thrun and T. M. Mitchell, "Lifelong robot learning," Robotics and autonomous systems, vol. 15, no. 1-2, pp. 25-46, 1995.
[3] G. M. Van de Ven and A. S. Tolias, "Three scenarios for continual learning," arXiv preprint arXiv:1904.07734, 2019.
[4] S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, "icarl: Incremental classifier and representation learning," in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2017, pp. 2001-2010.
[5] F. M. Castro, M. J. Marín-Jiménez, N. Guil, C. Schmid, and K. Alahari, "End-to-end incremental learning," in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 233-248.
[6] Y. Wu et al., "Large scale incremental learning," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 374-382.
[7] C. V. Nguyen, Y. Li, T. D. Bui, and R. E. Turner, "Variational continual learning," arXiv preprint arXiv:1710.10628, 2017.
[8] H. Shin, J. K. Lee, J. Kim, and J. Kim, "Continual learning with deep generative replay," Advances in neural information processing systems, vol. 30, 2017.
[9] P. Singh, P. Mazumder, P. Rai, and V. P. Namboodiri, "Rectification-based knowledge retention for continual learning," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 15282-15291.
[10] S. Yan, J. Xie, and X. He, "Der: Dynamically expandable representation for class incremental learning," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 3014-3023.
[11] C. Buciluǎ, R. Caruana, and A. Niculescu-Mizil, "Model compression," in Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, 2006, pp. 535-541.
[12] G. Hinton, O. Vinyals, and J. Dean, "Distilling the knowledge in a neural network," arXiv preprint arXiv:1503.02531, 2015.
[13] D.-H. Lee, "Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks," in Workshop on challenges in representation learning, ICML, 2013, vol. 3, no. 2, p. 896.
[14] Y. Teng, A. Choromanska, M. Campbell, S. Lu, P. Ram, and L. Horesh, "Overcoming Catastrophic Forgetting via Direction-Constrained Optimization," in European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, 2022.
[15] N. Kamra, U. Gupta, and Y. Liu, "Deep generative dual memory network for continual learning," arXiv preprint arXiv:1710.10368, 2017.
[16] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, "You only look once: Unified, real-time object detection," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779-788.
[17] C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7464-7475.
[18] I. Goodfellow et al., "Generative adversarial networks," Communications of the ACM, vol. 63, no. 11, pp. 139-144, 2020.
[19] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, "Image-to-image translation with conditional adversarial networks," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1125-1134.
[20] T. Karras, S. Laine, and T. Aila, "A style-based generator architecture for generative adversarial networks," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4401-4410.
[21] W.-C. Kuo, "非侵入式生醫斷層影像簡介," 物理雙月刊, vol. 28, no. 4, pp. 698-703, 2006.
[22] R. Weissleder, "Molecular imaging: exploring the next frontier," Radiology, vol. 212, no. 3, pp. 609-614, 1999.
[23] M. L. Mather and C. Baldock, "Ultrasound tomography imaging of radiation dose distributions in polymer gel dosimeters: Preliminary study," Medical physics, vol. 30, no. 8, pp. 2140-2148, 2003.
[24] D. G. Gadian, "NMR and its applications to living systems," (No Title), 1995.
[25] J. Li et al., "Signet ring cell detection with a semi-supervised learning framework," in Information Processing in Medical Imaging: 26th International Conference, IPMI 2019, Hong Kong, China, June 2–7, 2019, Proceedings 26, 2019: Springer, pp. 842-854.
[26] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, "The pascal visual object classes (voc) challenge," International journal of computer vision, vol. 88, pp. 303-338, 2010.
[27] http://host.robots.ox.ac.uk/pascal/VOC/voc2007/
[28] A. Aksac, D. J. Demetrick, T. Ozyer, and R. Alhajj, "BreCaHAD: a dataset for breast cancer histopathological annotation and diagnosis," BMC research notes, vol. 12, no. 1, pp. 1-3, 2019.
[29] R. Mehta and C. Ozturk, "Object detection at 200 frames per second," in Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018, pp. 0-0.
[30] T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, "Analyzing and improving the image quality of stylegan," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8110-8119.
[31] T. Karras, T. Aila, S. Laine, and J. Lehtinen, "Progressive growing of gans for improved quality, stability, and variation," arXiv preprint arXiv:1710.10196, 2017.