隨著物聯網快速的發展，大量的攝影機進行即時且廣域的監控將產生龐大的串流資料，因此機器學習結合邊緣計算之相關研究議題也更受重視。然而，目前既有的即時廣域監控上，通常未考量串流資料本身的適切性，特別是使用者之隱私。其次，串流影像在受到各種不確定因素導致概念飄移 (concept drift) 發生時，模型無法有效即時調適，導致準確率逐漸下降。再者，因串流資料無法立即取得其真實標籤，導致系統上線後模型無法因應概念飄移進行可靠的修正。最後，面臨新場域佈署時，若無法重用既有場域所學的知識，將導致無法快速地進行跨場域佈署。因此，本研究提出基於人類教育學的自我調整學習 (self-regulated learning) 之線上模型學習框架，改善既有機器學習在前述問題上的不足，且藉由邊緣設備與雲伺服器協同合作來發揮物聯網邊緣計算 (edge computing) 應有的效益。該框架具備習前預思 (forethought)、習間表現監控 (performance) 和習後反思 (reflection) 三個循環要素。首先，本研究透過「習前預思模組」之目標模型 (goal model) 針對學習目標進行前處理，現階段主要是有效降低使用者隱私的疑慮。再來藉由「習間表現監督模組」之主要辨識模型 (performance model)，套用所提出之一般化 (generalized) 後設認知 (meta-cognitive) 學習框架後可讓深度模型即時監控是否發生概念漂移並做出調適。而「習後反思模組」藉由專家模型 (expert model) 在無法得知真實標籤的狀況下協助主要辨識模型進行錯誤修正，並盡可能減少人為介入。為了彰顯自我調整學習在跨領域學習的優勢，本研究亦提出「跨領域知識轉移模組」，其藉由轉移學習將既有場域的資料或訓練好之模型在新場域盡可能重用，藉此加快系統的佈署。實驗結果顯示，「習前預思模組」可在對人物或是整張影像進行模糊化處理，但準確率可與模糊化前相當。而「習間表現監督模組」透過所提出之一般化後設認知學習框架讓YOLOv2深度模型能在主動偵測到概念飄移後能再透過知識蒸餾 (knowledge distillation) 來有效降低模型調適訓練時間達38%。「習後反思模組」之專家模型可將原本需要人為介入來標註資料的負擔額外降低約一半且能維持系統的準確率。最後，「跨領域知識轉移模組」藉由轉移學習跨場域模型可在相同訓練迭代次數下提高新場域的模型辨識的準確率達37%。

With the development of Interet of Things (IoT), real-time wide-area monitoring through a large number of cameras will generate a plethora of streaming data, so machine learning incorporating edge computing has attracted increasing attention. However, most prior machine learning for real-time wide-area monitorning ignored setting a learning goal, e.g., protecting user privacy. Next, when any concept drift or uncertainty occurs in the streaming data, the existing model to process the data often cannot effectively selfadapt, thus gradually compromising model performance. In addition, since streaming images do not have ground-truth labels, their online models cannot be reliably corrected once any concept drift occurs. Furthmore, a new wide-area monitorning system cannot be quickly deployed due to failing to reuse previously learned knowledge from the existing systems. To address above issues, this study incorporates human’s self-regulated learning. The study also uses edge devices or cameras to collaborate with cloud servers for balancing computation burdens. There are three elements in self-regulated learning, including Forethought, Performance, and Reflection, and this study proposes their couterparts in marchine learning. For the Forethought stage (a.k.a. mForethought), a goal model can be trained to effectively reduce user privacy concerns in streaming images, which is the preset goal for the demonstration system. Next, a performance model empowered by the proposed generalized meta-cognitive based learning framework for the Performance stage (a.k.a. m Performance) can monitor the occurrence of concept drift for later online model adaption. The Reflection stage (a.k.a. mReflection) uses an expert model to correct mistakes and effectively reduce human interventions given no ground-truth can be obtained. Finally, we propose cross-field knowledge transfer, which uses transfer learning to leverage the exiting data or learned knowledge for rapidly adapting to a new field. With the above enhancements, the experiment results show that the goal model of the mForethought can effiectively blur the detected pedestrians or the whole image yet at the cost of slightly compromsing precision to ensure user privacy. The generalized meta-cognitive based learning framework of the mPerformance enhanced a regular YOLOv2 deep network with the cognitive ability to detect concept drift, the we used knowledge distillation to reduce time of model adaptation by 38% even using blurred perdstrians or images. The expert model of the mReflection reduces about 50% of human intervention for labeling data, meanwhile maintaining as much precsion. The cross-field knowledge transfer stage uses transfer learning to improve model precision of a new field by 37% under the same training iteration for a new field.

[1] G. Ananthanarayanan, P. Bahl, P. Bodík, K. Chintalapudi, M. Philipose, L. Ravindranath, et al., "Real-Time Video Analytics: The Killer App for Edge Computing," Computer, vol. 50, pp. 58-67, 2017.
[2] M. Chiang and T. Zhang, "Fog and IoT: An Overview of Research Opportunities," IEEE Internet of Things Journal, pp. 1-1, 2016.
[3] J. S. Preden, K. Tammemae, A. Jantsch, M. Leier, A. Riid, and E. Calis, "The Benefits of Self-Awareness and Attention in Fog and Mist Computing," Computer, vol. 48, pp. 37-45, Jul 2015.
[4] Y. Chen, X. Yang, B. Zhong, S. Pan, D. Chen, and H. Zhang, "CNNTracker: online discriminative object tracking via deep convolutional neural network," Applied Soft Computing, vol. 38, pp. 1088-1098, 2016.
[5] J. Gama, I. Žliobaitė, A. Bifet, M. Pechenizkiy, and A. Bouchachia, "A survey on concept drift adaptation," ACM Computing Surveys (CSUR), vol. 46, p. 44, 2014.
[6] K. Kim and L. S. Davis, "Multi-camera tracking and segmentation of occluded people on ground plane using search-guided particle filtering," in European Conference on Computer Vision, 2006, pp. 98-109.
[7] X. Wang, "Intelligent multi-camera video surveillance: A review," Pattern Recognition Letters, vol. 34, pp. 3-19, 2013/01/01/ 2013.
[8] S. K. Shah and A. Bedagkar-Gala, "Person Re-identification in Wide Area Camera Networks," in Wide Area Surveillance: Real-time Motion Detection Systems, V. K. Asari, Ed., ed Berlin, Heidelberg: Springer Berlin Heidelberg, 2014, pp. 137-162.
[9] S. Ojha and S. Sakhare, "Image processing techniques for object tracking in video surveillance- A survey," in 2015 International Conference on Pervasive Computing (ICPC), 2015, pp. 1-6.
[10] Y. Yan, E. Ricci, R. Subramanian, G. Liu, O. Lanz, and N. Sebe, "A Multi-Task Learning Framework for Head Pose Estimation under Target Motion," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, pp. 1070-1083, 2016.
[11] J. Zhu, S. Liao, Z. Lei, and S. Z. Li, "Multi-label convolutional neural network based pedestrian attribute classification," Image and Vision Computing, vol. 58, pp. 224-229, 2017/02/01/ 2017.
[12] H. Li, Y. Li, and F. Porikli, "Deeptrack: Learning discriminative feature representations online for robust visual tracking," IEEE Transactions on Image Processing, vol. 25, pp. 1834-1848, 2016.
[13] J. Fan, W. Xu, Y. Wu, and Y. Gong, "Human Tracking Using Convolutional Neural Networks," IEEE Transactions on Neural Networks, vol. 21, pp. 1610-1623, 2010.
[14] N. Wang and D.-Y. Yeung, "Learning a deep compact image representation for visual tracking," in Advances in neural information processing systems, 2013, pp. 809-817.
[15] N. Chen, Y. Chen, Y. You, H. Ling, P. Liang, and R. Zimmermann, "Dynamic Urban Surveillance Video Stream Processing Using Fog Computing," in 2016 IEEE Second International Conference on Multimedia Big Data (BigMM), 2016, pp. 105-112.
[16] J. Wang, B. Amos, A. Das, P. Pillai, N. Sadeh, and M. Satyanarayanan, "A Scalable and Privacy-Aware IoT Service for Live Video Analytics," in Proceedings of the 8th ACM on Multimedia Systems Conference, 2017, pp. 38-49.
[17] P. Aditya, R. Sen, P. Druschel, S. Joon Oh, R. Benenson, M. Fritz, et al., "I-pic: A platform for privacy-compliant image capture," in Proceedings of the 14th Annual International Conference on Mobile Systems, Applications, and Services, 2016, pp. 235-248.
[18] S. A. Osia, A. S. Shamsabadi, A. Taheri, H. R. Rabiee, and H. Haddadi, "Private and Scalable Personal Data Analytics Using Hybrid Edge-to-Cloud Deep Learning," Computer, vol. 51, pp. 42-49, 2018.
[19] B. J. Zimmerman, "Chapter 2 - Attaining Self-Regulation: A Social Cognitive Perspective A2 - Boekaerts, Monique," in Handbook of Self-Regulation, P. R. Pintrich and M. Zeidner, Eds., ed San Diego: Academic Press, 2000, pp. 13-39.
[20] M. Pratama, G. Zhang, M. J. Er, and S. Anavatti, "An incremental type-2 meta-cognitive extreme learning machine," IEEE transactions on cybernetics, vol. 47, pp. 339-353, 2017.
[21] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell, "Deep domain confusion: Maximizing for domain invariance," arXiv preprint arXiv:1412.3474, 2014.
[22] I. Redko and Y. Bennani, "Kernel alignment for unsupervised transfer learning," in Pattern Recognition (ICPR), 2016 23rd International Conference on, 2016, pp. 525-530.
[23] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, "How transferable are features in deep neural networks?," in Advances in neural information processing systems, 2014, pp. 3320-3328.
[24] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei, "Large-scale video classification with convolutional neural networks," in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2014, pp. 1725-1732.
[25] I. Zliobaite, A. Bifet, B. Pfahringer, and G. Holmes, "Active learning with drifting streaming data," IEEE Trans Neural Netw Learn Syst, vol. 25, pp. 27-39, Jan 2014.
[26] G. Hinton, O. Vinyals, and J. Dean, "Distilling the knowledge in a neural network," arXiv preprint arXiv:1503.02531, 2015.
[27] J. Redmon and A. Farhadi, "YOLO9000: better, faster, stronger," arXiv preprint, 2017.
[28] G. S. Babu, R. Savitha, and S. Suresh, "A projection based learning in Meta-cognitive Radial Basis Function Network for classification problems," in The 2012 International Joint Conference on Neural Networks (IJCNN), 2012, pp. 1-8.
[29] Y. Zhang and M. J. Er, "Sequential active learning using meta-cognitive extreme learning machine," Neurocomputing, vol. 173, pp. 835-844, 2016.
[30] S. J. Pan and Q. Yang, "A Survey on Transfer Learning," IEEE Transactions on Knowledge and Data Engineering, vol. 22, pp. 1345-1359, 2010.
[31] S. Ruder. (2017, June 1, 2017). An Overview of Multi-Task Learning in Deep Neural Networks. ArXiv e-prints 1706. Available: http://adsabs.harvard.edu/abs/2017arXiv170605098R
[32] A. Bettge, R. Roscher, and S. Wenzel, "Deep Self-taught Learning for Remote Sensing Image Classification," CoRR, vol. abs/1710.07096, 2017.
[33] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, et al., "Domain-Adversarial Training of Neural Networks," Journal of Machine Learning Research, vol. 17, pp. 59:1-59:35, 2016.
[34] M. Johnson, M. Schuster, Q. V. Le, M. Krikun, Y. Wu, Z. Chen, et al., "Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation," TACL, vol. 5, pp. 339-351, 2017.
[35] W. Dai, Q. Yang, G.-R. Xue, and Y. Yu, "Self-taught clustering," presented at the Proceedings of the 25th international conference on Machine learning, Helsinki, Finland, 2008, pp. 200-207.
[36] Y. Zhan, Y. Chen, Q. Zhang, and X. Kang, "Image Forensics Based on Transfer Learning and Convolutional Neural Network," presented at the Proceedings of the 5th ACM Workshop on Information Hiding and Multimedia Security, Philadelphia, Pennsylvania, USA, 2017, pp. 165-170.
[37] M. Everingham, S. A. Eslami, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, "The pascal visual object classes challenge: A retrospective," International journal of computer vision, vol. 111, pp. 98-136, 2015.
[38] S. Han, H. Mao, and W. J. Dally, "Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding," arXiv preprint arXiv:1510.00149, 2015.
[39] I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio, "Binarized neural networks," in Advances in neural information processing systems, 2016, pp. 4107-4115.
[40] A. Aghasi, A. Abdi, N. Nguyen, and J. Romberg, "Net-trim: Convex pruning of deep neural networks with performance guarantee," in Advances in Neural Information Processing Systems, 2017, pp. 3177-3186.
[41] S. Han, J. Pool, J. Tran, and W. Dally, "Learning both weights and connections for efficient neural network," in Advances in neural information processing systems, 2015, pp. 1135-1143.
[42] C.-H. Lu, C.-H. Yu, C.-R. Chen, and S.-S. Huang, "Goal-Driven On-Line Imbalanced Streaming Data Preprocessing," in IEEE International Conference on Consumer Electronics (IEEE ICCE-TW 2018), Taichung, TAIWAN, 2018 (Accepted).