隨著物聯網 (IoT) 技術與人工智慧 (AI) 應用逐漸成熟，人工智慧物聯網 (AIoT) 便順勢飛快發展。AIoT以「低接觸服務」特質的無人介入之智慧生活應用最為熱門，目前為「無人商店」最貼近一般大眾的消費生活，其技術關鍵乃是基於影像識別與大量具邊緣運算能力的攝影機 (本研究稱為邊緣攝影機) 之結合運用。然而，佈署大量邊緣攝影機非常消耗時間與成本，包括收集有效標籤資料和訓練個別邊緣攝影機的邊緣模型 (edge model) 等。雖然轉移學習 (transfer learning, TL) 可以加快模型訓練，但既有應用於邊緣運算的轉移學習都須仰賴集中式伺服器，無法發揮物聯網邊緣運算應有的效益。因此，本研究提出「多邊緣攝影機之協同轉移學習框架」，讓邊緣攝影機之間能不假借伺服器的協助下，直接互相轉移資訊來建立邊緣模型。為了降低資訊傳輸的成本，本研究提出「輕量化轉移管理員 (lightweight e2e TL manager, Le2eTLM)」負責處理與其他邊緣攝影機的轉移程序，例如，轉移時所需的交換資訊、轉移類型 (transfer learning type, TL-type) 的選擇等。而該管理員首先包含「基於精英樣本相似度比對」，其根據測試資料的特性，本研究透過顏色和形狀相似度比對來篩選具代表性的精英樣本，有效降低邊緣攝影機之間圖片傳輸的網路成本。其次，根據圖片相似度匹配出來源邊緣攝影機後，本研究提出「一對多之邊緣對邊緣轉移技術」，其藉由讓主要一台邊緣攝影機作為多台相似目標邊緣攝影機知識轉移的來源，藉此提高知識或資訊的重用度，並快速建立初始模型。最後，為了進一步提升邊緣攝影機間的資訊交流，來充分發揮多邊緣攝影機環境資料的多元性，本研究更提出「多對一之邊緣對邊緣轉移技術」，讓新加入的邊緣攝影機可以探索更多同儕邊緣攝影機作為潛在的資訊來源，更可降低標籤資料收集的負擔。實驗結果顯示，本研究的「基於精英樣本相似度比對」在無人商店的情境下，能有效選擇知識來源以及訓練樣本提升來準確率，並有效減少邊緣對邊緣轉移所需傳送樣本數平均約70%。其中，「一對多之邊緣對邊緣轉移技術」可以在端對端轉移上比既有研究平均提升0.6%的準確度，且比一對一使用全部樣本之轉移學習者平均提升1.05%的準確度，因此可以更快速且準確建立多台新加入邊緣攝影機的初始模型.。另外，「多對一之邊緣對邊緣轉移技術」也比既有研究平均提升0.75%的準確度，並比一對一使用全部樣本之轉移學習者平均節省樣本數68.6%與平均提升5.95%的準確度。且在二對一轉移上可節省傳送75.4%所需來源樣本數，三對一時則可節省61.9%。因此，本研究提出的方法可以在稍微提升準確率的情況下，在端對端間轉移學習上能更充分重用資訊並降低資料傳送的頻寬需求。

The technologies of the Internet of Things (IoT) and artificial intelligent (AI) have been becoming more mature. The Artificial Intelligence of Things (AIoT) is one of the most popular smart living applications, and it can facilitate low-touch services; therefore, unmanned stores have now closely related to our daily lives. The key technology of the unmanned stores is based on image recognition via a large number of IoT-enabled cameras (one with the ability to leverage edge intelligence, hereafter referred to edge camera). However, it is very time-consuming and costly to deploy edge cameras and to collect labeled data for training a model of each edge camera. Although existing transfer learning can speed up model training, it must depend on a centralized sever’s assistance, thus failing to bring benefits into IoT-enabled edge computing. To address the above issues, our study proposes collaborative transfer learning for training edge cameras without assistance from centralized servers. The core of collaborative transfer learning is lightweight edge-to-edge TL manager, which consists of three key technologies. The first one is the elite-instance based matching, which utilizes color histogram and perceptual hash to filter representative samples for effectively decreasing network communication cost among edge cameras. The second one is one-to-many edge-to-edge transfer learning, which can transfer a selected edge camera's knowledge to multiple targets based on elite-instance based matching. This can increase knowledge reusability on an edge camera to rapidly build its initial model. The last one is many-to-one edge-to-edge transfer learning, which enables an edge camera to reuse multiple source information based on elite-instance based matching, thus decreasing the effort for labeled data collection. The experimental results show that the elite-instance based matching can effectively save 70% source samples on average that need to be transmitted and help the one-to-many edge-to-edge transfer learning to improve the accuracy by 0.6% on average w.r.t. the existing research. It can also improve the accuracy by 1.05% on average comparing with one-to-one transfer learning. Finally, the elite-instance based matching also helps the many-to-one edge-to-edge transfer learning to improve the accuracy by 0.75% on average w.r.t. the existing research. It can also improve the accuracy by 5.95% on average and save 68.6% source samples on average that need to be transmitted on average comparing with one-to-one transfer learning. It can save 75.4% source samples that need to be transmitted under 2-to-1 transfer. It can save 61.9% under 3-to-1 transfer. Therefore, our proposed method can slightly improve accuracy on edge-to-edge transfer learning and reuse the existing information more sufficiently to decrease the requirement of data transmitting bandwidth.

[1] D. Reinsel, J. Gantz, and J. Rydning, "The digitization of the world: from edge to core," Framingham: International Data Corporation, 2018.
[2] G. Ananthanarayanan et al., "Real-time video analytics: The killer app for edge computing," computer, vol. 50, no. 10, pp. 58-67, 2017.
[3] J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu, "Squeeze-and-Excitation Networks," ArXiv e-prints, Accessed on: September 01, 2017Available: https://ui.adsabs.harvard.edu/#abs/2017arXiv170901507H
[4] M. Bilkhu and H. Ayyubi, "Human Activity Recognition for Edge Devices," arXiv preprint arXiv:1903.07563, 2019.
[5] H. Li, K. Ota, and M. Dong, "Learning IoT in edge: Deep learning for the Internet of Things with edge computing," IEEE Network, vol. 32, no. 1, pp. 96-101, 2018.
[6] P. P. Angelov and X. Gu, "Toward Anthropomorphic Machine Learning," Computer, vol. 51, no. 9, pp. 18-27, 2018.
[7] N. Garun, "Amazon just launched a cashier-free convenience store," The Verge.(Retrieved April 17, 2017 from http://www. theverge. com/2016/12/5/13842592/amazongo-new-cashier-less-convenience-store), 2016.
[8] Z. Tao and Q. Li, "esgd: Communication efficient distributed deep learning on the edge," in {USENIX} Workshop on Hot Topics in Edge Computing (HotEdge 18), 2018.
[9] R. Sharma, S. Biookaghazadeh, B. Li, and M. Zhao, "Are existing knowledge transfer techniques effective for deep learning with edge devices?," in 2018 IEEE International Conference on Edge Computing (EDGE), 2018, pp. 42-49: IEEE.
[10] S. J. Pan and Q. Yang, "A Survey on Transfer Learning," IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp. 1345-1359, 2010.
[11] L. Torrey and J. W. Shavlik, "Transfer Learning," 2009.
[12] W. Dai, Q. Yang, G.-R. Xue, and Y. Yu, "Boosting for transfer learning," presented at the Proceedings of the 24th international conference on Machine learning, Corvalis, Oregon, USA, 2007.
[13] N. D. Lawrence and J. C. Platt, "Learning to learn with the informative vector machine," in Proceedings of the twenty-first international conference on Machine learning, 2004, p. 65: ACM.
[14] R. Raina, A. Battle, H. Lee, B. Packer, and A. Y. Ng, "Self-taught learning: transfer learning from unlabeled data," in Proceedings of the 24th international conference on Machine learning, 2007, pp. 759-766: ACM.
[15] L. Mihalkova, T. Huynh, and R. J. Mooney, "Mapping and revising Markov logic networks for transfer learning," in AAAI, 2007, vol. 7, pp. 608-614.
[16] K. Weiss, T. M. Khoshgoftaar, and D. Wang, "A survey of transfer learning," Journal of Big Data, vol. 3, no. 1, p. 9, 2016.
[17] S. Ruder, "An Overview of Multi-Task Learning in Deep Neural Networks," ArXiv e-prints, vol. 1706, Accessed on: June 1, 2017Available: http://adsabs.harvard.edu/abs/2017arXiv170605098R
[18] A. Bettge, R. Roscher, and S. Wenzel, "Deep Self-taught Learning for Remote Sensing Image Classification," CoRR, vol. abs/1710.07096, 2017.
[19] Y. Ganin et al., "Domain-Adversarial Training of Neural Networks," Journal of Machine Learning Research, vol. 17, pp. 59:1-59:35, 2016.
[20] M. Johnson et al., "Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation," TACL, vol. 5, pp. 339-351, 2017.
[21] W. Dai, Q. Yang, G.-R. Xue, and Y. Yu, "Self-taught clustering," in ICML, 2008.
[22] Z. Chen, J. Zhuang, X. Liang, and L. Lin, "Blending-Target Domain Adaptation by Adversarial Meta-Adaptation Networks," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 2248-2257.
[23] W. Ying, Y. Zhang, J. Huang, and Q. Yang, "Transfer learning via learning to transfer," in International Conference on Machine Learning, 2018, pp. 5072-5081.
[24] Q. Wu, X. Zhou, Y. Yan, H. Wu, and H. Min, "Online transfer learning by leveraging multiple source domains," Knowledge and Information Systems, vol. 52, no. 3, pp. 687-707, 2017.
[25] Y. Freund and R. E. Schapire, "A decision-theoretic generalization of on-line learning and an application to boosting," Journal of computer and system sciences, vol. 55, no. 1, pp. 119-139, 1997.
[26] Y. Zhu, F. Zhuang, and D. Wang, "Aligning domain-specific distribution and classifier for cross-domain classification from multiple sources," in Proceedings of the AAAI Conference on Artificial Intelligence, 2019, vol. 33, pp. 5989-5996.
[27] D. Zhang, X. Chen, D. Wang, and J. Shi, "A Survey on Collaborative Deep Learning and Privacy-Preserving," in 2018 IEEE Third International Conference on Data Science in Cyberspace (DSC), 2018, pp. 652-658: IEEE.
[28] K. Lin, S. Wang, and J. Zhou, "Collaborative deep reinforcement learning," arXiv preprint arXiv:1702.05796, 2017.
[29] D. Benditkis, A. Keren, L. Mor-Yosef, T. Avidor, N. Shoham, and N. Tal-Israel, "Distributed deep neural network training on edge devices," in Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019, pp. 304-306: ACM.
[30] J. Konečný, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh, and D. Bacon, "Federated learning: Strategies for improving communication efficiency," arXiv preprint arXiv:1610.05492, 2016.
[31] S. A. Rokni and H. Ghasemzadeh, "Autonomous Training of Activity Recognition Algorithms in Mobile Sensors: A Transfer Learning Approach in Context-Invariant Views," IEEE Transactions on Mobile Computing, 2018.
[32] T. Xing, S. S. Sandha, B. Balaji, S. Chakraborty, and M. Srivastava, "Enabling Edge Devices that Learn from Each Other: Cross Modal Training for Activity Recognition," in Proceedings of the 1st International Workshop on Edge Systems, Analytics and Networking, 2018, pp. 37-42: ACM.
[33] O. Valery, P. Liu, and J. Wu, "CPU/GPU Collaboration Techniques for Transfer Learning on Mobile Devices," in 2017 IEEE 23rd International Conference on Parallel and Distributed Systems (ICPADS), 2017, pp. 477-484.
[34] H. Daga, P. K. Nicholson, A. Gavrilovska, and D. Lugones, "Cartel: A System for Collaborative Transfer Learning at the Edge," in Proceedings of the ACM Symposium on Cloud Computing, 2019, pp. 25-37: ACM.
[35] M. Ghifary, W. B. Kleijn, and M. Zhang, "Domain adaptive neural networks for object recognition," in Pacific Rim international conference on artificial intelligence, 2014, pp. 898-904: Springer.
[36] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell, "Deep domain confusion: Maximizing for domain invariance," arXiv preprint arXiv:1412.3474, 2014.
[37] M. Long, Y. Cao, J. Wang, and M. I. Jordan, "Learning transferable features with deep adaptation networks," arXiv preprint arXiv:1502.02791, 2015.
[38] H. Yu, M. Hu, and S. Chen, "Multi-target unsupervised domain adaptation without exactly shared categories," arXiv preprint arXiv:1809.00852, 2018.
[39] J. Li, W. Wu, D. Xue, and P. Gao, "Multi-Source Deep Transfer Neural Network Algorithm," Sensors, vol. 19, no. 18, p. 3992, 2019.
[40] M. J. Swain and D. H. Ballard, "Color indexing," International journal of computer vision, vol. 7, no. 1, pp. 11-32, 1991.
[41] C. Zauner, "Implementation and benchmarking of perceptual image hash functions," 2010.
[42] M. Norouzi, D. J. Fleet, and R. R. Salakhutdinov, "Hamming distance metric learning," in Advances in neural information processing systems, 2012, pp. 1061-1069.
[43] Z. Wang, "The SSIM index for image quality assessment," https://ece. uwaterloo. ca/~ z70wang/research/ssim, 2003.
[44] S. Zagoruyko and N. Komodakis, "Learning to compare image patches via convolutional neural networks," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 4353-4361.
[45] P. Chamoso, A. Rivas, J. J. Martín-Limorti, and S. Rodríguez, "A hash based image matching algorithm for social networks," in International Conference on Practical Applications of Agents and Multi-Agent Systems, 2017, pp. 183-190: Springer.
[46] S. Chapman, "Google rolls out reverse image search: RIP TinEye," ed, 2011.
[47] K. R. Rao and P. Yip, Discrete cosine transform: algorithms, advantages, applications. Academic press, 2014.
[48] H. Shimodaira, "Improving predictive inference under covariate shift by weighting the log-likelihood function," Journal of statistical planning and inference, vol. 90, no. 2, pp. 227-244, 2000.
[49] J. Huang, A. Gretton, K. Borgwardt, B. Schölkopf, and A. J. Smola, "Correcting sample selection bias by unlabeled data," in Advances in neural information processing systems, 2007, pp. 601-608.
[50] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, "Domain adaptation via transfer component analysis," IEEE Transactions on Neural Networks, vol. 22, no. 2, pp. 199-210, 2010.
[51] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770-778.
[52] C. Dwork, F. McSherry, K. Nissim, and A. Smith, "Calibrating noise to sensitivity in private data analysis," in Theory of cryptography conference, 2006, pp. 265-284: Springer.
[53] R. Xu, Z. Chen, W. Zuo, J. Yan, and L. Lin, "Deep cocktail network: Multi-source unsupervised domain adaptation with category shift," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 3964-3973.
[54] I. Hussain, Q. He, and Z. Chen, "AUTOMATIC FRUIT RECOGNITION BASED ON DCNN FOR COMMERCIAL SOURCE TRACE SYSTEM."
[55] D. P. Kingma and J. Ba, "Adam: A method for stochastic optimization," arXiv preprint arXiv:1412.6980, 2014.
[56] Z. Zhang and M. Sabuncu, "Generalized cross entropy loss for training deep neural networks with noisy labels," in Advances in neural information processing systems, 2018, pp. 8778-8788.