Prototyping Opportunistic Learning in Resource Constrained Mobile Devices
Prototyping Opportunistic Learning in Resource Constrained Mobile Devices
复制标题
DOI：
10.1109/percomworkshops53856.2022.9767493
复制
发表时间：
2022-03
期刊：
影响因子：
--
通讯作者：
Haoxiang Yu;Hsiao-Yuan Chen;Sangsu Lee;Xi Zheng;C. Julien
中科院分区：
文献类型：
--
作者：
Haoxiang Yu;Hsiao-Yuan Chen;Sangsu Lee;Xi Zheng;C. Julien
随着普适计算设备能力的不断增强，在设备上训练机器学习模型已经成为可能。同时，对增加用户隐私和减少通信开销的需求使去中心化机器学习成为人们关注的焦点。在这些范例中，单个设备机会性地协作，使用本地可用的数据来训练模型。在本文中，我们考察了这种机会学习的实际可行性。在基本的机会学习方法中，当一个设备（学习者）遇到另一个设备（邻居）时，它可以请求邻居使用邻居自己的本地数据代表学习者执行训练。为了在现实世界中实现机会学习，必须解决两个挑战：(1)利用设备到设备的通信来发现邻居和交换模型；(2)在受资源和延迟约束的设备上训练模型。在本文中，我们使用不同iOS设备的小型网络测试平台，研究了实现机会学习来学习卷积神经网络（CNN）模型用于图像分类任务的可行性。我们展示了在执行完全分散的方法方面取得的成功，并描述了未来的挑战和机遇。
With the increasing capabilities of pervasive computing devices, training machine learning models on-device has become feasible. At the same time, demands for increased user privacy and reduced communication overhead have brought decentralized machine learning to the forefront. In these paradigms, individual devices collaborate opportunistically to train models using locally available data. In this paper, we examine the practical feasibility of such opportunistic learning. In a basic opportunistic learning approach, when a device (the learner) encounters another device (the neighbor), it can request the neighbor to perform training on the learner’s behalf using the neighbor’s own local data. To realize an opportunistic learning in the real world, one must solve two challenges: (1) leveraging device-to-device communication to discover neighbors and exchange models and (2) training models on the device subject to resource and latency constraints. In this paper, we examine the feasibility of implementing opportunistic learning to learn a convolutional neural network (CNN) model for an image classification task using a small network testbed of diverse iOS devices. We demonstrate success in implementing a completely decentralized approach and characterize the challenges and opportunities that lie ahead.