Towards Ubiquitous Learning: A First Measurement of On-Device Training Performance

Abstract

We are witnessing the emergence of ubiquitous learning, where each device (smartphones, wearables, IoTs, etc) can learn from their environments either alone or collaboratively. Such a new paradigm is enabled by deep learning techniques, or more specifically, on-device training. Given its popularity in the machine learning community, unfortunately, there are no systematic understandings of a critical question: how much cost does it take to train typical deep models on commodity end devices? Therefore, this work performs comprehensive measurements of on-device training with the state-of-the-art training library, 6 mobile phones, and 5 classical neural networks. Our measurements report metrics of training time, energy consumption, memory footprint, hardware utilization, and thermal dynamics, thus help reveal a complete landscape of the on-device training performance. The observations from the measurements help guide us to several promising future directions to efficiently enable ubiquitous learning.