CAUDA: Cloud-assisted Active Unsupervised Domain Adaptation for On-device Learning

The recent developments in mobile computing have facilitated the application of on-device machine learning on edge devices. However, complex heterogeneous environmental factors such as network conditions, spatial locations, and device types, can introduce significant unlabeled domain shift, leading to performance degradation in the real-world deployment of on-device learning. In this work, we design CAUDA, a Cloud-assisted Active Unsupervised Domain Adaptation framework for on-device machine learning, to address unlabeled domain shift issue in mobile computing. We propose active filtering method to perform data valuation and select hard-to-adapt samples to label with theoretical guarantees during adaptation. To targetedly process the filtered hard-to-adapt data, we leverage prototype contrastive learning to align them from the perspective of metric learning. The cloud assists on-device learning by annotating filtered unlabeled data and generating corresponding prototypes. Extensive experiments on four tasks for two representative mobile computing applications demonstrate the remarkable advantages of CAUDA with up to 21.38% increase in average inference accuracy compared with other unsupervised domain adaptation and active learning methods.