Scenario Adaptive Edge Data Reduction

Abstract

It becomes common to deploy a pre-trained machine learning model on the edge devices to improve their intelligence. Considering the dynamic nature of the edge environment, for ensuring decision accuracy, edge devices always need to collect the latest samples and upload them to the cloud to get an updated model. During this process, it is crucial to determine which samples are necessary to be uploaded considering the communication cost. We propose a scenario adaptive edge data reduction strategy to filter samples differently from existing approaches by measuring whether they can affect current decision accuracy. First, we put forward a novel adaptive data reduction framework for cloud-edge collaborative scenarios. Then, we present the implementation algorithms for filtering samples based on scenarios, identifying candidate scenarios emerging in edge environments, and updating edge scenarios. Experiment results show that in the best case, our approach can discard 70% samples while keeping the same inference accuracy with the original sample set.