(An) efficient retraining scheme for split learning based VA in edge computing

Abstract

In recent years, edge video analytics (VA) has emerged as a technology that attracts attention. Recent researches have reported the data drift issue in edge video analytics. Continuous retraining with acquired new data has been suggested as a solution to the data drift issue. Despite its need, there hasn’t been a retrainable system for an edge computing environment with low communication cost and practical retrain time. We propose an efficient split retraining scheme for edge video analytics that can improve accuracy by selecting the best split point and retraining the partial model in the edge server. We evaluated and compared our scheme with other baselines as a proof-of-concept. While keeping inference time limit, our scheme shows better retrain accuracy, and at best, its retrain accuracy growth is 1.96%p higher than baseline.