Reuse Experts in Continuous Learning to Cope with Data Drift

XJTU 2024 CSUC Submission12 Authors

31 Mar 2024 (modified: 03 Apr 2024)XJTU 2024 CSUC SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: video analysis, edge compute, continuous learn- ing, model reuse
Abstract: Continuous learning has become a famous method by retraining light-weighted experts with sampled video frames from each video stream to cope with the data drift in real time video analysis area. However, retraining models consumes a considerable amount of computational resources, affecting video inference and leading to a decline in inference accuracy and throughput. Given that video streams often display temporal or spatial consistency (for instance, vehicles on the same route will pass through the same video scenes; similar video scenes will be brought about by the same lighting and weather in different places), reusing previous expert models on different cameras holds potential.To validate the potential of reusing models, I carried out a comparative experiment between reusing and retraining models on the Cityscapes dataset. It was observed that as the number of cameras increased progressively, model reuse brought about an improvement of around 10 mAP compared to model retraining.
Submission Number: 12
Loading