Latency-constrained DNN architecture learning for edge systems using zerorized batch normalization

Shuo Huai, Di Liu, Hao Kong, Weichen Liu, Ravi Subramaniam, Christian Makaya, Qian Lin

Abstract: Highlights•Optimize DNN to fulfil the latency constraint and maintain high accuracy.•A one-shot training procedure to avoid the pre-training/re-training cost.•A dynamic Zero-Recovery process to extend the search space for better architecture.•A machine learning latency predictor to avoid time-expensive on-device measurements.