DQN Based Exit Selection in Multi-Exit Deep Neural Networks for Applications Targeting Situation Awareness
Abhishek Vashist, Sharan Vidash Vidya Shanmugham, Amlan Ganguly, Sai Manoj Pudukotai Dinakarrao · 2022
Smart infrastructure targeting situation awareness for first responders enables intelligent and faster response time. Such an application requires an edge device to process information. The edge hardware is limited by computation capability for running Deep Neural Networks (DNNs). Multi-exit DNNs are used to provide static exit selection for faster inference. We propose using Deep Q Network (DQN) based technique for dynamic exit selection in multi-exit DNNs. The DQN learns an optimal policy using hardware and multi-exit DNN based state information. Our system achieves performance improvement with 63.5 % decrease in inference time, 33 % reduction in energy and classifies 2.2 x more inputs.