Characterizing the Execution of Deep Neural Networks on Collaborative Robots and Edge Devices
Machine Learning/Artificial Intelligence
TimeWednesday, July 312:30pm - 3pm
DescriptionIn this paper, we examined and analyzed the effects of executing deep neural networks collaboratively on edge devices. Our method uses model parallelism to distribute tasks from DNN layers to Raspberry Pis, which are mounted on and powered by iRobots. Our measurements showed that while adding additional devices to a network of edge devices increases static energy, it decreases the dynamic energy used by each device. It also increases the average number of inferences performed by each device. We found that execution of DNNs on a distributed robot system can lead to unpredictable power consumption, which may in turn worsen Raspberry Pi performance. Another unreliable aspect is the network latency, which increases and decreases often based on a variety of often uncontrollable factors. Finally, through our measurement of the Raspberry Pi temperature during DNN execution, we found that DNN computation can increase temperature by as much as 16 degrees Celsius.