Energy-Constrained Bayesian Neural Architecture Search (EC-BNAS) for Edge AI Deployment

Abstract

Automatic methods will be used to identify an optimal architecture for constrained-resource edge devices. Bayesian optimization methods combined with energy-aware searching will be used to optimize accuracy and all constraints (power consumption, latency, and memory usage). Each time an architectural option is substituted, learning occurs. Significant optimization will occur to find the most efficient design based on the prior relationships. The expected result is a collection of models that will yield compact, high performance designs when implemented onto heterogeneous edge hardware to enable real time performance and operation for IoT sensors and mobile intelligent systems. In conclusion, the EC-BNAS concept will implement a scalable and adaptable framework supporting sustainable and energy-efficient development and deployment of Edge AI systems.