Neural-Architecture-Search

In this project, I explored neural architecture search (NAS) methods to design efficient neural networks for microcontrollers, using techniques adapted from MIT HAN Lab. The assignment involved deploying Once-for-All (OFA) NAS, enabling me to search within a large "super network" to find optimized sub-networks that meet performance requirements without retraining. The design space included varying depths, kernel sizes, and other network parameters.

By leveraging both accuracy and efficiency predictors, I rapidly estimated the effectiveness of potential network configurations, minimizing computational cost. I implemented random search and evolutionary search algorithms to identify architectures that meet real-world constraints (like MACs and memory usage), allowing deployment on microcontrollers with limited resources.

This assignment highlighted the potential of NAS to democratize AI by making deep learning feasible on resource-constrained devices, significantly advancing efficiency-focused model design.