CREST Revolutionizes Neural Architecture Search for Embedded Systems

Joseph Q. Zales and his team have introduced CREST (Cross-platform Runtime Evaluation and Search Tool), a novel Hardware-in-the-Loop (HIL) Neural Architecture Search (NAS) framework designed specifically for low-power microcontroller (MCU) applications in embedded sensing systems. By addressing the limitations of existing workflows, which often rely on static proxy costs, CREST enables a more realistic evaluation of model architectures under tight memory, latency, and energy constraints. The framework allows for the optimization of various configurable factors, including workload, model family, target backend, and scheduling, ensuring that deployment effects are experimentally separable within a reusable workflow.

In practical evaluations, CREST has shown significant improvements in energy efficiency. For instance, in inertial odometry applications, it achieved a median per-inference energy reduction of 41.7% compared to traditional FLOPs-based selection methods, while also avoiding infeasible deployments on memory-constrained targets. The framework's ability to adaptively select different architectures based on the specific MCU and deployment schedule further enhances its utility. By jointly optimizing model architecture, platform, runtime schedule, and deployment policy, CREST represents a significant advancement in the field of embedded AI systems, paving the way for more efficient and effective deployments in real-world applications.