Energy-Efficient Neural Architecture Search for Embedded Intelligent Systems

Abstract

Neural architecture search (NAS) began as an extravagantly expensive alternative to hand-designed networks, with the
earliest reinforcement-learning and evolutionary methods consuming thousands of GPU-days to find a single competitive
architecture. This article synthesizes a decade of subsequent work along two efficiency axes that the phrase energy-efficient
NAS names jointly: making the search process itself cheap, through weight sharing and differentiable relaxation, and making
the architectures it discovers cheap to run once deployed, by incorporating measured latency, energy, and memory footprint
directly into the search objective. We trace this evolution from NASNet and DARTS through hardware-aware,
latency-constrained methods including MnasNet, ProxylessNAS, FBNet, and Once-for-All, before turning to the
microcontroller-class embedded regime, where MCUNet's joint design of search space and inference engine has enabled the
first sub-megabyte, ImageNet-capable models to run entirely on commercial microcontrollers. We situate this work alongside
the complementary literature on pruning, quantization, and knowledge distillation with which it is routinely combined in
deployed systems. Read together, the evidence shows search cost falling by three to four orders of magnitude while deployment
efficiency gains compound across the pipeline, though genuinely energy-targeted, as opposed to latency-targeted, search
objectives and on-device adaptation after deployment remain substantially open problems.