We introduce ENAS, a hardware‑aware Neural Architecture Search (NAS) framework. It combines a static feasibility check, a cell‑based search space that supports standard, depthwise‑separable, and bottleneck blocks with optional skip connections, and a three‑stage hybrid search strategy (random → top‑K → mutation) equipped with persistent cross‑run caching. Unlike many NAS systems that rely on GPU acceleration, ENAS is designed to run efficiently without GPUs, making it suitable for resource‑constrained development environments.
We evaluate ENAS on two TinyML benchmarks—Visual Wake Words and Melanoma Cancer—across eight microcontrollers (20 KB to 1 MB SRAM) and nine input image resolutions. Results show mean search‑time speedups of $2.41\times$ on Visual Wake Words and $1.70\times$ on Melanoma Cancer, while maintaining test accuracy comparable to the recent NanoNAS framework. A resource analysis reveals that ENAS‑selected models use substantially lower peak activation RAM, the primary constraint for microcontroller deployment at matched accuracy.
On an STM32H743‑based microcontroller, ENAS achieves $79.4\%$ test accuracy, outperforming the greedy CPU‑only baseline by $2.6$ percentage points. The framework is released as open‑source at https://github.com/EdgeIntelligenceLab/ENAS.
Review: ENAS’s hardware‑aware search space and efficient three‑stage strategy significantly reduce search time and memory footprint without sacrificing accuracy, offering a practical solution for TinyML practitioners.