Research & Papers

New AI Tool Squeezes Smart Features Into Tiny, Cheap Chips

⚡Your next gadget could think for itself — without phoning home to the cloud.

Deep Dive

Researchers present ENAS, a hardware-aware Neural Architecture Search framework for TinyML that combines a static feasibility check, a cell-based search space supporting standard, depthwise-separable, and bottleneck blocks with optional skip connections, and a three-stage hybrid search (random, then top-K, then mutation) with persistent cross-run caching. Unlike many NAS frameworks that rely on GPU acceleration, ENAS is designed to work efficiently without GPUs. Tested on two TinyML benchmarks — Visual Wake Words and Melanoma Cancer — across eight microcontrollers and nine input image resolutions, it achieved mean search-time speedups of 2.41x and 1.70x respectively, with test accuracy competitive against the recent NanoNAS framework. ENAS-selected models also used substantially lower peak activation RAM, the binding constraint for microcontroller deployment at matched accuracy. On an STM32H743-based microcontroller it reached 79.4% test accuracy, beating the greedy CPU-only baseline by 2.6 percentage points. The framework is released as open source.

Key Points
  • ENAS finds AI designs that fit on chips with as little as 20 KB of memory — roughly the size of a short email
  • It searches 2.41x faster than the NanoNAS method on person-detection tasks, with no GPU required
  • A model running on a single STM32H743 chip hit 79.4% accuracy on skin-cancer images, beating a simpler baseline by 2.6 points

Why It Matters

Cheaper, private gadgets that work offline — your watch, doorbell or clinic device gets smarter without the cloud.

📬 Get the top 10 AI stories daily