Exploring Machine Learning Topologies at Home with Tiny Constraints for Presence Classification
Simone Tognocchi, Alessandro Tomasoni, Mohammadreza Bakhshizadeh Mohajer, Daniele Lo Iacono, Danilo Pietro Pau · Lecture notes in computer science · 2026
This paper presents a fully automated, edge multiprocessor hardware-aware machine learning topology search for human presence detection using Wi-Fi signals. The proposed method, called WiFiNAS, is a derivative-free, cell-based neural architecture search technique that automatically designs lightweight 1D convolutional neural networks suitable for deployment on microcontrollers with limited memory and computational resources. The process iteratively explores topology configurations by adjusting the number of convolutional filters and computational layers, selecting architectures based on validation accuracy while respecting hardware constraints evaluated each time a topology is generated. This evaluation is performed through an automated code generation tool called unified core technology, which imports and profiles memory usage, operations, and latency on off-the-shelf microcontrollers, feeding this information back into the search process to guide subsequent iterations. The final model achieved a global classification accuracy of 98.57% on a custom WiFi-based presence classification dataset and was post-training quantized to 8-bit integers to reduce deployment costs. The resulting topology demonstrated competitive performance with reduced memory footprint and computational requirements, enabling real-time inference on various STM32 microcontrollers (STM32H7, STM32MP1, and STM32MP2), with or without neural processing acceleration. This work advocates on-premises, hardware-aware neural search and inference techniques for personalized, compact machine learning designs for edge IoT applications, ensuring full privacy by keeping data local, such as within a home environment.