Designing and deploying Artificial Neural Networks in super-integrated sensors represents enormous challenges in terms of computing and storage capacities, imposed by the physical limitations on available memory and silicon space, with a power envelope on the scale of microwatts. These constraints are two orders of magnitude less than typical microcontroller requirements. While Deeply Quantized Neural Networks (DQNNs) offer the most promising answer to in-sensor artificial intelligence, the current state-of-the-art is not yet capable of incorporating the inherent advantages of DQNN in sensors. The most viable approach to neural computing in sensors in terms of programmability in C, area and energy efficiency is therefore the design of reduced instruction set programmable processors. The resulting quantized AI sensor is able to support full precision to 1-bit neural networks without compromising the accurate classification of common human activities or the detection of anomalies from inertial data. We call this innovation the Intelligent Sensor Processing unit, or ISPU, with optimized ultra-low power hardware circuitry for real-time Hybrid Precision Neural Network (HPNN) execution, embodying STRED DSP on 130µm CMOS technology, presented here together with associated demo results.”
Simone Ferri is Director of the Consumer MEMS Business Unit in STMicroelectronics’ MEMS Sensor Division and has held this position since February 2016.Ferri joined STMicroelectronics in 1999 as an engineer working in Central R&D. After 2 years, he moved to the Audio Division as a digital designer, then into product management 5 years later. In 2014, Ferri added responsibility for all ST’s MEMS consumer sensors, and more recently for all MEMS-sensor related Marketing and Application activities globally across markets and segments.Simone Ferri was born in Milan, Italy, in 1972, and graduated with a degree in Microelectronics from the Polytechnic of Milan. He also completed his MBA at the Polytechnic of Milan.