Enabling autonomous learning and decisioning in tiny devices

Two electronic trends characterize the past decade: 1.) the trend of ubiquitous electronics, in which electronics get more and more integrated into everyday objects around us. 2.) the advent of machine learning (ML), which enables to find patterns and make informed decisions based on data. Next decade will bring the integration of these two trends, denoted by “tinyML”: bringing ML to small electronic devices, ubiquitously present in our environment. Currently, tinyML is restricted to only run inference tasks (recognition, regression, …) in the tiny device, while the heavy training and decision-making tasks are still performed in the cloud. To enable true tinyML, including training and decision making workloads, hardware and algorithms have to be tightly co-optimized. Yet, the totally different innovaton pace of these two domains (fast SW changes and slow HW development) prevents aligned progress. This project aims to overcome this by: 1.) creating a HW-aware ML cost estimation framework, which allows algorithmic designers to quickly assess their innovations in terms of HW and system performance metrics, such as task latency or energy consumption; 2.) building high level synthesis hardware templates for tinyML processors, which allows to quickly adapt these processors to new ML algorithms. This can break the SW/HW barrier towards true tinyML systems, that can locally learn from new data, and actively collect knowledge through reinforcement learning.