Heterogeneous integration of imprecise memory devices to enable learning from a very small volume of noisy data

The artificial intelligence community, inspired by the tremendous progress made in neuroscience, has recently proposed powerful algorithms to enable effective real-time decision making based on a limited volume of noisy sensory data. However, implementing such algorithms in low-power devices remains a challenge due to the energy inefficiency that comes from separating logic and memory in current electronic systems. For the past 10 years, research groups have been developing alternative electronic components and systems, such as brain inspired computing architectures and novel resistive memory technologies to address this design bottleneck. The critical feature for these new technologies to perform at their best is a very high-density, reliable, non-volatile memory with infinite endurance. This ideal memory does not exist today, and it is unlikely it will ever exist. This project takes inspiration from the insects nervous system. The general aim of DIVERSE is to enable learning from a very limited volume of noisy data based on imperfect, limited density, low endurance, resistive memories. Unlike digital systems, insects are not very good at performing precise calculations, but they excel at making extremely energy-efficient real time decisions by combining sensory data recorded in noisy environments. I thus propose to take inspiration from the well-studied crickets nervous system and to use my experience and skills in resistive memories to develop a new technology that expresses robust cognitive behaviour while interacting with the environment. This cross-disciplinary work will lead to the fabrication of an innovative hardware/software platform with extremely high power efficiency and robust cognitive computing capabilities. This new technology will open new perspectives in dynamically developing areas including service and consumer robotics, implantable medical diagnostic microchips and wearable electronics.