Adaptive neuromorphic sensor-fusion for low-power and low-latency applications

Current autonomous systems are equipped with several sensors which are sensing the environment continuously, actively or passively, and generating an awful lot of data. While most of the traditional processing until now has been taking place in the cloud or edge-cloud platforms due to the amount of computation and caching resources required for these data, one of the milestones of recent research is to enable fusion at or near the sensors. The challenge is to devise both algorithms and as well as compute capabilities that can match and cope with the data generation rates of the sensors, while at the same time keeping the energy footprint and latency low.

Under the neuromorphic computing paradigm, multi-modal sensors operate with asynchronous sampling rates. Special bio-, chemistry- and nature- inspired algorithms not only promise to effectively align the sensor outputs but also compute based on them efficiently under tight resource constraints. There is a lot of space for a holistic exploration of application-level optimization either with hardware in the loop or by considering available compute resources in optimize and improve performance metrics like power, latency, cost, and of course application efficacy. This project aims to explore, design, optimize, and benchmark such sensor fusion algorithms for neuromorphic processing platforms.