Neuromorphic Computing for Sensing and Control

Neuromorphic engineering uses spiking neural networks (SNNs) to enable new algorithms, devices, architecture and applications allowing for low power computing with low latency, when the demand for accuracy is not significantly high. Despite recent demand of decentralized computing and the rise of neuromorphic architecture even in mainstream technical companies, there is inadequate evidence and benchmarking available to test the true possibilities and limits of these applications. Some areas of relevance in this regard include anomaly detection and control. Our project investigates design and testing of neuromorphic architectures for (a) sudden changes in SNN-based filtered sensor data for a leak introduced in a pipe with water flow and related power consumption comparison among various neuromorphic hardware, and (b) time evolution plots for the pole angle of a cartpole balancing task using a single spiking neuron and an ensemble of spiking neurons. Spiking neurons were observed to perform signal representation using significantly lower power than its non-spiking counterparts. It was also observed that control can be reasonably achieved with spiking neurons. Precision increases with the number of neurons and a future comparisons with conventional control methods can provide boundaries of operation and usefulness of a neuromorphic approach.