Energy Efficient Learning and Adaptation of Multivariate Time Series Using Neuromorphic Computing

The dynamics of the physical world can be captured by multi-channel time-varying analog signals. From sensing to actuation, to interact with the physical world, IoTs and edge devices must have the ability to detect, classify, and generate patterns in multivariate time series with rich temporal and spatial dynamics. However, limited hardware resources and battery capacity pose significant challenges in information representation and processing. Additionally, the constantly changing environment and mission requirements demands the ability for online learning and adaptation. Inspired by the structure and behavior of biological neural systems, spiking neural network (SNN) models and neuromorphic computing hardware incorporate many energy-efficient features of biological systems making them effective for mobile and edge applications. The neuron and synapse states maintained by membrane potentials provide rich temporal dynamics for pattern detection and generation, making the model ideal for in-memory computing. In this talk I will introduce SNNs and neuromorphic computing techniques for multivariate time series processing. Using neurons modeled as a network of infinite impulse response filters, the SNN can either work as a classifier to detect patterns in the input temporal sequences or as a generator to generate desired temporal sequences. The ability to discern temporal patterns allows for very sparse input representation, where information is encoded by the intervals between spike events. When combined with event-driven computing and communication, such temporal coding results in significant energy savings.