CSR: Small: Energy-efficient Embedded Signal-processing Inference Systems

Princeton University, Princeton NJ

Investigators

Abstract

Machine-learning algorithms enable pattern recognition from data that are too complex to model analytically. This pattern recognition is of fundamental importance in diverse domains. These algorithms are becoming an essential part of embedded systems that find use in infrastructure, environmental monitoring, personal health monitoring, energy management, food supply chain, assembly lines, etc. This research has the potential to enable significant advances in such systems by enabling highly energy-efficient on-sensor inference to be performed. With its plans for involving students from underrepresented groups, industrial engagement, outreach to the broader public, and online distribution of tools, it is expected to have a broad impact. The aim of the proposed work is to explore the energy savings achievable by embedded signal-processing inference systems through random projections. Random projections have previously been employed in the context of compressive sensing to reduce system energy. We have found that when random projections are used to compress Nyquist signals, the compression mechanism is far more robust, while offering the possibility of two orders of magnitude system energy savings. We term this mechanism compressed signal processing. We propose work on bringing this concept to fruition through new methodologies and signal-processing architectures. In addition, we propose the use of genetic programming and error-aware inference to tackle the nonlinear signal-processing problem. We plan extensive evaluations of the system-level energy-accuracy tradeoffs the proposed mechanisms offer.