The statistical processing of sensor data using conventional digital computers is inefficient in terms of time, energy usage, and communication bandwidth, among others. Therefore, new approaches are sought to create context and make sense of the sensor data using special-purpose computers that excel in specific computation tasks. Herein we discuss the requirements for physical systems to perform sophisticated nonlinear computations needed for real-time pattern recognition in data, specifically sensor data. Our focus is on physical reservoir computing as a neuromorphic computing approach. Considering energy flow as the coupling mechanism between nonlinear dynamic systems, we demonstrate that many physical systems satisfy the basic requirements for building reservoir computers. Using physical reservoir computers brings up exciting opportunities for near- or in-sensor computing as to how new data is collected and processed. We demonstrate the concepts through a novel physical computation platform, where off-the-shelf, temperature-sensitive resistors are employed to perform various standard and specific computational tasks. This platform is used as a near-sensor processor to detect particular events. We further discuss how a similar platform may be used for in-sensor neuromorphic computations.
Member of collection