Echo state neural networks (ESNs) provide an efficient classification technique for spatiotemporal signals. The feedback connections in the ESN enable feature extraction in both spatial and temporal components in time series data. This property has been used in several application domains such as image and video analysis, anomaly detection, and speech recognition. The software implementations of the ESN demonstrated efficiency in processing such applications, and have low design cost and flexibility. However, hardware implementation is necessary for power constrained resources applications such as therapeutic and mobile devices. Moreover, software realization consumes an order or more power compared to the hardware realization. In this work, a hardware ESN architecture with neuromemristive system is proposed. A neuromemristive system is a brain inspired computing system that uses memristive devises for synaptic plasticity. The memristive devices in neuromemristive systems have several interesting properties such as small footprint, simple device structure, and most importantly zero static power dissipation. The proposed architecture is reconfigurable for different ESN topologies. 2-D mesh architecture and toroidal networks are exploited in the reservoir layer. The relation between performance of the proposed reservoir architecture and reservoir metrics are analyzed. The proposed architecture is tested on a suite of medical and human computer interaction applications. The benchmark suite includes epileptic seizure detection, speech emotion recognition, and electromyography (EMG) based finger motion recognition. The proposed ESN architecture demonstrated an accuracy of $90\%$, $96\%$, and $84\%$ for epileptic seizure detection, speech emotion recognition and EMG prosthetic fingers control respectively.
Library of Congress Subject Headings
Neural networks (Computer science); Machine learning; Memristors
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Saleh, Qutaiba Mohammed, "Design of a Neuromemristive Echo State Network Architecture" (2015). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus