A high accuracy and robust machine learning network for pattern recognition based on binary RRAM devices

Utilizing the binary RRAM devices, a hardware implemented network based on the modified k-nearest neighbor (KNN) algorithm is proposed for pattern recognition. Regarding to the recognition of the handwritten digits, the hardware network shows brilliant performance with simple training scheme, high tolerance of input noise (up to 40%) and variation (up to 60%), and high recognition accuracy rate (more than 90%). The simulations show that the stable binary rather than multilevel resistance characteristics of RRAM enable the better coupling of device and algorithm for the proposed learning network.