Handwritten numeral recognition utilizing optoelectronic reservoir computing with double reservoir layers

Abstract

Optoelectronic reservoir computing (RC) is a supervised training algorithm implanted in an optoelectronic time-delay system, which possesses simple structure and can be utilized to realize pattern recognition. In this work, based on double reservoir layers composed of two Mach-Zehnder modulators (MZMs), a novel optoelectronic RC system is proposed and the system performances for processing handwritten numeral recognition (HNR) are analyzed. For such a system, a masked handwritten numeral information is injected into the first reservoir layer, the different value between two adjacent node states of the first reservoir layer is sent to the second reservoir layer, and the virtual node states of the second reservoir layer are extracted for training and testing. The simulated results show that, by optimizing the system parameters, a word error rate (WER) of 0.11 for processing HNR can be achieved. By comparing with an optoelectronic RC with a single reservoir layer, the optoelectronic RC with two reservoir layers possesses better performances for processing HNR.