A facile photonics reconfigurable memristor with dynamically allocated neurons and synapses functions

Abstract

The dynamic neural network function realized by reconfigurable memristors to implement artificial neurons and synapses is an effective method to complete the next generation of neuromorphic computing. However, due to the limitation of reconfiguration conditions, there are inconsistencies in the turn-on voltage and operating current before and after the reconfiguration of neuromorphic devices, which leads to huge difficulties in hardware application development and is an urgent problem to be solved. In this work, we introduced light as a regulatory means in the memristor and achieved the reconfiguration of volatile (endurance ~10⁶ cycles) and non-volatile (retention ~10⁴ s) characteristics with a unified working parameter through the photoelectric coupling mode. The switching voltage of the device can be controlled 100% by this method without any limiting current. This will allow neurons and synapses to be dynamically allocated on demand. We completed the verification such as Morse code decoding, Poisson coded image recognition, denoising in the image recognition process, and intelligent traffic signal recognition hardware system under different work modes. It is verified that the device can dynamically adjust the neuromorphic according to needs, providing a new idea for the further integration of neuromorphic computing in the future.