Memristive-based electro-optical neuromorphic hardware takes advantage of both the high-density of electronic circuits and the high bandwidth of their photonic counterparts, thus showing potential for low-power artificial intelligence applications. In this Perspective paper, we introduce a class of electro-optical memristors that can emulate the key properties of synapses and neurons, which are essential features for the realization of electro-optical neuromorphic functionalities. We then describe the challenges associated with existing technologies and finally give our viewpoint on possible developments toward an energy-efficient neuromorphic platform.

The typically nonlinear and asymmetric response of synaptic memristors to positive and negative electrical pulses makes the realization of accurate deep neural networks very challenging. Here, we integrate a two-terminal valence change memory (VCM) into a photonic/plasmonic circuit and show that the switching properties of this memristor become more gradual and symmetric under light irradiation. The added optical input acts on the VCM as a third, independent modulation channel. It locally heats the active area of the device, which enhances the generation of oxygen vacancies and broadens the resulting nanoscale conductive filaments. The measured conductance modulation of the VCM is then inserted into a neural network simulator. Using the MNIST data set of handwritten digits as an application, a light-enhanced recognition accuracy of 93.53% is demonstrated, similar to ideally performing memristors (94.86%) and much higher than those without light (67.37%). Notably, the optical signal does not increase the overall energy consumption by more than 3.2%. Finally, an approach to scale up our electro-optical technology is proposed, which could allow high-density, energy-efficient neuromorphic computing chips.