Neuromorphic Visual Receptive Field Hardware with Vertically Integrated Indium-Gallium-Zinc-Oxide Optoelectronic Memristors over Silicon Neuron Transistors.

Abstract

Event-driven processing in neuromorphic vision systems, utilizing spiking neural networks, can offer improved energy efficiency compared to conventional von Neumann systems. This study proposes an artificial retinal neuron with a vertically integrated optoelectronic memristor (optomemristor) and a neuron transistor (neuristor), inspired by the visual receptive field (VRF) of the biological retina. This design performs pre-processing in the sensor to extract essential image features, such as edges. The optomemristor on top consists of an In-Ga-Zn-O thin film, which detects light, while the neuristor at the bottom is made of a Si field-effect transistor (FET), converting the spikes into an electrical signal. The VRF hardware comprises excitatory (ON-type) and inhibitory (OFF-type) cells. The spiking frequency of the Si FET increases in response to light exposure for ON-type cells, which are composed of a serially connected optomemristor and neuristor. In contrast, OFF-type cells, composed of parallelly connected devices, decrease the spiking frequency under light exposure. The dual-type configuration, which incorporates both ON- and OFF-type cells, achieves a remarkable 99.8% accuracy in fingerprint pattern classification due to the efficient extraction of edge information. This represents a significant improvement over the 56.1% accuracy of the single-type configuration that relies solely on ON-type cells.