Artificial Electrochromic Synapses for Optical Neuromorphic Computing

光学神经形态计算具有速度快、并行处理、容量大、抗干扰能力强等优点，因此在视觉感知、智能语音系统和生物医学图像处理方面大有可为。然而，对于触觉、嗅觉和听觉等感官来说，从感官输入转换而来的电反应无法输入到光学神经形态计算的框架中。如何实现光电转换是光学神经形态计算面临的巨大挑战。文章提出了一种基于人工电致变色突触的电光转换途径，由普鲁士蓝(Prussian Blue, PB)和氧化钨(Tungsten Oxide, WO3)组成的突触在不同电压下表现出可调的光学透射率，根据电压脉冲下的透射率响应，可以展示典型的突触行为，包括短期记忆、长期记忆和双脉冲易化。然后制作了一个3 × 4的人工电致变色突触阵列，用于显示盲文数字图像。为了进行盲文识别，图像被输入到一个衍射深度神经网络(Diffraction Deep Neural Network, D2NN)，这是一个结合了光学衍射原理和深度学习的光学神经形态计算框架。实验显示盲文数字0~9的识别准确率达到100%。这些结果表明，将人工电致变色突触与光学神经形态计算相结合，为实现低成本、快速、高效的多模态人工智能系统提供了新的方向。Optical neuromorphic computing offers advantages such as high speed, parallel processing, large capacity, and strong resistance to interference, making it promising for visual perception, intelligent speech systems, and biomedical image processing. However, for sensories of tactility, olfactory, and auditory, electrical responses converted from sensory inputs cannot be inputted into frameworks of optical neuromorphic computing. It is a tremendous challenge for optical neuromorphic computing to achieve electrical-optical conversion. In this paper, a pathway of electrical-optical conversion based on artificial electrochromic synapses is proposed. The synapse, consisting of Prussian blue and tungsten oxide, exhibits tunable optical transmittances under different voltages. Based on the transmittance responses under voltage pulses, typical synaptic behaviors, including short-term memory, long-term memory, and paired-pulse facilitation, can be demonstrated. Next, a 3 × 4 array of artificial electrochromic synapses is fabricated to display images of Braille numbers. For Braille recognition, the images are input to a diffractive deep neural network, an optical neuromorphic computing framework combining the principle of optical diffraction with deep learning. The recognition accuracy of Braille numbers 0~9 reaches 100%. The results imply that integrating artificial electrochromic synapses with optical neuromorphic computing provides a new direction for realizing low-cost, fast, and efficient multimodal artificial intelligence systems.