Sensory Neural Networks for Advanced Photodetectors

这个项目的重点是以单片集成电路的形式实现光电探测器阵列和相关的第一级感觉神经网络，目的是将视野应用于先进光电探测器的工程应用，并产生可用于研究视觉系统中乘法侧抑制的物理模型。光电探测器和电子电路采用砷化镓技术制造，以实现与集成光电子系统的向上兼容，并允许构建的探测器组件在更恶劣的操作环境中运行。此外，光电探测器阵列与第一级神经网络的单片集成在尺寸、可靠性、成本方面具有优势，因此具有商业适用性。生物视觉系统执行许多与电子硬件目前的能力相去甚远的要求苛刻的任务。许多生物系统的成功在很大程度上可以归因于在感觉系统的最外围水平对视觉图像进行很大程度的预处理。这一前处理已经被许多模型所描述，其中乘性侧抑制具有在电子电路中实现简单的优点，本课题对其进行了研究。由于识别是在完全并行的背景下进行的，基于这一发展中的技术的先进的光子传感器有可能接近生物视网膜的空间识别性能，并在特征获取的时间性能上大大超过生物视网膜。

This project focuses on implementing arrays of photodetectors and associated first-level sensory neural networks in monolithic integrated circuit format for purposes of applying the field of vision to engineering applications of advanced photodetectors and for producing a physical model that can be used in the study of multiplicative lateral inhibition in visual systems. The photodetectors and electronic circuitry are fabricated using gallium-arsenide technology to allow for upward compatibility with integrated optoelectronic systems and to allow the constructed detector assemblies to function in harsher operating environments. Additionally, the monolithic integration of an array of photodetectors with a first level neural network offers advantages in size, reliability, cost, and hence commercial applicability. Biological visual systems perform many demanding tasks that are far from the present capabilities of electronic hardware. The success of many biological systems can mostly be attributed to a large degree of pre-processing of the visual image at the most peripheral level of the sensory system. This pre-processing has been described by many models, one of which is multiplicative lateral inhibition, which has the advantage of simple implementation in electronic circuitry and is studied in this project. Since recognition is studied in a fully parallel context, advanced photonic sensors based upon this developing technology have the potential to approach the spatial recognition performance of biological retinas and to greatly surpass them in temporal performance of feature acquisition.