Novel devices for optical convolutional neural networks

光学神经网络是具有光学组件的人工神经网络的物理实现。实现卷积函数在这方面特别有利，因为光学处理器可以容易地在傅立叶域中执行卷积。该项目的重点是开发为此类网络提供输入的新型光学设备。特别是，我们将开发一系列空间光调制器来提供2D输入功能。每个调制器将基于一个非对称的法布里-珀罗调制器，其反射率可以通过载波注入来控制。学生将优化这些调制器的设计，在约克纳米中心洁净室制造它们，然后在光子学实验室进行测试。目标是实现反射率从接近0%到接近100%的调制，根据初步计算，这可以在非常低的电流(10-100A)和高速(MHz-GHz)下实现。在单个设备的基础上进行优化后，该项目将实现大规模阵列(例如32x32或64x64)并在系统环境中对其进行测试。该项目是与Optalysys有限公司的合作，Optalysys Ltd.将提供系统专业知识。

An optical neural network is the physical implementation of an artificial neural network with optical components. Realising the convolution function is particularly advantageous in this regard because optical processors can readily perform convolutions in the Fourier domain. This project focusses on the development of novel optical devices that provide the input to such networks. In particular, we will develop an array of spatial light modulators to provide the 2D input function. Each modulator will be based on an asymmetric Fabry-Perot modulator, the reflectivity of which can be controlled by carrier injection. The student will optimise the design of these modulators, fabricate them in the York Nanocentre cleanroom, then test them in the Photonics laboratories. The goal is to achieve a modulation of the reflectivity from near 0% to near 100%, which, according to initial calculations, can be realised with a very low (10-100A) current and at high speed (MHz-GHz). Following the optimisation on the basis of individual devices, the project will then realise large scale arrays (e.g. 32x32 or 64x64) and test them in a systems context. The project is a collaboration with the company Optalysys Ltd. who will provide the systems expertise.