All optical neural networks

全光学神经网络

• 批准号: 2589868
• 金额: --
• 依托单位: University of Glasgow
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2020
• 资助国家: 英国
• 起止时间: 2020 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2589868
• 关键词: optical; neural; networks

Machine learning (ML) techniques have seen unprecedented success in a range of fields in recent years and, accordingly, significant investment from academia and industry alike. Today, large data centres train ML models on enormous volumes of data to achieve marginal gains in select, lucrative fields, from financial forecasting to autonomous vehicles. In academia, ML has provided a new toolbox for solving otherwise challenging problems and opened new avenues in all areas of science and technology research. The success of such methods relies on scale - large volumes of data, computing power and processing time are used to solve even modest problems. Energy consumption is significant, with individual data centres using the equivalent of a small city's worth in electricity, and the demand for power from the tech sector predicted to increase fifteenfold in the next ten years. The computing hardware used is also extremely exploitative of natural resources, in particular water and rare earth metals, again magnified by the scale with which this hardware is being manufactured to meet demand.Optical neural networks and neuromorphics aim to replicate some of the properties of traditional silicon-based ML using optical hardware, by encoding information in light and processing it through interactions with configurable optical elements. Examples include spatial light modulators (SLM) and random scattering media. Cameras can read out information from such a system, allowing for optical hardware and silicon computing to be paired in a system capable of solving general ML tasks.Optical implementations of machine learning provide several advantages over silicon computing: Energy efficiency is improved by leveraging passive optical elements, and at the limit, single photons can be used as carriers of information. Calculations are performed through light-matter interaction as opposed to individual powered transistors in the case of silicon computing. These calculations are carried out at the speed of light in most cases, giving further advantage over silicon. The ability to easily scale optical hardware in three-dimensional space is also desirable, with silicon computing beginning to see scaling problems as Moore's Law starts to slow. The scope to introduce quantum effects is a clear advantage, giving options for solving classes of problem which may be intractable with classical computing.This research aims to build optical neuromorphic systems to solve specific problems in imaging where optical setups are already in place, and systems which could provide advantage over silicon computing in established ML benchmarks. The primary challenge is developing efficient methods for training such systems in situ with minimal silicon computing overhead, where traditional optimisation techniques aren't feasible. This aims to mimic the natural learning processes which occur in biological systems such as the human brain. Photonic neuromorphics sits at the junction between several highly active areas of research, including machine learning, quantum computing, imaging and even neuroscience. Bringing together developments in all of these areas will open new avenues of research and influence the future of machine learning computing hardware.

近年来，机器学习（ML）技术在一系列领域取得了前所未有的成功，相应地，学术界和工业界都进行了大量投资。如今，大型数据中心在海量数据上训练机器学习模型，以在从金融预测到自动驾驶汽车等精选的利润丰厚的领域实现边际收益。在学术界，机器学习为解决其他具有挑战性的问题提供了一个新的工具箱，并在所有科学和技术研究领域开辟了新的途径。这些方法的成功依赖于规模——大量的数据、计算能力和处理时间被用来解决即使是很小的问题。能源消耗是巨大的，单个数据中心使用的电力相当于一个小城市的电力，预计未来十年科技行业的电力需求将增长15倍。所使用的计算硬件也极大地消耗了自然资源，特别是水和稀土金属，这又被为了满足需求而制造这种硬件的规模所放大。光学神经网络和神经形态旨在利用光学硬件复制传统硅基机器学习的一些特性，方法是在光中编码信息，并通过与可配置光学元件的相互作用对其进行处理。例如空间光调制器（SLM）和随机散射介质。相机可以从这样的系统中读出信息，允许光学硬件和硅计算在一个能够解决一般机器学习任务的系统中配对。与硅计算相比，机器学习的光学实现提供了几个优势：通过利用无源光学元件提高能源效率，并且在极限情况下，单个光子可以用作信息载体。计算是通过光与物质的相互作用来完成的，而不是像硅计算那样使用单个的晶体管。在大多数情况下，这些计算是在光速下进行的，这比硅有进一步的优势。在三维空间中轻松缩放光学硬件的能力也是可取的，随着摩尔定律开始放缓，硅计算开始看到缩放问题。引入量子效应的范围是一个明显的优势，为解决经典计算可能难以解决的问题提供了选择。这项研究旨在建立光学神经形态系统，以解决光学装置已经到位的成像中的特定问题，以及在已建立的机器学习基准中提供优于硅计算的系统。主要的挑战是开发有效的方法，以最小的硅计算开销在现场训练这样的系统，传统的优化技术是不可行的。其目的是模仿发生在生物系统（如人脑）中的自然学习过程。光子神经形态学位于几个高度活跃的研究领域的交汇处，包括机器学习、量子计算、成像甚至神经科学。将所有这些领域的发展结合起来，将开辟新的研究途径，并影响机器学习计算硬件的未来。