Modern information technology (it) relies on division of labour. Photons carry data around the world and electrons process them. But, before optical fibres, electrons did both—and some people hope to complete the transition by having photons process data as well as carrying them.

现代信息技术依赖劳动分工。光子传输全球的数据，电子处理数据。但在光纤问世之前，两项任务都由电子完成。有人希望实现彻底的转变，使数据的传输和处理都由光子来完成。

Unlike electrons, photons (which are electrically neutral) can cross each others’ paths without interacting, so glass fibres can handle many simultaneous signals in a way that copper wires cannot. An optical computer could likewise do lots of calculations at the same time. Using photons reduces power consumption, too. Electrical resistance generates heat, which wastes energy. The passage of photons through transparent media is resistance-free.

不同于电子，光子（具有电中性）可以在不相互作用的情况下穿过彼此的路径，所以玻璃纤维能处理许多同步信号，而铜线做不到。同样，光学计算机能同时进行大量的计算。利用光子还能减少功耗，电阻发热会有能量损失，而光子在透明介质中传输是没有电阻的。

For optical computing to happen, though, the well-established architecture of digital electronic processing would have to be replaced by equivalent optical components. Or maybe not. For some people are working on a novel optical architecture that uses analogue rather than digital computing (that is, it encodes data as a continuous signal rather than as discrete “bits”). At the moment, this architecture is best suited to solving one particular class of problems, those of a branch of maths called linear algebra. But that is a potentially huge market, for linear algebra is fundamental to, among other matters, artificial neural networks, and they, in turn, are fundamental to machine learning—and thus artificial intelligence (ai).

但要想实现光计算，成熟的数字电子运算架构必须由等效的光学器件来代替。也许没这个必要，因为有人正在研发一种新型光学架构，它使用模拟而非数字计算（将数据编码成连续的信号而不是离散的“比特”）。目前，这种架构最适合解决一类问题：称为线性代数的数学分支。但这是潜在的巨大市场，因为线性代数是人工神经网络的基础，而人工神经网络是机器学习——进而成为人工智能的基础。

The power of the matrix

矩阵的力量

Linear algebra manipulates matrices. These are grids of numbers (representing coefficients of simultaneous equations) that can be added and multiplied a bit like individual numbers. Among the things which can be described by matrices are the equations governing the behaviour of electromagnetic radiation (such as light) that were discovered in the 19th century by James Clerk Maxwell. Light’s underlying Maxwellian nature makes it easy, using appropriate modulating devices, to encode matrix data into light beams and then manipulate those data.

线性代数计算矩阵。矩阵是数字网格（代表联立方程式的系数）有点类似于相加和相乘的单个数字。矩阵可用来描述的事物包括19世纪詹姆斯·克拉克·麦克斯韦发现的控制电磁辐射行为（例如光）的方程组。光具有潜在的麦克斯韦特性，所以利用适当的调制设备易于将矩阵数据编码成光束，然后运算这些数据。

Artificial neural networks are programs that represent layers of nodes, the connections between which represent numbers in matrices. The values of these change in response to incoming signals in a way that results in matrix multiplication. The results are passed on to the next layer for another round of processing, and so on, until they arrive at a final output layer, which synthesises them into an answer. The upshot is to allow a network to recognise and learn about patterns in the input data.

人工神经网络是表示多层节点的程序，节点之间的连接表示矩阵数据，它们的数值随着输入信号发生变化，结果导致矩阵相乘。计算结果传递给下一层，进行新一轮运算，依此类推，直到传递至最终输出层，将这些计算结果合成一个答案，最终使神经网络能够识别和学习输入数据中的模式。

The idea of turning neural networks optical is not new. It goes back to the 1990s. But only now has the technology to make it commercially viable come into existence. One of the people who has observed this transition is Demetri Psaltis, an electrical engineer then at the California Institute of Technology (Caltech) and now at the Swiss Federal Institute of Technology in Lausanne. He was among the first to use optical neural networks for face recognition.

神经网络向光学转变不是新理念。它可以追溯到20世纪90年代，但直到现在才出现使它具有商业可行性的技术。杰梅特里·普萨尔蒂斯目睹了转变过程，他当时在加州理工学院担任电气工程师，现在任职于洛桑联邦理工学院，他率先将光神经网络应用于人脸识别。
原创翻译：龙腾网 http://www.ltaaa.cn 转载请注明出处


The neural networks of Dr Psaltis’s youth were shallow. They had but one or two layers and a few thousand nodes. These days, so-called deep-learning networks can have more than 100 layers and billions of nodes. Meanwhile, investments by the telecoms industry—the part of it that ships data around through all those optical fibres—have made it possible to fabricate and control optical systems far more complex than those of the past.

神经网络在普萨尔蒂斯的青年时期比较简单，只有一两层和几千个节点，现在所谓的深度学习网络拥有100多层和几十亿个节点。同时，电信行业的投资——利用光纤传输数据的那部分业务——使人们得以制造和控制远比过去更复杂的光学设备。

That is the technological push. The financial pull derives from shedding the cost of the vast amount of electricity consumed by modern networks as they and the quantities of data they handle get bigger and bigger.

这是技术上的动力。随着现代网络及其处理的数据量越来越大，耗电量十分巨大，而经济上的吸引力是降低电力的成本。

Most efforts to build optical neural networks have not abandoned electrons entirely—they pragmatically retain electronics where appropriate. For example, Lightmatter and Lightelligence, two firms in Boston, Massachusetts, are building hybrid “modulators” that multiply matrices together by manipulating an optically encoded signal according to numbers fed back electronically. This gains the benefit of parallelism for the optical input (which can be 100 times what electronics would permit) while using more conventional kit as what Nicholas Harris, Lightmatter’s founder, describes as the puppet master.

建造光神经网络的大多数行动没有完全抛弃电子——他们务实地在适当的位置保留电子元件。例如，位于美国马萨诸塞州波士顿的两家公司Lightmatter和Lightelligence正在制造混合型“调制器”，它们根据电子反馈的数据来处理光学编码信号，从而实现矩阵相乘。好处是使用比较常规的配套元件就能达到光纤输入量（是电子元件输入量的100倍），Lightmatter公司创始人尼古拉斯·哈里斯称其为傀儡主人 。
原创翻译：龙腾网 http://www.ltaaa.cn 转载请注明出处


The modulators themselves are made of silicon. Though this is not the absolute best material for light modulation, it is by far the best-developed for electronics. Using silicon allows hybrid chips to be made with equipment designed for conventional ones—perhaps even affording it a new lease of life. For, as Maurice Steinman, vice-president of engineering at Lightelligence, observes, though the decades’ long rise in the performance of electronics is slowing down, “we’re just at the beginning of generational scaling on optics”.

这些调制器本身是由硅制成的。硅不是光调制的绝佳材料，但它是迄今针对电子元件开发的最佳材料。如果使用硅材料，设计用于制造常规芯片的设备可以制造混合芯片——甚至可能使它重获新生。因为正如Lightelligence公司的工程副总裁莫里斯·斯泰因曼所说的，电子元件持续数十年的性能提升正在放缓，“但光学元件的更新换代才刚刚开始”。

Ground zero

起点

Ryan Hamerly and his team at the Massachusetts Institute of Technology (the organisation from which Lightelligence and Lightmatter were spun out) seek to exploit the low power consumption of hybrid optical devices for smart speakers, lightweight drones and even self-driving cars. A smart speaker does not have the computational and energetic chops to run deep-learning programs by itself. It therefore sends a digitised version of what it has heard over the internet to a remote server, which does the processing for it. The server then returns the answer.

瑞恩·哈姆利和他的团队来自麻省理工学院（Lightelligence和Lightmatter公司的前身），他们设法将混合光学设备的低功耗特性应用于智能音箱、轻型无人机、甚至无人驾驶汽车。智能音箱不具备运行深度学习程序的运算能力和能量，所以从互联网接收数据后发送至远端服务器来运算数据，然后将运算结果反馈给智能音箱。

All this takes time, though, and is insecure. An optical chip put in such a speaker could perform the needed linear algebra there and then, with low power consumption and without having to transfer potentially sensitive data elsewhere.

但这一过程需要时间，并且不安全。智能音箱内置的光学芯片可以进行必要的线性代数运算，同时做到低功耗，避免了将潜在敏感的数据传输到别的地方。

Other researchers, including Ugur Tegin, at Caltech, reckon optical computing’s true benefit is its ability to handle large data sets. At the moment, for example, image-recognition systems are trained on low-resolution pictures, because high-res versions are too big for them to handle efficiently, if at all. As long as there is an electronic component to the process, there is limited bandwidth. Dr Tegin’s answer is to forgo electronics altogether and use an all-optical machine.

加州理工学院的乌格·泰吉纳等其他研究人员认为，光学运算的真正益处是能够处理大型数据集。例如，目前的图像识别系统被训练识别低分辨率图片，因为高分辨率图片对它们来说太大了，即使能够处理，效率也很低。只要有电子元件参与处理这一过程，带宽就会受限。泰吉纳博士的解决办法是彻底放弃电子元件，使用全光学设备。

This has, however, proved tricky—for what allows neural networks to learn pretty well any pattern thrown at them is the use, in addition to all the linear processing, of a non-linear function in each of their nodes. Employing only linear functions would mean that only linear patterns could be learned.

但事实证明难以做到——因为要想让神经网络学会几乎所有的模式，那么除了所有的线性运算，还需要使用每个节点中的非线性函数。如果只用线性函数，意味着神经网络只能学会线性模式。
原创翻译：龙腾网 http://www.ltaaa.cn 转载请注明出处


Fortunately, although light does behave mostly in a linear fashion, there is an exception. This, Dr Tegin explains, is when an extremely short and intense pulse of it is shone through a so-called multi-mode fibre, which exploits multiple properties of light to enhance its ability to carry parallel signals. In these circumstances, the pulse’s passage changes the properties of the material itself, altering the behaviour of the passing light in a non-linear manner.

可喜的是，虽然光以线性模式为主，但也有例外。泰吉纳博士解释说，当利用所谓的多模光纤传输极短和极强的光脉冲时就会出现例外，多模光纤利用光的多种特性来增强其传输并行信号的能力。在这种情况下，光脉冲的传输改变了材料自身的特性，以非线性的方式改变了光的行为。
原创翻译：龙腾网 http://www.ltaaa.cn 转载请注明出处


Dr Tegin exploited this feature in what is, save its final output layer, an all-optical network. He describes this in a paper published last year in Nature Computational Science. He is able to keep all of the information in an optical form right up until its arrival at the last layer—the one where the answer emerges. Only then is it converted into electronic form, for processing by the simpler and smaller electronic network which makes up this layer.

泰吉纳博士将这种特性应用于全光网络，最终输出层除外。去年，他在《自然计算科学》杂志上发表的论文中阐述了这一原理。他能以光学形式保存所有的数据，直到传输至最后一层，即得出运算结果的那层。直到这时数据转化为电子形式，由构成该层的比较简单和小型的电子网络进行运算。

Meanwhile, at the University of California, Los Angeles, Aydogan Ozcan is taking yet another approach to all-optical matrix processing. In a paper published in Science in 2018, he and his collaborators describe how to create optical devices that do it without involving electrons at all.

同时在洛杉矶的加州大学，埃道甘·奥兹坎正在采取另一种方法实现全光学矩阵运算。在2018年《科学》杂志上发表的论文中，他与合作伙伴描述了如何在完全不使用电子元件的情况下制造出这样的光学设备。

The magic here lies in the use of thin sheets of specially fabricated glass, each the size of a postage stamp, laid on top of each other in stacks analogous to the layers of an artificial neural network. Together, these sheets diffract incoming light in the way that such a neural network would process a digital image.

神奇之处是使用特制的薄玻璃片，每片都有邮票大小，相互层叠在一起，就像人工神经网络的多层结构一样。这些玻璃片会以神经网络处理数字图像的方式衍射输入光。

In this case, the optics work passively, like the lens of a camera, rather than receiving active feedback. Dr Ozcan says that provides security benefits. The system never captures images or sends out the raw data—only the inferred result. There is a trade-off, though. Because the sheets cannot be reconfigured they must, if the inference algorithm changes, be replaced.

在这种情况下，光学器件就像照相机镜头一样被动工作，而不是接收主动反馈。奥兹坎博士说这能带来安全方面的益处，该系统绝不会采集图像或发送原始数据，而只会发送推导的结果。但这样做也有弊端，由于无法重新配置玻璃片，所以如果推理算法发生变化，必须更换玻璃片。

How far optical computing of this sort will get remains to be seen. But ai based on deep learning is developing fast, as recent brouhaha about Chatgpt, a program that can turn out passable prose (and even poetry) with only a little prompting, shows. Hardware which can speed up that development still more is thus likely to find favour. So, after decades in the doldrums, the future of optical computing now looks pretty bright.

这种光学计算能走多远仍有待于观察。但是，基于深度学习的人工智能发展很快，最近火爆的Chatgpt就是明证，只需略加提示，这种程序就能创作出差强人意的散文（甚至诗歌）。能够进一步加快其发展的硬件可能因此受到青睐。所以经过几十年的低迷阶段后，现在看来光计算的前景非常光明。