校园文苑|AlphaGo背后，一个“人工智能+”的时代

【财新网】（火线评论/记者 张而弛）3月12日下午，时针走过4点14分。即使再不关心围棋的中国人可能也已经知道：谷歌研发的阿尔法围棋(AlphaGo)以3:0的大比分击败了人类的世界冠军李世石。

(wire caixin 】【 comments/reporter Zhang Erchi) on March 12 in the afternoon, clockwise through 4 PM. Even if he did not care about Chinese go may already know: Google research and development of alpha go (AlphaGo) 5-1 3-0 win over the human world champion lee se-dol.

本次对决前，被称为“棋圣”的聂卫平等众多围棋高手曾经预测，李世石将赢得这场人机大战的胜利。在他们看来，人工智能还远没有进化到能够打败人类顶级高手的水平。但3月9日AlphaGo先下一局后，许多人的问题便瞬间变成了：人类是怎么走到这一步的？人工智能的未来在哪里？

Before the battle, known as the "bear comparison with chess saint" wei-ping nie and so on go ace predicted, lee se-dol will win the man-machine war victory. In their view, artificial intelligence evolution is far from enough to defeat the human master's top level. On March 9, but after AlphaGo next game first, many problems will instantly becomes: man is how to get to this step? Where is the future of artificial intelligence?

从2015年下半年起，许多人工智能领域的专家便开始讨论一个话题：人工智能的拐点是否已经来临？

Since the second half of 2015, many experts in the field of artificial intelligence began to discuss a topic: the turning point of artificial intelligence has come?

随着演算法、机器学习等各方面的进步，过去的五年是人工智能突飞猛进的阶段。以图片识别为例。从2010年起，斯坦福大学每年都会组织一场大规模视觉识别挑战赛。这个比赛的一个项目要求计算机识别图片内容，并将其进行分类。2012年，获胜机器的错误率尚高达15.315%。但到了2015年，这个数字已降到3.567%。此前人类在这个测试中，错误率大约为5.1%。

As algorithms, machine learning and other aspects of progress, in the past five years is artificial intelligence by leaps and bounds. In image recognition, for example. Every year since 2010, Stanford university organized a large-scale visual identification challenge. A project of this game requires computer recognition image content, and classify them. In 2012, winning the error rate is as high as 15.315% of the machine. But by 2015, that number has dropped to 3.567%. After the human in this test, error rate is about 5.1%.

这些“超越人类”的进步都要归功于近年来得到广泛应用的深度学习模型。深度学习是机器学习的一种，它的核心是“深层神经网络”技术，模仿的是人脑中神经网络的工作模式。

The progress of these "beyond human" are credited to deep learning models are widely applied in recent years. Deep learning is a kind of machine learning, it is the core of "deep neural network" technique, in imitation of the brain neural network model of work.

还给机器的自由

负责微软研究院新体验与新技术部门（NExT）的彼得李（Peter Lee）曾向财新记者解释深度学习的原理。

Free of the machine

Responsible for Microsoft research new experience and new technical department (NExT) Peter Lee (Peter Lee) to gain new reporter explained the principle of deep learning.

过去几年，微软研究院一直在探索深度学习在语音识别上的应用，包括他们推出的虚拟机器人“小冰”和实时语音翻译应用Skype Translator。

In the past few years, Microsoft research has been exploring deep learning application in speech recognition, including their launch virtual robot "little ice" and real-time voice translation application Skype the Translator.

为了让机器学会一种语言，彼得解释，微软搭建了一个五类分层的神经网络，通过长期训练，使这个系统能够完成非常准确的语音识别。

In order to let the machine learn a language, Peter explained, Microsoft built a 5 kinds of layered neural network, through long-term training, enables the system to complete very accurate speech recognition.

“想象有一个多层的蛋糕，每一层都在学习人类怎么说话。在蛋糕的最底层，我们让它学习最基础的音素。在上面一层，可能学习连在一起的几个音素，知道声音是如何连在一块儿的。再上面一层是一个个词语，再到词与词的搭配和组合，最后到整个句子的理解。每往上一层，机器学的就越多。”彼得说。

"Imagine that a multi-layer cake, each layer in learning how to speak. Human in the bottom of the cake, we let it study the most basic phonemes. On a layer above, may learn together several phonemes, know how the sound is even together. Then a layer above is a word, and then to the word and the word collocation and combination, and finally to the understanding of the whole sentence. Each layer upwards, machine learning." Peter said to him.

在学会人类语言后，每次听到有人说话，机器就会将声音分割为一个个音素，与系统中已有的数据做对比，然后识别出人类在说什么。

In the society of human language, every time I hear someone speak, which can be the voice for each phoneme segmentation, compared with existing data in the system to do, and then identify the human in what to say.

彼得告诉财新记者：五层的神经网络只是针对语音识别，图片识别等其它任务需要设定的层数每个都会有所不同。

Peter told fortune new reporter: five layer neural network just for speech recognition, image recognition, and other tasks need to set the number of layers in each will be different.

根据谷歌旗下人工智能公司DeepMind发布的资料，这次比赛，AlphaGo一共搭建了12层神经网络，里面有上百万个像人类神经元一样的连接。DeepMind给这个神经网络输入了3000万步人类的围棋走法，使AlphaGo能在57%的时候预测出，人类要走的下一步是什么。

According to Google's artificial intelligence company DeepMind published data, this competition, AlphaGo a total of 12 layer neural network set up, there are millions of like human neuronal connections. DeepMind step input to the neural network of 30 million human go way, make AlphaGo can predict 57% of the time, what was next for the human to go.

不过，这还不是这台机器致胜的关键。DeepMind之所以成名，还是因为它在机器学习的另一个领域有突破性的进展，这就是AlphaGo的杀手锏强化学习（Reinforcement learning）。

However, this is not the key to winning this machine. DeepMind became famous, or because it is in the field of machine learning another breakthrough, this is a killer - AlphaGo Reinforcement learning (Reinforcement learning).

简单来说，强化学习是给人工智能设立目标后，让其向着这个目标不断尝试，通过给予持续的激励，让人工智能自己找到最佳的解决方案。

In simple terms, reinforcement learning is to set up a goal of artificial intelligence, let it toward this goal constantly trying, by giving incentives, make yourself find the best solution for artificial intelligence.

这个领域的突破性在于，以前人们都是通过编程，让机器严格执行人类的命令。为了解决一个问题，人类首先需要找到解决方案，再通过机器能够理解的方式教给它。而强化学习则把这个探索的过程交给了神经网络，让人工智能在一次次的摸索中自己找寻到最优路径，而这个方法有可能是人类都没有想到过的。

Breakthrough in this field is that people are programmed before, let the machine strictly implement human command. In order to solve a problem, the human need to find the solution first, then teach by means of machine can understand it. And reinforcement learning is the process of the exploration to the neural network, make the artificial intelligence in time for yourself to find the optimal path, and this method it is possible that humans are not thought of.

因此，早在这次人机大战之前，DeepMind团队便反复强调：AlphaGo背后不是“手工制作”的程序，而是机器自己琢磨出来的致胜之道，这个能力是“通用”的，并不仅限于在围棋比赛中取胜。

Therefore, as early as before the man-machine war, DeepMind team then repeated: AlphaGo behind is not "handmade" program, but his own ideas out of the way of winning machine, this ability is "general", is not limited to win at chess game.

“人工智能+”

此次AlphaGo获胜的基础在于拥有过去海量的数据和不断试错改进的能力。对于其它坐拥大数据的行业来说，这套方法或许同样适用。

"artificial intelligence"

The AlphaGo victory lies in the basis of the past vast amounts of data and the ability to trial and error to improve continuously. For other industries with large data, this method may also apply.

《连线》杂志创始主编凯文凯利告诉财新记者，他认为，人类即将迎来与互联网类似的“人工智能+”阶段，即在各个行业中加入人工智能，对其加以改造。

Founding editor of wired magazine Kevin kelly told fortune new reporter, he thinks, the human will soon celebrate + "artificial intelligence" phase similar to that of the Internet, that join in various industries, artificial intelligence, transform them.

“从人工智能的架构来说，我的设想是，人工智能会永久地以云服务或互联网服务的形式出售，这和购买电力的道理是一样的。”凯文凯利说，“当你买下这个服务后，你只需要享受即可。想象一下，拥有一个24小时不间断工作、任人差遣的智能管家，该是一件多么幸福的事情。”

"From the architectures of artificial intelligence, my idea is that artificial intelligence would be permanently in the form of cloud services or Internet service to sell, this is the same reason and buying power." Kevin kelly said, "when you buy this service, you only need to enjoy. Imagine having a 24-hour non-stop work, be sent intelligent housekeeper, and this is a how happy things."

同时，他认为，在人工智能技术云端化后，大量数据可以共享，这又将促使人工智能进一步发展，加速新事物的发明进程。

At the same time, he thought, in the artificial intelligence technology after the cloud, a large amount of data can be Shared and this in turn will lead to the further development of artificial intelligence, and quickens the process of the invention of new things.

“从云端数据的角度来说，当人工智能越‘智能’，就有越多的人想要使用它；越多人使用它，它的智能化水平就越高，发展越快，应用范围越广。”凯文凯利对财新记者说。

"From the cloud data point of view, the" smart "artificial intelligence, the more people want to use it; the more people use it, it's the higher intelligent level, the faster the development, application range is wider." Kevin kelly said to gain new reporters.

如此诱人的商业前景，IBM、微软和谷歌等美国科技企业都看得十分清楚，且已开始布局。

So attractive business prospects, the us technology companies such as IBM, Microsoft and Google to see very clear, and has started to layout.

2015年10月，IBM把公司的整个转型方向之一定为“认知计算”，大力推广自己的人工智能系统Watson。在智能家居、医药和机器人领域，IBM分别与美国家电厂商惠而浦、医疗设备公司美敦力、日本的软银等展开合作，共同探索人工智能的变现途径。

In October 2015, IBM rated one of the transformation direction of the company as "cognitive computing," Watson, vigorously promote their own artificial intelligence system. In the field of intelligent household, medicine, and robot, IBM, respectively, with the American electrical appliances manufacturer whirlpool, medtronic medical equipment company, Japan's softbank, such as cooperation, jointly explore the realizable way of artificial intelligence.

2016年1月，DeepMind的联合创始人Demis Hassabis在谷歌博客上表示，他们将把AlphaGo的技术用于解决现代社会最棘手和最紧迫的问题，包括气候建模和复杂疾病分析。2月24日，DeepMind建立健康团队，宣布将与伦敦帝国学院和伦敦皇家NHS信托基金会合作，探索人工智能在医疗领域的实践。

In January 2016, co-founder of DeepMind Demis Hassabis on Google blog said, they will turn AlphaGo technology is used to solve the most difficult and most pressing problem in modern society, including climate modeling and the analysis of complex diseases. On February 24, DeepMind build health team, announced that it would with imperial college London and the royal London NHS trust, cooperation and exploration in the field of artificial intelligence in the medical practice.

不过，虽然前景值得憧憬，但作为一项技术来说，人工智能尚处在发展的初级阶段，如何应用这些最新成果还需要大量的摸索和试错。

However, although the future worth looking forward to, but as a technology, artificial intelligence is still at the primary stage of development, how to apply the latest achievements also requires a lot of grope and trial and error.

在谈论自己家的Watson时，IBM的研究人员把现在的人工智能比成1960年代的大型电脑。

When talking about his home Watson, the IBM researchers have now than in the 1960 s large computer artificial intelligence.

“我们在1960年代的时候，不可能想象到今天计算机会发展成这样。同样，我们今天也不敢想象，20年后的认知计算会给我们带来的震撼。”IBM中国研究院院长沈晓卫对财新在内的媒体表示，“从这个角度来讲，无论是技术还是商业模式，人工智能都还处在婴儿阶段，是非常不成熟的。但这也是其会为行业带来很多兴奋点和震撼点的原因之一。”

"When we were in the 1960 s, it is impossible to imagine today development like this. Also, we also dare not imagine today, after 20 years of cognitive computing will bring us." IBM China dean Shen Xiaowei caixin media said that "from this perspective, both in technology and business model, and artificial intelligence are still in the infant stage, it is very immature. But it is also the will bring the industry a lot of excitement and shock point one of the reasons why."

山大商院外语系

　　　　　　　　　　　　　 查看评论 回复