神经网络谷歌人工智能文章中英文对照解答

非专业人士翻译如有错误请谅解

Google's Al Reas ons Its

Way around the Lon don

Un derground

伦敦地铁系统的路线

DeepMind's latest technique uses external memory to solve tasks that requirelogic and reas oning—astep toward more huma nlike AI

深度思维最新技术使用了外部存储来解决需要逻辑思

维和推理能力的任务

By Elizabeth Gibney,Nature magazine on October 14,2016

伊丽莎白•吉布尼2016年10月14日发表于《自然》杂志

Artificial-in tellige nee(Al) systems known as n eural n etworks can recog nize images, tran slatelan guages and eve n master the an cie nt game of Go.But their limited ability to represe nt complex relati onships betwee n data or variables has preve nted them from conq ueri ng tasks that require logic and reasoning.

人工智能Al系统被认为是神经网络可以识别图片翻译甚至精通古老的游戏。但他们描绘数据或变量之间的复杂关系的能力有

限这妨碍了他们克服需要逻辑思维和推理能力的任务。

In a paper published in Nature on October 12, the

Google-ow ned compa ny DeepM ind in London reveals that it has take n a step towards overco ming thishurdle by creati ng a n eural n etwork with an exter nal memory.The comb in ati on allows the n euralnetwork not only to learn,but to use memoryto store and recall facts to make inferen ces like a conven ti onalalgorithm.This in turn en ables it to tackle problems such as n avigati ng the London Un dergro und withoutany prior kno wledge and solv ing logic puzzles.Though solv ing these problems would not be impressivefor an algorithm programmed to do so, the hybrid system man ages to accomplish this without anypredefined rules.

在10月12日 《自然》杂志中发表的一篇论文中谷歌在伦敦的子公司深度思维展示了他们通过结合外部存储创造了一个神经网络 来进一步克服这些障碍。这种和外部存储的结合不仅允许神经网络学习 还可以通过存储器来存储和回忆事件并以此来像正常情况那样做推断。这反过来能够让它解决难题 比如在没有任何经验的情况下操控伦敦地铁 比如解决逻辑谜题。尽管对于一个算法程序来说做到这点并不会令人印象深刻但这个混合系统在没有任何先决条件的情况下做到了这点。

Although the approach is not en tirely new —DeepM ind itselfreported attempting a similar feat in a preprint in 2014 —“the progressmade in this paper is remarkable says Y osh ua Ben gio,a computer scie ntist at the Uni versity of Mon trealin Can ada.

虽然这个方法不是一个全新的技术一一深度思维自己就在 2014年报告过他们尝试了一种相似的技术——但 “在论文中的这个进步是非凡的”加拿大蒙特利尔的计算机学家本吉奥.本希奥赞叹道。

MEM ORY MAGIC

记忆魔法

A n eural n etwork lear ns by stre ngthe ning conn ecti ons betwee n virtual n euro n-like un its.Without a memory,such a n etwork might need to see a specific London Undeground map thousands of times to learn the best way to n avigatethe tube.

神经网络通过加强虚拟神经元之间的联系来学习。如果没有存储器这样一个网络可能需要看一副特定的伦敦地铁地图数千次来学习最佳路线。

DeepM in d's new system—which they call a'differe ntiable n eural computer'—ca n make sense of a map it hasn ever see n before. It first trains its n eural n etwork on ran domly gen erated map-like structures(which could represe ntstati ons conn ected by lin es,or other relati on ships), i n the process lear ning how to store descripti ons of these relati onships in its exter nal memory as well as an swer questi ons about them.Confron ted with a new map, the DeepM ind systemcan write these new relati on ships —conn ecti ons betwee n

Un dergro und stati ons, in one example from the paper—to memory,and recall it to pla n a route.

深度思维的新系统一一他们称它为微分神经计算机一一可以理解它从未见过的地图。第一次训练神经网络是在随机生成的类似结构的地图上(被铁路线链接的车站或者其他关系)在这个过程中学习如何将这些关系的描述存储在它的外部存储器并且回答问题。面对一个新的地

图深度思维的系统可以把这些新关系一一按照一个图纸上例子来连接各地铁站之间的关系一一写到存储器 并能够回忆这些关系然后计划路

线。

DeepMind's AI system used the same technique to tackle puzzles that require reas oning.After training on 20differe nt types of questi on-an d-a nswer problems, it lear nt to make accurate deducti ons.

For example, the system deduced correctly that a ball is in a playgro und,hav ing bee n in formed that “Joh n picked upthe footballand“Joh n is inthe playgro und It got such problems right morethan 96%of the time.The system performed better than „ recurrentneural networks ',which also have a memcbyt one that is in the fabric of the n etwork itself,and so is less flexible tha n anexter nal memory.

深度思维的人工智能系统使用同样的方法来处理需要推理能力的智力游戏。在通过20种不同类型的问答训练之后它学会了做出准确的推论。例如系统通过被告之“约翰抓着足球”和“约翰在操场上”准确的推断出一个球在操场上。答对问题的概率超过了96%这个系统

的效率比拥有一个内部存储器的周期神经网络更高也更灵活。

Although the DeepMind technique has proven itself on only artificial probl ems, it could be applied to real-worldtasks that in volve making inferen ces from huge amounts of data.This could solve questi ons whose an swers are notexplicitly stated in the data set, says Alex Graves,a computer scie ntist at DeepM ind and a co-author on the paper.Forexample, to determ ine whether two people lived in the same country at the same time, the system might collate facts fromtheir respective Wikipedia pages.

虽然深度思维的技术已被证明只针对人工问题但它能够被应用

到需要通过海量数据来进行推断的真实世界的工作。 这能够解决那些在

数据中没有明确答案的问题。 来自深度思维的计算机科学家 研究报告

的合著者亚历克斯•格雷夫斯介绍说。例如对于判断两人是否在同一时间住在同一个国家系统可能会核对他们各自在维基百科上的事项。

Although the puzzles tackled by DeepMind 's Al are simple,

Ben gio sees the paper as a sig nal that n eural n etworks are adva ncing bey ond mere patter n recog niti on to huma n-liketasks such as reas oning.human-level Al. ”

虽然对于深度思维的人工智能来说智力游戏很简单但本希奥认为该论文是一个信号它表明神经网络正在跨越单纯的模式识别成长到能够做人类才能做的任务例如推理。 “如果我们想实现像人一样的人工智能这次突破是非常重要的。 ”

This article is reproduced with permissi on and was first published on October 13,2016.

这篇文章允许被转载首次发表是2016年10月13日。