谷歌神经网络识别猫脸

时间:2012-07-05 09:32:04来源网络人气:()

谷歌全球计算网络的强大信息处理能力以及神秘的Google X实验室中的技术天才很少为外界所知。但上周我们有幸一睹该公司的强大头脑(不管是人脑还是电脑)在想什么:猫。

谷歌科学家们用1.6万块电脑处理器构建了全球最大的电子模拟神经网络,并通过向其展示自YouTube上随机选取的1000万段视频,考察其能够学到什么。结果显示,在无外界指令的自发条件下,该人工神经网络自主学会了识别猫的面孔。

也许这看起来只是琐碎的成就,除了表明猫主人们热衷于上传宠物视频之外,说明不了更多问题。但实际上该成果表明谷歌在人工智能领域已取得重大进展。对电脑用户而言,人工智能研究一直前景广阔,但迄今成果寥寥。

在爱丁堡一个关于机器学习的会议上,谷歌研究人员所作的演示表明该公司在人工智能领域雄心勃勃,并有极其强大的计算资源作为支撑。

标准的机器学习以及图像识别技术依靠数以千计带标签的图片,对电脑进行初始"训练",使电脑从一开始就对猫脸长什么样有一个概念。但是给图片加标签需要耗费大量人力,并且正如谷歌研究人员所说,"带标签的数据相对有限。

为将搜索能力拓展至面向海量非结构化及无标签数据的图像识别领域,谷歌需要掌握其所谓的"自学"或"深度学习"技术。借助此类技术,未来如果某人有一幅出自不知名画家的描绘不知何处风景的画作,他可将此画的照片上传谷歌,经谷歌将其与互联网上数十亿计的图像进行比对后,此人有相当好的机会获知风景所在地与画家身份。

谷歌上周展示的研究成果,就是向开发此类技术迈出的一步。研究人员借助谷歌数据中心,构建具有10亿个连接的人工神经网络,并用一周时间让这个"新生大脑"接触YouTube视频片段,而未以任何方式贴标签。

英语原文:

The immense processing power of Google's global computing network and the brainpower of its secretive Google X research labs remain largely hidden from a curious world. But this week we were given a glimpse of what the company's great minds, human and electronic, are thinking about: cats.

Google scientists built the world's biggest electronic simulation of a brain, running on 16,000 computer processors, and discovered what it would learn when exposed to 10m clips randomly selected from YouTube videos. Unprompted, the computer brain taught itself to identify the feline face.

That might seem a trivial accomplishment, demonstrating little more than the obsession of cat owners with posting videos of their pets. But in fact Google has made a significant advance in artificial intelligence, a research field that has promised much but delivered little to computer users.

In their presentation at a machine learning conference in Edinburgh, the Google researchers demonstrated the company's ambitions in AI as well as the strength of its computing resources.

Standard machine learning and image recognition techniques depend on initial "training" of the computer with thousands of labelled pictures, so it starts off with an electronic idea of what, say, a cat's face looks like. Labelling, however, requires a lot of human labour and, as the Google researchers say, "there is comparatively little labelled data out there".

"Google needs to master what it calls "self-taught learning" or "deep learning", if it is to extend its search capabilities to recognise images among the vast volume of unstructured and unlabelled data. That would enable someone who, for example, owned an unidentified portrait painted by an unknown artist to submit a photograph of it to a future Google – and stand a reasonable chance of having both the scene and the painter identified through comparison with billions of images across the internet.

The study presented this week is a step towards developing such technology. The researchers used Google data centres to set up an artificial neural network with 1bn connections and then exposed this "newborn brain" to YouTube clips for a week, without labelling data of any sort.

Copyright © 2012   企运网络   地址:衢州市世纪大道新青年大厦413室   浙ICP备11016129号