谷歌推出史上最強人臉識別系統

“I never forget a face,” some people like to boast. It’s a claim that looks quainter by the day as artificial intelligence research continues to advance. Some computers, it turns out, never forget 260 million faces.

有些人總喜歡誇口說：“我從來不會忘記別人的長相。”在人工智能研究突飛猛進的今天，還要這麼誇口就有點奇怪了。事實上，現在有些電腦能記住2.6億張臉。

Last week, a trio of Google GOOG -0.66% researchers published a paper on a new artificial intelligence system dubbed FaceNet that it claims represents the most-accurate approach yet to recognizing human faces. FaceNet achieved nearly 100-percent accuracy on a popular facial-recognition dataset called Labeled Faces in the Wild, which includes more than 13,000 pictures of faces from across the web. Trained on a massive 260-million-image dataset, FaceNet performed with better than 86 percent accuracy.

上週，谷歌公司的三位研究人員發表了一篇有關全新人工智能系統的研究論文。這一系統名爲FaceNet，谷歌號稱它是迄今爲止最精確的人臉識別技術。面對一個名爲“人面數據庫”（Labeled Faces in the Wild）的常用人臉識別數據庫時，FaceNet識別的準確率近乎百分之百。這個數據庫容納了網上搜集的一萬三千多張人臉照片。而在面對一個含有2.6億張人臉照片的龐大數據庫時，這個系統的準確率也超過86%。

Researchers benchmarking their facial-recognition systems against Labeled Faces in the Wild are testing for what they call “verification.” Essentially, they’re measuring how good the algorithms are at determining whether two images are of the same person.

研究人員聲稱，面對“人面數據庫”時，他們主要測試該系統的“確認能力”。就本質而言，他們衡量的是這套算法在判斷兩張照片是否同屬一人時到底有多準確。

In December, a team of Chinese researchers also claimed better than 99 percent accuracy on the dataset. Last year, Facebook researchers published a paper boasting better than 97 percent accuracy. The Facebook FB 1.66% paper points to researchers claiming that humans analyzing images in the Labeled Faces dataset only achieve 97.5 percent accuracy.

去年12月，一箇中國研究團隊也聲稱，對這套數據庫的識別準確率超過99%。去年，Facebook公司的研究人員發表論文稱，他們也能做到超過97%的準確率。根據這篇論文援引的一些研究人員的說法，人類對該數據庫的識別準確率僅有97.5%。

However, the approach Google’s researchers took goes beyond simply verifying whether two faces are the same. Its system can also put a name to a face—classic facial recognition—and even present collections of faces that look the most similar or the most distinct.

不過，谷歌研究人員採用的方法絕不只是確認兩張臉是否一樣這麼簡單。這套系統還能將人名和臉匹配——經典的人臉識別技術，甚至能把看起來最像或最不像的臉歸集在一起。

This is all just research, but it points to a near future where the types of crime-fighting, or surveillance-enhancing, computers we often see on network television and blockbuster movies will be much more attainable. Or perhaps a world where online dating is even simpler (and shallower) than swiping left or right on Tinder.

目前這還僅僅是研究而已，但它預示着，在不遠的將來，我們經常在網上視頻或大片裏看到的那種能懲治犯罪、加強監控的電腦將更加觸手可及。比起在交友應用Tinder上劃來劃去，它可能會使網上交友更加簡單（也更停留於表面）。

Have a thing for Brad Pitt circa 1998? Here are the 500 profiles that look the most like him.

很喜歡1998年左右時的布拉德o皮特？這個數據庫裏有500張看起來很像他的臉。

At first we’ll see systems like Google’s FaceNet and Facebook’s aforementioned system (dubbed “DeepFace”) make their way onto those company’s web platforms. They will make it easier, or more automatic, for users to tag photos and search for people, because the algorithms will know who’s in a picture even when they’re not labeled. These types of systems will also make it easier for web companies to analyze their users’ social networks and to assess global trends and celebrity popularity based on who’s appearing in pictures.

一開始，我們會看到谷歌的FaceNet及Facebook的DeepFace系統在各自的網絡平臺上運行。它們會讓用戶更加方便地（或者說更加自動化地）給照片貼上標籤，找到要找的人，因爲這些算法知道照片中的這個人是誰，即使這些照片並沒有姓名標記。此外，這類系統還能讓網絡公司更加方便地基於照片人物的身份，來分析它們的用戶社交網絡，評判全球流行趨勢及名人的受歡迎程度。

Though Google and Facebook’s advances in facial recognition are relatively new, computer systems like this can be found all around us today. They incorporate an artificial intelligence technique called deep learning, which has proven remarkably effective at so-called machine perception tasks such as recognizing objects (by some metrics, machines are now better at this than are people), recognizing voices, and understanding the content of written text.

儘管谷歌和Facebook在人臉識別技術上最近才取得這類進步，但與之類似的電腦系統早就無處不在。它們都含有一種名爲“深度學習”的人工智能技術。事實證明，這種技術能夠極其有效地完成識別物體（按照某些標準來看，機器在這方面已經比人類要強了）、識別語音及理解書面文字等機器辨別任務。

Aside from Google and Facebook, companies including Microsoft MSFT 0.32% , Baidu, and Yahoo YHOO 0.63% are also investing heavily in deep learning research. The algorithms already power everyday features such as voice control on smartphones, Skype Translate, predictive text-messaging applications, and advanced image-searching. (If you have images uploaded to a Google+ account, go ahead and search them for specific objects.) Spotify and Netflix NFLX -0.82% are investigating deep learning to power smarter media recommendations. PayPal EBAY -0.13% is using it to fight fraud.

除了谷歌和Facebook外，微軟、百度和雅虎也在“深度學習”研究上投入重金。這種算法已經應用在一些我們常用的功能上了，比如智能手機語音控制、Skype實時翻譯、短信預測輸入法及先進的圖像搜索等（如果你已經將一些圖片上傳至Google+賬戶裏，你就可以試試用它們來搜索特定目標）。Spotify和Netflix公司正在研究如何利用深度學習技術更智能地推薦視頻。貝寶公司則將其用於打擊欺詐。

There are also several technology startups using deep learning to analyze medical images in real time, and to provide capabilities such as text analysis, computer vision, and voice recognition as cloud computing services. Twitter, Pinterest, Dropbox, Yahoo, and Google have all acquired deep learning startups in recent years. And IBM IBM -0.08% just bought a Denver-based startup called AlchemyAPI to help make its Watson system smarter and bolster its new Bluemix cloud platform. (The idea: Developers can easily connect mobile and web applications to cloud services and therefore build smart applications without ever studying the complex computer science that underpins artificial intelligence.)

還有幾家科技創業公司正將深度學習技術用於實時分析醫療圖像，並提供諸如文本分析、計算機視覺及語音識別這類雲計算服務項目。近年來，Twitter、Pinterest,、Dropbox、雅虎和谷歌等公司都收購了一些專攻深度學習技術的創業公司。IBM公司剛剛收購了一家位於丹佛，名爲AlchemyAPI的初創企業，用以提升其Watson超級計算機的智能水平，並支持其全新的Bluemix雲平臺（該平臺的理念是：開發者可以方便地將移動和網絡應用與雲服務連接起來，藉以打造一些智能應用，而無需再鑽研人工智能背後複雜的計算機科學）。

That’s not all. As consumer robots, driverless cars and smart homes become real, deep learning will be there, too, providing the eyes, ears, and some of the brains for our new toys. DARPA, the U.S. Department of Defense’s research agency, is also investigating how deep learning techniques might be able to help it make sense of the streams of communications crossing intelligence networks everyday.

不止於此。隨着消費級機器人、無人駕駛汽車及智能家居逐漸成爲現實，深度學習技術也將如影隨形，爲我們這些新玩具提供耳目和一些頭腦功能。美國國防部高級研究計劃局（DARPA）也在探索如何藉助深度學習技術來實時理解龐大的情報信息流。

Something tells me it’s looking at Google’s FaceNet and getting pretty excited, too.

我猜想，DARPA正在關注谷歌的FaceNet系統，併爲之激動。