“谷歌助手”把人工智能帶給大眾

Google’s big bet on computers that can teach themselves is about to face its most significant examination.

谷歌(Google)押注計算機可以自主學習的賭局，即將面臨最重大的考驗。

“Machine learning” has brought artificial intelligence (AI) back into the technology mainstream which, for Google, means using its computing resources to analyse mountains of data to identify patterns and make predictions, from calculating the adverts users are likely to find relevant to whether a digital image shows a cat or a dog.

“機器學習”把人工智能(AI)帶回到科技主流中，對谷歌而言，這意味著利用它的計算能力來分析海量數(shù)據(jù)以識別模式并作出預(yù)測，從計算用戶可能覺得相關(guān)的廣告，到一幅數(shù)字圖像顯示的是貓還是狗。

“It’s now solving problems we don’t know how to solve in any other way,” said Jeff Dean, the engineer who has spearheaded Google’s efforts since it began to focus on the area nearly five years ago. About 100 product teams at Google now apply the technology, he added. The latest — and most visible — product of the push is an intelligent digital assistant, intended to usher in a more natural and intelligent form of human-computer interaction, based on the use of everyday language. The feature — called Assistant — is due to appear, in different guises, in a range of Google products and services in the coming weeks.

“它現(xiàn)在正在解決我們完全不知道如何解決的問題，”自谷歌在近5年前開始聚焦該領(lǐng)域以來一直引領(lǐng)研究的工程師杰夫•迪恩(Jeff Dean)表示。他補充稱，谷歌如今約有100個產(chǎn)品團隊正在應(yīng)用這項技術(shù)。最新(也最顯眼)的產(chǎn)品是一個智能數(shù)字助理，旨在開啟一個更自然、更智能的人機交互模式，基于日常語言的使用。被稱為“助手”(Assistant)的這項功能將于未來幾周以不同形式出現(xiàn)在谷歌一系列產(chǎn)品和服務(wù)中。

That will give it a central place in the company’s efforts to steal users away from some of its rivals’ most successful recent ventures. These include Amazon’s voice-activated home device, Echo; Apple’s smart assistant, Siri; and Facebook’s messaging services, Messenger and WhatsApp.

它將有助于谷歌從某些競爭對手最成功的新項目奪取用戶。這些包括亞馬遜(Amazon)的家庭聲控設(shè)備Echo;蘋果(Apple)的智能助手Siri;以及Facebook的通訊服務(wù)——Messenger和WhatsApp。

But even for a company with Google’s massive computing power and engineering brains, teaching computers to act more naturally and intelligently has required it to confront some of the most intractable computer science problems.

但是，即使是對于像谷歌那樣擁有龐大計算能力和工程設(shè)計人才的公司來說，教會計算機更自然更智能地行動，也需要面對一些最棘手的計算機科學問題。

“Google certainly has the bench strength to make a dent in this problem but no one has cracked the code yet,” said Tim Tuttle, chief executive of MindMeld, an AI start-up that is building its own platform for “conversational” computing.

“谷歌當然擁有足夠強大的人才實力來挑戰(zhàn)這個問題，但是迄今還沒人能完全破解，”AI初創(chuàng)企業(yè)MindMeld的首席執(zhí)行官蒂姆•塔特爾(Tim Tuttle)表示。該公司正在打造自己的“對話式”計算平臺。

Many experts in the AI field credit Google with having edged ahead of its main rivals in machine learning.

AI領(lǐng)域的很多專家承認，谷歌在機器學習方面領(lǐng)先于其主要競爭對手。

It has been showing “leading edge” results in the field, said Oren Etzioni, head of artificial intelligence at the research institute of Microsoft co-founder Paul Allen. He credits it with taking a more open approach than rivals, publishing its research and making its technologies freely available. This open-sourcing has helped it build a wider ecosystem around its approach. “Amazon has adopted a much more closed model and is playing catch-up in machine learning,” said Mr Etzioni. “The people that they have attracted are not at the same level.”

在微軟(Microsoft)共同創(chuàng)始人保羅•艾倫(Paul Allen)的研究所負責AI研究的奧倫•埃齊奧尼(Oren Etzioni)稱，谷歌在該領(lǐng)域展現(xiàn)了“前沿”成果。他認為，這是由于谷歌采取了比對手更開放的姿態(tài)，發(fā)表研究結(jié)果，并使其技術(shù)可以免費獲得。這種開源模式幫助它圍繞自己的方法建立了一個更大的生態(tài)系統(tǒng)。“亞馬遜采用了更封閉的模式，在機器學習領(lǐng)域正追趕谷歌，”埃齊奧尼稱，“他們吸引到的人才不是同一水平的。

All of this has served to raise expectations that Google’s Assistant will reach new standards in understanding language and supplying more intelligent guidance, from answering direct questions to steering users through tasks such as finding a restaurant for dinner or arranging a flight. But the heightened expectations have also greatly elevated the risks. Users are often quick to impute high levels of intelligence to computers that appear to understand language, leaving plenty of room for disappointment when the results fall short.

所有這一切都起到了提高期望值的作用，即谷歌“Assistant”在理解語音和提供更智能的指引上將達到新水平，從回答直接的問題，到指導用戶完成尋找餐廳或安排航班等任務(wù)。但是，期望值提高也大大提升了風險。用戶往往很快認為似乎理解語言的計算機具有高智能，當結(jié)果不盡人意時會非常失望。

Google first disclosed its plans for Assistant at its annual developer conference in May. The technology will take different forms, depending on the device or service where it is used. It is set to be used in a product called Home, a voice-activated gadget modelled on Amazon’s breakthrough Echo. Google also said in May that it would power a text-based intelligent service to appear inside Allo, an app launched yesterday (see below) that is intended to propel Google, belatedly, into messaging.

谷歌于今年5月在年度開發(fā)者大會上首次透露了“Assistant”計劃。該技術(shù)將根據(jù)使用的設(shè)備或服務(wù)而采取不同形式。預(yù)計將用于一款被稱為Home的語音工具產(chǎn)品(效仿亞馬遜的Echo)。谷歌5月時還表示，該技術(shù)將用于在應(yīng)用軟件Allo中驅(qū)動基于文本的智能服務(wù)。近日已發(fā)布的Allo旨在推動谷歌進入即時信息領(lǐng)域。

With these new approaches, the search company is betting that many people are ready to try new ways of interacting with digital devices. Around 20 per cent of searches on Android devices in the US are already conducted by voice, according to Google.

憑借這些新方法，這家搜索公司押注很多人都已準備好嘗試與數(shù)字化設(shè)備交互的新方式。據(jù)谷歌表示，在美國，Android設(shè)備上進行的搜索約20%通過語音完成。

Advances in the quality of techniques like speech recognition have brought the technology to a stage where it is ready for a mass market, said Mr Dean. For instance, Google says its error rate in understanding spoken words, even in a noisy room, has fallen to 8 per cent.

迪恩稱，語音識別等技術(shù)的進步，使得AI達到了可以面向大眾市場的階段。例如，谷歌稱其理解口語單詞的錯誤率(即使是在嘈雜的房間內(nèi))已降至8%。

The company has done a “remarkable job” in areas such as speech recognition and the text-to-speech feature that turns search results into spoken answers, said Mr Tuttle.

塔特爾稱，該公司還在語音識別和文本轉(zhuǎn)換語音(將搜索結(jié)果轉(zhuǎn)換為語音回答)等領(lǐng)域取得了“出色的表現(xiàn)”。

Each of these draws on Google’s roots in internet search, which supplies it with mountains of data about general language usage to fuel its core language engines. “In these contexts, Google has an advantage,” says Mr Tuttle.

這一切成功都利用了谷歌在互聯(lián)網(wǎng)搜索方面的根基，后者使其可以利用有關(guān)一般語言用法的海量數(shù)據(jù)來推動其核心語言引擎。“在這些方面，谷歌具有優(yōu)勢，”塔特爾表示。

However, understanding language at the deeper level involves grasping the context of a statement, which is often not obvious, or being able to follow a sequence of comments that follow human but not computer logic. These are things that trip up general-purpose tools such as Assistant, said Mr Tuttle.

然而，若要在更深層面上理解語言，就必然涉及掌握一句話的背景(往往不明顯)或是能夠理解一系列遵循人類(而非計算機)邏輯的評論。塔特爾稱，這些任務(wù)會使“Assistant”等通用工具出錯。

In taking on the more intractable challenges, Google is looking to draw on deep learning, the most advanced form of machine learning. Patterned on the workings of the human brain, deep learning systems use multiple processing layers, like artificial neural networks, to filter data to reach their results. The technology is particularly well suited to things that computers have traditionally found impossible, such as image recognition, and has been applied most strikingly in Google’s Photos app to automatically identify people or objects in users’ albums.

為了應(yīng)對更棘手的挑戰(zhàn)，谷歌正在尋求利用深度學習——機器學習的最高級形式。深度學習系統(tǒng)借鑒人類大腦的工作方式，利用多個處理層(就像人工神經(jīng)網(wǎng)絡(luò)那樣)來過濾數(shù)據(jù)以得到結(jié)果。這項技術(shù)特別適合于處理傳統(tǒng)電腦不可能完成的任務(wù)，比如圖像識別。該技術(shù)迄今最引人矚目的應(yīng)用是在谷歌相冊(Photos)的用戶相簿中自動識別人或物體。

According to Mr Dean, the sort of breakthroughs made in image recognition are now beginning to be seen in language, divining context and meaning where other programs have foundered. “What’s happened recently is the deep learning approaches have started showing an ability to understand language for many different tasks,” he said.

據(jù)迪恩表示，圖像識別上的這種突破，如今已經(jīng)開始出現(xiàn)在語音、語境和語意推測方面;在這些方面，其他程序已失敗。“最近出現(xiàn)的情況是，深度學習方法開始在很多不同的任務(wù)中表現(xiàn)出了理解語言的能力，”他稱。

He concedes, though, that Google’s computers are still far from matching human levels of language comprehension, or replicating the broad understanding of the world that people draw on when holding a conversation. “We have a pretty good ability to understand shorter sentences or utterances,” said Mr Dean. “But we don’t have the ability in long-range context, or the deep background models a human has from other areas when you are talking.”

盡管如此，他承認谷歌的計算機距離人類語言理解能力、或者人類在對話時利用深厚背景知識的程度仍然很遠。“我們在理解較短的句子或表達時擁有相當出色的能力，”迪恩稱，“但是我們無法理解長程語境和人類在說話時來自其他方面的深層背景模式。

A further challenge will be to restrict the situations in which Assistant can handle tasks automatically, limiting it to areas where there is little chance of it making a mistake. It is one thing to unleash a deep learning program to identify pictures of cats, said Mr Dean, but it is another to set the same program free to make changes to your travel itinerary, where a slight misunderstanding would cause deep inconvenience.

還有一個挑戰(zhàn)將限制“Assistant”自動處理任務(wù)的情形，把它限制在犯錯幾率很小的領(lǐng)域。迪恩稱，釋放一款深度學習程序來識別貓咪照片是一回事，而放手讓同樣的程序來更改你的行程則是另一回事。在后面一種情形中，細微的誤解都會造成極大的不便。

As a result, the packaging of the new Assistant technology — finding a useful set of tasks that it can do well, without over-promising or disappointing — is likely to be as important to its success as the underlying technical achievements themselves. “The best technologies don’t always translate to the best product or the winner in the market place,” said Mr Etzioni.

其結(jié)果是，新“Assistant”技術(shù)的包裝——在不過度承諾或讓人失望的情況下，找到一套它可以順利完成的任務(wù)——可能會和它本身作為根本性技術(shù)成就的成功同樣重要。“最好的技術(shù)并不總是轉(zhuǎn)化為最棒的產(chǎn)品或市場上的贏家，”埃齊奧尼稱。

Google has already seen Amazon steal a march with the groundbreaking Echo, and Apple catch the popular imagination with Siri. With Assistant, it is time to get back into the conversation.

在眼看著亞馬遜以開創(chuàng)性的Echo先聲奪人、蘋果以Siri抓住大眾想象力之后，谷歌是時候在“Assistant”的幫助下重新成為關(guān)注焦點。