人工智能真的不能理解人類的情感嗎？為了突破這個(gè)領(lǐng)域，中國的一家頂尖的科技公司正在試圖“教”電腦看懂上世紀(jì)90年代的經(jīng)典愛情災(zāi)難片——《泰坦尼克號(hào)》。

這家中國公司的名字叫商湯科技，該公司希望它的人工智能技術(shù)能夠區(qū)分電影中的戀愛場(chǎng)景與災(zāi)難場(chǎng)景。大多數(shù)人看過這部片子后，都知道它講的是杰克和羅絲在沉船上搞對(duì)象的故事。然而對(duì)于計(jì)算機(jī)來說，要理解這種微妙的人類情感卻是一件極復(fù)雜的任務(wù)。

在《麻省理工科技評(píng)論》上周一在舊金山舉辦的EmTech數(shù)字峰會(huì)上，這項(xiàng)技術(shù)的表現(xiàn)相當(dāng)搶眼，正確地分辨出了各種場(chǎng)景。這充分說明了人工智能技術(shù)的進(jìn)步，不過到目前為止，它要想理解其他更復(fù)雜的電影情節(jié)，還是有很長(zhǎng)的路要走的。

會(huì)上，商湯科技與香港中文大學(xué)聯(lián)合實(shí)驗(yàn)室主任林達(dá)華（他也是港中大的副教授）播放了電影《泰坦尼克號(hào)》的片段。片中，杰克（萊昂納多·迪卡普里奧飾）輕輕地從身后抱住羅絲（凱特·溫斯萊特飾），二人在船頭擺出了經(jīng)典的“我心飛翔”造型。視頻下方有一張小圖表，顯示該公司的電腦認(rèn)為這個(gè)場(chǎng)景是浪漫的還是刺激的。

電腦中的人工智能程序想來已經(jīng)接受了數(shù)千個(gè)視頻和圖像的“訓(xùn)練”，它在分析了影片的數(shù)據(jù)后，認(rèn)為這個(gè)場(chǎng)景是“浪漫”而非“災(zāi)難”。之后林達(dá)華又播放了泰坦尼克號(hào)沉沒的片段，電腦很快確定，這個(gè)場(chǎng)景是“災(zāi)難”而非“浪漫”。??

隨著深度學(xué)習(xí)等人工智能技術(shù)的進(jìn)步，研究人員已經(jīng)可以“訓(xùn)練”計(jì)算機(jī)分辨圖片和視頻中的物體。而從演示的情況看，商湯科技的AI程序已經(jīng)不僅可以分辨視頻中的物體，甚至可以理解視頻片斷的情緒和語境。據(jù)稱，Netflix等美國科技公司也在攻關(guān)類似的技術(shù)，以使用AI程序解析視頻，然后向觀眾推送最吸引他們的推廣視頻。

林達(dá)華并沒有解釋商湯科技是怎樣教會(huì)計(jì)算機(jī)分辨電影場(chǎng)景的語境的，他只是大而化之地談了談公司AI技術(shù)的進(jìn)展，并表示他們的技術(shù)可以用于識(shí)別人類面部表情等領(lǐng)域。

林達(dá)華表示，商湯科技的AI技術(shù)主要賣給了中國的“視頻服務(wù)公司”，也就是那些山寨了YouTube服務(wù)的視頻網(wǎng)站。他表示，這些企業(yè)客戶想知道用戶最喜歡哪類電影場(chǎng)景，好鼓勵(lì)他們觀看更多類似影片，不過他并沒有解釋這些企業(yè)客戶打算如何實(shí)現(xiàn)這個(gè)目標(biāo)。

有人擔(dān)心，中國對(duì)人工智能的使用或許會(huì)令一些美國人產(chǎn)生誤解。在被問到這個(gè)問題時(shí)，林達(dá)華提到了政府利用面部識(shí)別技術(shù)進(jìn)行監(jiān)控的問題。有些人擔(dān)心，這種先進(jìn)的AI監(jiān)控?cái)z像頭將具有面部識(shí)別功能，從而帶來隱私問題。林達(dá)華則表示，中國應(yīng)用人工智能的興趣很廣泛，面部識(shí)別只是其中的“一小部分”，人工智能還可用于醫(yī)療等產(chǎn)業(yè)。（財(cái)富中文網(wǎng)）

譯者：樸成奎?

One of China’s top tech companies is trying to push the frontiers of artificial intelligence by teaching computers to understand scenes from the 1990’s romance-disaster epic Titanic.

The technology, from China company SenseTime, is supposed to distinguish Titanic’sromantic scenes from disaster scenes. Although most humans would have no problem distinguishing Jack and Rose’s blooming love from the Titantic’s sinking, the feat is highly complex for computers.

In a demonstration at MIT Technology Review‘s EmTech Digital conference in San Francisco on last Monday, the technology performed well and was able to classify the scenes correctly. It highlights the advancement of artificial intelligence, but also how far it still has to go before becoming able to understand more complex movie scenes outside of public demonstrations.

Dahua Lin, the director of a joint research lab between SenseTime and the Chinese University of Hong Kong where he’s also an assistant professor, played a video of the scene from Titanic in which Jack (Leonardo DiCaprio) gently holds Rose (Kate Winslet) as she leans over the bow of the doomed passenger ship like she’s flying. Beneath the video was a small chart indicating whether the company’s computers thought the scene was romantic or action-packed.

After crunching data, presumably taken from thousands of videos and image stills in video clips, the computer determined that the scene was more “romantic” than a “disaster.” Then later, when Lin briefly showed the clip of the Titanic sinking, the computers quickly identified the scene as more of a “disaster” than “romantic.”

Advances in AI technologies like deep learning have led to researchers “training” computers to understand objects in photos and videos. SenseTime’s computers, at least as demonstrated, appear to be able to understand the context behind video clips besides merely identifying the objects. U.S. tech companies like Netflix, are also reportedly exploring the use of AI in similar ways to parse videos and then show viewers promotional clips filled with scenes more likely to appeal to them.

Lin didn’t explain how SenseTime taught its computers to distinguish the context of movie scenes, but instead explained more broadly the company’s work developing AI technologies that can do things like recognize human facial expressions.

SenseTime sells its AI technology to Chinese “video services,” he said, likely referring to YouTube copycats. These corporate customers, he said, want to know which movie scenes individual users prefer in order to encourage them to watch more, although he didn’t explain how those corporate customers accomplish that.

Asked to describe possible misunderstandings Americans may have about China’s use of AI technologies, Lin cited the country’s use of facial recognition for government surveillance.Some people?worry that more sophisticated surveillance cameras could recognize individual faces and create privacy problems, among other issues. But Lin minimized the potential pitfalls, saying that facial recognition is just a “small part” of China’s interest in using AI and that AI could also be used to improve industries like healthcare.