想要提高托福閱讀能力，我們一定要在日常生活中有意識地增加英語閱讀量，提升語感和熟練度，這其中比較常用也比較方便地一個方式就是利用各類英文報刊雜志文章進行精讀與泛讀練習。下面我們來看一篇經濟學人文章：谷歌的海馬體。
Artificial intelligence
Google’s hippocampus
Alphabet has plenty of AI expertise, so why does it need DeepMind?
人工智能
谷歌的海馬體
Alphabet已擁有大量人工智能專門技術，為何還需要DeepMind?
DEEPMIND’S office is tucked away in a nondescript building next to London’s Kings Cross train station. From the outside, it doesn’t look like something that two of the world’s most powerful technology companies, Facebook and Google, would have fought to acquire. Google won, buying DeepMind for ￡400m ($660m) in January 2014. But why did it want to own a British artificial-intelligence (AI) company in the first place? Google was already on the cutting edge of machine learning and AI, its newly trendy cousin. What value could DeepMind provide?
人工智能(以下簡稱AI)科技公司DeepMind的辦公室藏身于倫敦國王十字火車站旁邊一座不起眼的建筑物內，從外看去，完全不像是Facebook和谷歌這兩大科技巨頭爭相收購的對象。最終，谷歌勝出，在2014年1月以4億英鎊(6.6億美元)成功收購了DeepMind。但谷歌當初為何要收購這樣一家英國AI公司呢?在機器學習及與之相近的AI技術方面，谷歌早已走在前列。DeepMind能給谷歌帶來什么價值?
That question has become a little more pressing. Before October 2015 Google’s gigantic advertising revenues had cast a comfortable shade in which ambitious, zero-revenue projects like DeepMind could shelter. Then Google conjured up a corporate superstructure called Alphabet, slotting itself in as the only profitable firm. For the first time, other businesses had their combined revenues broken out from Google’s on the balance-sheet, placing them under more scrutiny (see next article). But understanding DeepMind’s worth is not a simple financial question. Its value is deeper than that.
這個問題現在變得更迫切了些。2015年10月之前，谷歌的巨額廣告收入為DeepMind這類雄心勃勃的零收入項目提供了充足的庇蔭。而后谷歌構建了名為Alphabet的母公司架構，并成為公司旗下唯一盈利的公司。其他業務的綜合營收首次從谷歌的資產負債表中拆分出來，因而會受到更多審視。但要了解DeepMind的價值所在并非一個簡單的財務問題。其意義更為深遠。
DeepMind’s most immediate benefit to Google and Alphabet is the advantage it gives in the strategic battle that technology companies are waging over AI (see chart). It hoovers up talent, keeping researchers away from competitors like Facebook, Microsoft and Amazon. The Kings Cross office already houses about 400 computer scientists and neuroscientists, and there is talk of expanding that to 1,000.
DeepMind對谷歌和Alphabet最直接的好處是使其在科技公司圍繞AI展開的戰略競爭中處于有利位置(見圖表)。它吸納了眾多人才，令Facebook、微軟、亞馬遜等競爭對手對其研究人員求之而不得。公司在國王十字火車站旁的辦公樓內現有約400名計算機科學家及神經科學家，據說規模將擴至1000人。
Another boost to the mother ship comes in the form of prestige. DeepMind has reached the cover of Nature, a highly regarded academic journal, twice since it was acquired. Gigantic copies of the relevant covers adorn the walls of the office lobby. The first was for a video-game-playing AI programme the second for one that learned to play the ancient Asian board game of Go. Named AlphaGo for its parent, that software went on to make headlines around the world when it beat Lee Sedol, a South Korean champion, in March 2016 (the match is pictured here).
DeepMind為母公司帶來的另外一個好處是聲望的提升。被收購后，DeepMind已兩次登上權威學術期刊《自然》的封面，相關封面的巨幅復制品就張貼在公司大堂的墻上。首次登上封面是因為一款能玩電子游戲的AI程序，第二次則是由于一款學會了下古老的亞洲棋盤游戲圍棋的程序。這一以母公司名字命名的軟件AlphaGo在2016年3月擊敗了韓國圍棋冠軍李世石(如圖)，一舉登上世界各地的新聞頭條。
DeepMind’s horizons stretch far beyond talent capture and public attention, however. Demis Hassabis, its CEO and one of its co-founders, describes the company as a new kind of research organisation, combining the long-term outlook of academia with “the energy and focus of a technology startup”—to say nothing of Alphabet’s cash. He founded it in 2010, along with Mustafa Suleyman and Shane Legg. Mr Legg and Mr Hassabis met as neuroscience researchers at University College, London; Mr Suleyman is a childhood friend of Mr Hassabis’s.
然而，DeepMind的眼光遠不止于吸引人才和公眾關注。其CEO及聯合創始人德米斯·哈薩比斯(Demis Hassabis)將公司描述為一種新型的研究機構：既擁有學術領域的長遠眼界，也具備“科技創業公司的活力和專注”，而Alphabet的資金就更不用說了。哈薩比斯在2010年與穆斯塔法·蘇萊曼(Mustafa Suleyman)和謝恩·列格(Shane Legg)一起創立了DeepMind。列格與哈薩比斯在倫敦大學學院(University College, London)從事神經科學研究時相識，蘇萊曼則是哈薩比斯兒時的玩伴。
The firm’s overall mission, as Mr Hassabis puts it, is to “solve intelligence”. This would allow the firm to create multifunctional, “general” artificial intelligence that can think as broadly and effectively as a human. Being bought by Google had several attractions. One was access to the technology firm’s computing power. Another was Google’s profitability; a weaker buyer would have been more likely to require DeepMind to make money. This way Mr Hassabis can focus on research rather than the detail of running a firm. And by keeping DeepMind in London, at a safe distance from Google’s Silicon Valley base in Mountain View, he can retain more control over the operation.
正如哈薩比斯所說，公司的整體使命是“解密智能”。這將使公司創造能像人類那樣廣泛高效思考的多功能“通用型”人工智能。公司接受谷歌收購有幾個誘因。一是可藉此獲得谷歌的計算能力。另一個則是谷歌的盈利能力：如果是由財力較弱的買家來收購，則更可能對DeepMind設下盈利要求。而谷歌沒有這樣的要求，哈薩比斯便可專注于研究，而非公司的運營細節。通過把DeepMind留在倫敦，與谷歌位于山景城的硅谷總部保持一段安全距離，他還可以對運營保留更大的控制權。
Were he to succeed in creating a general-purpose AI, that would obviously be enormously valuable to Alphabet. It would in effect give the firm a digital employee that could be copied over and over again in service of multiple problems. Yet DeepMind’s research agenda is not—or not yet—the same thing as a business model. And its time frames are extremely long. Mr Hassabis says the company is following a 20-year road map. DeepMind aims to invent new kinds of AI algorithms, he adds, that are inspired by the way the human brain works. This explains the firm’s large number of neuroscientists. Mr Hassabis claims that seeking inspiration from the brain sets his firm far apart from other machine-learning research units and in particular from “deep learning”, the powerful branch of machine-learning that is being used by the Google Brain unit.
假如他成功實現了通用AI技術，顯然將會為Alphabet帶來巨大的價值，等于為之提供了一名可以被無窮復制的數字化員工，用于解決各種問題。但DeepMind的研究計劃并不是——或者說尚未成為——一種商業模式，而且其未來規劃極為長遠。哈薩比斯表示公司正在執行一個20年期的規劃。他補充道，DeepMind的目標是發明類似人腦運作方式的AI新算法。正因如此，公司聘用了大批神經科學家。哈薩比斯聲稱，從人腦尋求靈感使DeepMind大大有別于其他機器學習研究團隊，尤其是“深度學習”這一正為“谷歌大腦”團隊使用的機器學習的強大分支。
Even if DeepMind never achieves human-level (or indeed, superhuman) artificial intelligence, however, the learning software that it creates along the way can still benefit other Alphabet businesses. This has already happened. In July the company announced that its learning software had found a way to reduce the quantity of electricity that is needed to cool Google data centres, by two-fifths. The software learned about the task by crunching data-centre operation logs, and then optimised the process by running it over and over again in a simulation.
即便DeepMind從來都沒研發出達到人類水平(或甚至超人類)的人工智能，但在研究過程中創建的學習軟件仍可為Alphabet的其他業務帶來好處，而且效果已經顯現。今年七月，公司宣布其學習軟件已找到方法將谷歌數據中心的制冷用電量減少五分之二。該軟件先是分析數據中心的操作日志來理解任務，然后通過反復模擬運行來優化過程。
DeepMind is also applying its AI research to solve problems in its own right. Mr Suleyman, who leads these efforts, has expressed an ambition for DeepMind to help manage energy infrastructure, hone health-care systems and improve access to clean water, in return for revenue streams. The company has already started on health care. Its first paid work came in November in the form of a five-year deal with the Royal Free London, an NHS Foundation Trust, to process 1.7m patient records. Earlier this year it gained access to two data sets from other London hospitals: one million retina scans that it can mine and thereby identify early signs of degenerative eye conditions, and head and neck cancer imagery which, fed into its models, will allow DeepMind’s AI to distinguish between healthy and cancerous tissues.
DeepMind也在應用AI研究來自主解決問題。主管這些工作的蘇萊曼曾表達過此種抱負：希望DeepMind能幫助管理能源基礎設施，完善醫療保健系統，改善潔凈水的供給，以此開拓公司的收入來源。DeepMind已經啟動了醫療保健方面的工作。今年11月，公司獲得了首個付費工作，與NHS公立醫院皇家自由倫敦醫院(Royal Free London)簽下五年的合同，為其處理170萬份病歷。今年早前，DeepMind從倫敦其他醫院獲得了兩組數據集：100萬份視網膜掃描圖，可從中挖掘并辨別出退行性眼病的早期征兆;頭頸部癌癥病例的醫學影像，可輸入到DeepMind的模型中，讓其AI系統學習區分健康和癌變組織。
Da Neu Ron Ron
Skilful programmers and powerful computers are crucial to this applied AI business. But access to data about the real-world environment is also vital. When systems like hospitals, electricity grids and factories are targeted for improvement using AI and machine learning, data about their specific operations are needed.
神經網絡在延展
熟練的程序員及強大的計算機是這類應用型AI業務的關鍵，不過獲取現實世界的數據也至關重要。運用AI及機器學習技術改進醫院、電網及工廠等系統時，獲取其具體操作數據是必需的。
Alphabet, of course, holds huge volumes of data that can be mined for these purposes. But DeepMind will have to acquire lots more in each of the fields it aims to examine. In the case of a recent project it was involved in on lip-reading, for example, it was the acquisition of an unprecedentedly large data set that made it a success. A group of researchers at the University of Oxford, headed by Andrew Zisserman, a computer-vision researcher, led the work. The BBC gave the researchers hundreds of thousands of hours of newscaster footage, in the absence of which they would not have been able to train their AI systems.
當然，在這些方面，Alphabet公司擁有大量數據可供挖掘，但DeepMind必須還要從其有意探究的各個領域獲取更多數據。例如，最近它參與一個關于唇讀的項目之所以取得成功，就是因為獲得了前所未有的大數據集。由計算機視覺專家安德魯·基澤曼(Andrew Zisserman)帶領的一組牛津大學的科研人員負責了該項目。BBC向這些研究者提供了數十萬小時的新聞播音員錄像。沒有這些數據，他們就無法訓練其AI系統。
Alphabet, of course, holds huge volumes of data that can be mined for these purposes. But DeepMind will have to acquire lots more in each of the fields it aims to examine. In the case of a recent project it was involved in on lip-reading, for example, it was the acquisition of an unprecedentedly large data set that made it a success. A group of researchers at the University of Oxford, headed by Andrew Zisserman, a computer-vision researcher, led the work. The BBC gave the researchers hundreds of thousands of hours of newscaster footage, in the absence of which they would not have been able to train their AI systems.
關于數據采集之于DeepMind未來的重要性，哈薩比斯輕描淡寫地表示，人類工程師只要能就有待解決的問題構建模擬情境就足夠了，然后DeepMind便可將學習主體置于這些模擬情境中。但目前運行的大多數機器學習系統并非如此操作。AlphaGo本身就是先在收錄了16萬盤人類棋局、包含數百萬著棋的數據庫中學習之后，才反復自我對弈訓練，加以改進。不過，DeepMind如果真的需要掌握大量個人信息，就必須解決消費者對于企業獲取數據的顧慮。
If it can solve these problems, however, DeepMind will hold immense value as something entirely new for Alphabet: an algorithm factory. That would go far beyond simply being the technology giant’s long-term AI research outfit and talent-holding pool. The data that DeepMind processes can remain the property of the organisations they come from (which should help to allay concerns about privacy), but the software that learns from that data will belong to Alphabet.
但如果這些問題得到解決，DeepMind將為Alphabet帶來巨大的價值，成為其一個全新的部分：一家算法工廠。這樣一來，DeepMind將遠不止是該科技巨頭的AI技術長遠研究機構及人才儲備庫。DeepMind處理的數據的所有權可歸其來源機構(這應有助于減輕人們對隱私外泄的擔憂)，但通過學習這些數據而打造出的軟件將屬于Alphabet。
DeepMind may not ever make significant revenue of its own by applying AI programmes to complex problems. But the knowledge it sends into learning software from those same sets of data may justify the bidding war that brought it into Alphabet’s compass.
DeepMind自己運用AI程序解決復雜問題也許永遠賺不了大錢，但學習軟件從那些數據集中獲取的知識卻意義重大。科技巨頭們掀起收購戰，Alphabet把DeepMind納入麾下，原因或許就在于此。
