Let me tell you a story. It goes back 200 million years. It's a story of the neocortex, which means "new rind." So in these early mammals, because only mammals have a neocortex, rodent-like creatures. It was the size of a postage stamp and just as thin, and was a thin covering around their walnut-sized brain, but it was capable of a new type of thinking. Rather than the fixed behaviors that non-mammalian animals have, it could invent new behaviors. So a mouse is escaping a predator, its path is blocked, it'll try to invent a new solution. That may work, it may not, but if it does, it will remember that and have a new behavior, and that can actually spread virally through the rest of the community. Another mouse watching this could say, "Hey, that was pretty clever, going around that rock," and it could adopt a new behavior as well.
首先,我想與大家分享一個故事。 時鐘撥回到兩億年前, 我們的故事, 與新皮層(neocortex)有關。 早期哺乳動物 (實際上只有哺乳動物才有新皮層) 比如齧齒類動物, 擁有一種尺寸和厚度與郵票相當的新皮層, 它像一層薄膜, 包覆著這些動物核桃大小的大腦。 新皮層的功能不可小覷, 它賦予動物新的思考能力。 不像非哺乳類動物, 牠們的行為基本上固定不變, 擁有新皮層的哺乳動物能發明新的行為。 比如,老鼠逃避天敵的追捕時, 一旦發現此路不通, 牠會嘗試去找新的出路。 最終可能逃之夭夭,也可能落入貓口, 但僥倖成功時,牠會記取成功的經驗, 最終形成一種新的行為。 值得一提的是,這種新近習得的行為, 會迅速傳遍整個鼠群。 我們可以想像,一旁觀望的老鼠會說: “哇,真是急中生智,居然想到繞開石頭來逃生!” 然後,輕而易舉也掌握了這種技能。
Non-mammalian animals couldn't do any of those things. They had fixed behaviors. Now they could learn a new behavior but not in the course of one lifetime. In the course of maybe a thousand lifetimes, it could evolve a new fixed behavior. That was perfectly okay 200 million years ago. The environment changed very slowly. It could take 10,000 years for there to be a significant environmental change, and during that period of time it would evolve a new behavior.
但是,非哺乳動物 對此完全無能為力, 牠們的行為一成不變。 準確地說,牠們也能習得新的行為, 但不是在一朝一夕之間, 可能需要歷經一千個世代, 整個種群才能形成一種新的固定行為。 在兩億年前的蠻荒世界, 這種進化節奏並無大礙。 那時,環境變遷步履蹣跚, 大約每一萬年, 才發生一回滄海桑田的巨變, 在這樣一個漫長的時間跨度裏, 動物才形成了一種新的行為。
Now that went along fine, but then something happened. Sixty-five million years ago, there was a sudden, violent change to the environment. We call it the Cretaceous extinction event. That's when the dinosaurs went extinct, that's when 75 percent of the animal and plant species went extinct, and that's when mammals overtook their ecological niche, and to anthropomorphize, biological evolution said, "Hmm, this neocortex is pretty good stuff," and it began to grow it. And mammals got bigger, their brains got bigger at an even faster pace, and the neocortex got bigger even faster than that and developed these distinctive ridges and folds basically to increase its surface area. If you took the human neocortex and stretched it out, it's about the size of a table napkin, and it's still a thin structure. It's about the thickness of a table napkin. But it has so many convolutions and ridges it's now 80 percent of our brain, and that's where we do our thinking, and it's the great sublimator. We still have that old brain that provides our basic drives and motivations, but I may have a drive for conquest, and that'll be sublimated by the neocortex into writing a poem or inventing an app or giving a TED Talk, and it's really the neocortex that's where the action is.
往後,一切安好。 直到,禍從天降。 時間快進到6500萬年前, 地球遭遇一場突如其來的環境遽變, 後人稱之為“白堊紀物種大滅絕”。 恐龍遭受滅頂之災; 75%的地球物種 走向滅絕; 而哺乳動物 趁機佔領了其他物種的生存地盤。 我們可以假託這些哺乳動物的口吻, 來評論這一進化過程: “唔,關鍵時候我們的新皮層真派上用場了。” 此後,新皮層繼續發育。 哺乳動物個頭也日漸見長, 大腦容量迅速擴大, 其中新皮層的發育堪稱突飛猛進, 已經逐步形成獨特的溝回和褶皺, 這可以進一步增加其表面積。 人類的新皮層, 如果充分展開平鋪, 尺寸可達一張餐巾大小。 但它仍然保持了纖薄的結構, 厚度也與餐巾不相上下。 外形曲折複雜,呈現千溝萬壑, 新皮層已佔據大腦體積的80%左右, 不僅肩負思考的重任, 還約束和昇華個人的行為。 今天,我們的大腦 仍然製造原始的需求和動機。 但是,對於我們內心狂野的征服欲望, 這個新皮層起著春風化雨、潤物無聲的作用, 最終將這種欲望化作創造詩歌、開發APP、 甚至是發表TED演講這樣的文明行為。 對於這一切, 新皮層功不可沒。
Fifty years ago, I wrote a paper describing how I thought the brain worked, and I described it as a series of modules. Each module could do things with a pattern. It could learn a pattern. It could remember a pattern. It could implement a pattern. And these modules were organized in hierarchies, and we created that hierarchy with our own thinking. And there was actually very little to go on 50 years ago. It led me to meet President Johnson. I've been thinking about this for 50 years, and a year and a half ago I came out with the book "How To Create A Mind," which has the same thesis, but now there's a plethora of evidence. The amount of data we're getting about the brain from neuroscience is doubling every year. Spatial resolution of brainscanning of all types is doubling every year. We can now see inside a living brain and see individual interneural connections connecting in real time, firing in real time. We can see your brain create your thoughts. We can see your thoughts create your brain, which is really key to how it works.
50年前,我完成了一篇論文, 探究大腦的工作原理, 我認為大腦是一系列模塊的有機結合。 每個模塊按照某種模式各司其職, 但也可以學習、記憶新的模式, 並將模式付諸應用。 這些模式以層級結構進行組織, 當然,我們借助自己的思考 假設了這種層級結構。 50年前,由於各種條件限制, 研究進展緩慢, 但這項成果使我獲得了 約翰遜總統的接見。 50年來,我一直潛心研究這個領域, 就在一年半前,我又發表了一部新的著作 ——《心智的構建》。 該專著探討了同一個課題, 幸運的是,我現在擁有充足的證據支撐。 神經科學為我們貢獻 大量有關大腦的數據, 還在以逐年翻倍的速度劇增; 各種腦部掃描技術的空間解析度, 也在逐年翻倍。 現在,我們能親眼窺見活體大腦的內部, 觀察單個神經間的連接, 目睹神經連接、觸發的實時發生。 我們親眼看到大腦如何創造思維, 或者反過來說,思維如何增強和促進大腦, 思維本身對大腦進化至關重要。
So let me describe briefly how it works. I've actually counted these modules. We have about 300 million of them, and we create them in these hierarchies. I'll give you a simple example. I've got a bunch of modules that can recognize the crossbar to a capital A, and that's all they care about. A beautiful song can play, a pretty girl could walk by, they don't care, but they see a crossbar to a capital A, they get very excited and they say "crossbar," and they put out a high probability on their output axon. That goes to the next level, and these layers are organized in conceptual levels. Each is more abstract than the next one, so the next one might say "capital A." That goes up to a higher level that might say "Apple." Information flows down also. If the apple recognizer has seen A-P-P-L, it'll think to itself, "Hmm, I think an E is probably likely," and it'll send a signal down to all the E recognizers saying, "Be on the lookout for an E, I think one might be coming." The E recognizers will lower their threshold and they see some sloppy thing, could be an E. Ordinarily you wouldn't think so, but we're expecting an E, it's good enough, and yeah, I've seen an E, and then apple says, "Yeah, I've seen an Apple."
接下來,我想簡單介紹大腦的工作方式。 實際上,我統計過這些模塊的數量。 我們總共有大約三億模塊, 分佈在不同的層級中。 讓我們來看一個簡單的例子。 假設我有一組模塊, 可以識別大寫字母“A”中間的短橫線, 它們的主要職責就在於此。 無論周遭播放著美妙的音樂, 還是一位妙齡女郎翩然而至, 它們都渾然不覺。但是,一旦發現“A”的短橫線, 它們就興奮異常,異口同聲喊出:“短橫線!” 同時,它們立即報告神經軸突, 識別任務已經順利完成。 接下來,更高級別的模塊—— 概念級別的模塊,將依次登場。 級別越高,思考的抽象程度越高。 例如,較低的級別可識別字母“A”, 逐級上升後,某個級別能識別“APPLE”這個單詞。 同時,信息也在持續傳遞。 負責識別“APPLE”的級別,發現A-P-P-L時, 它會想:“唔,我猜下一個字母應該是E吧。” 然後,它會將信號傳達到 負責識別“E”的那些模塊, 並發出預警:“嘿,各位注意, 字母E就要出現了!” 字母“E”的識別模塊於是降低了閥值, 一旦發現疑似字母,便認為是“E”。 當然,這並非通常情況下的處理機制, 但現在我們正在等待“E”的出現, 而疑似字母與它足夠相似, 所以,我們斷定它就是“E”。 “E”識別後,“APPLE”識別成功。
Go up another five levels, and you're now at a pretty high level of this hierarchy, and stretch down into the different senses, and you may have a module that sees a certain fabric, hears a certain voice quality, smells a certain perfume, and will say, "My wife has entered the room."
如果我們再躍升五個級別, 那麼,在整個層級結構上, 就到達了較高水平。 這個水平上,我們具有各種感知功能, 某些模塊能夠感知特定的布料質地, 辨識特定的音色,甚至嗅到特定的香水味, 然後告诉我:妻子剛進到房间!
Go up another 10 levels, and now you're at a very high level. You're probably in the frontal cortex, and you'll have modules that say, "That was ironic. That's funny. She's pretty."
再上升10級, 我們就到達了一個很高的水平, 可能來到了額葉皮層。 在這兒,我們的模塊已經能夠臧否人物了, 比如:這事有點滑稽可笑!她真是秀色可餐!
You might think that those are more sophisticated, but actually what's more complicated is the hierarchy beneath them. There was a 16-year-old girl, she had brain surgery, and she was conscious because the surgeons wanted to talk to her. You can do that because there's no pain receptors in the brain. And whenever they stimulated particular, very small points on her neocortex, shown here in red, she would laugh. So at first they thought they were triggering some kind of laugh reflex, but no, they quickly realized they had found the points in her neocortex that detect humor, and she just found everything hilarious whenever they stimulated these points. "You guys are so funny just standing around," was the typical comment, and they weren't funny, not while doing surgery.
大家可能覺得,這整個過程有點複雜。 實際上,更讓人費解的是 是這些過程的層級結構。 曾經有位16歲的姑娘,當時正接受腦部手術。 由於手術過程中醫生需要跟她講話, 所以就讓她保持清醒。 保持清醒的意識,這對於手術並無妨礙, 因為大腦內沒有痛覺感受器。 我們驚奇地發現,當醫生刺激新皮層上 某些細小區域時,就是圖中的紅色部位, 這個姑娘就會放聲大笑。 起初,大家以為, 可能是因為觸發了笑反應神經。 他們很快意識到事實並非如此, 這些新皮層上的特定區域能夠理會幽默, 只要醫生刺激這些區域, 她就會覺得所有的一切都滑稽有趣。 “你們這幫人光站在那裏,就讓人想笑。” 那位姑娘典型的解釋道。 我們知道,這個場景並不滑稽可笑, 因為大家都在進行緊張的手術。
So how are we doing today? Well, computers are actually beginning to master human language with techniques that are similar to the neocortex. I actually described the algorithm, which is similar to something called a hierarchical hidden Markov model, something I've worked on since the '90s. "Jeopardy" is a very broad natural language game, and Watson got a higher score than the best two players combined. It got this query correct: "A long, tiresome speech delivered by a frothy pie topping," and it quickly responded, "What is a meringue harangue?" And Jennings and the other guy didn't get that. It's a pretty sophisticated example of computers actually understanding human language, and it actually got its knowledge by reading Wikipedia and several other encyclopedias.
現在,我們又有哪些新的進展呢? 計算機日益智能化, 利用功能類似新皮層的先進技術, 它們可以學習和掌握人類的語言。 我曾描述過一種算法, 與層級隱含式馬爾可夫模型類似, (馬爾可夫模型是用於自然語言處理的統計模型) 上世紀90年以來我一直研究這種算法。 “Jeopardy”(危境)是一個 自然語言類的智力競賽節目, IBM研發的沃森計算機在比賽中 勇奪高分,總分超過兩名最佳選手的總和。 連這個難題都被它輕鬆化解了: “定義:由起泡的派餡料發表的冗長而乏味的演講。 請問:這定義的是什麼?” 它迅速回答道:愛開腔的蛋白霜。 而詹尼斯和另外一名選手卻一頭霧水。 這個問題難度很大,極富挑戰性, 向我們展示了計算機 正在掌握人類的語言。 實際上,沃森是通過廣泛閱讀維基百科 及其他百科全書來發展語言能力的。
Five to 10 years from now, search engines will actually be based on not just looking for combinations of words and links but actually understanding, reading for understanding the billions of pages on the web and in books. So you'll be walking along, and Google will pop up and say, "You know, Mary, you expressed concern to me a month ago that your glutathione supplement wasn't getting past the blood-brain barrier. Well, new research just came out 13 seconds ago that shows a whole new approach to that and a new way to take glutathione. Let me summarize it for you."
5至10年以後, 我們的搜索引擎 不再只是搜索詞語和鏈接這樣的簡單組合, 它會嘗試去理解信息, 通過涉獵浩如煙海的互聯網和書籍, 攫取和提煉知識。 想像有一天,你正在悠閒地散步, 智能設備端的 Google 助理突然和你說: “瑪麗,你上月提到,正在服用的谷胱甘肽補充劑 因為無法透過血腦屏障,所以暫時不起作用。 告訴你一個好消息!就在13秒鐘前, 一項新的研究成果表明, 可以透過一个新的途徑來補充谷胱甘肽。 讓我給你概括一下這個報告。”
Twenty years from now, we'll have nanobots, because another exponential trend is the shrinking of technology. They'll go into our brain through the capillaries and basically connect our neocortex to a synthetic neocortex in the cloud providing an extension of our neocortex. Now today, I mean, you have a computer in your phone, but if you need 10,000 computers for a few seconds to do a complex search, you can access that for a second or two in the cloud. In the 2030s, if you need some extra neocortex, you'll be able to connect to that in the cloud directly from your brain. So I'm walking along and I say, "Oh, there's Chris Anderson. He's coming my way. I'd better think of something clever to say. I've got three seconds. My 300 million modules in my neocortex isn't going to cut it. I need a billion more." I'll be able to access that in the cloud. And our thinking, then, will be a hybrid of biological and non-biological thinking, but the non-biological portion is subject to my law of accelerating returns. It will grow exponentially. And remember what happens the last time we expanded our neocortex? That was two million years ago when we became humanoids and developed these large foreheads. Other primates have a slanted brow. They don't have the frontal cortex. But the frontal cortex is not really qualitatively different. It's a quantitative expansion of neocortex, but that additional quantity of thinking was the enabling factor for us to take a qualitative leap and invent language and art and science and technology and TED conferences. No other species has done that.
20年以後,我們將迎來奈米機器人, 目前,科技產品正在日益微型化, 這一趨勢愈演愈烈。 科技設備將通過毛細血管 進入我們的大腦, 最終,將我們自身的新皮層 與雲端的人工合成新皮層相連, 使它成為新皮層的延伸和擴展。 今天, 智慧型手機都內置了一台計算機。 假如我們需要一萬台計算機, 在幾秒鐘內完成一次複雜的搜索, 我們可以通過訪問雲端來獲得這種能力。 到了2030年,當你需要更加強大的新皮層時, 你可以直接從你的大腦連接到雲端, 來獲得超凡的能力。 舉個例子,我正在漫步,遠遠看到一個人。 “老天,那不是克里斯.安德森(TED主持人)嗎? 他正朝我這邊走來。 我要抓住這個機遇,一鳴驚人! 但是,我只有三秒鐘, 我新皮層的三億個模塊 顯然不夠用。 我需要借來10億模塊增援!” 於是,我會立即連通雲端。 我的思考,綜合了生物體和非生物體 這兩者的優勢。 非生物部分的思考能力, 將受益於“加速回報定律”, 這是說,科技帶來的回報 呈指數級增長,而非線性。 大家是否還記得,上次新皮層大幅擴張時 發生了哪些重大變化? 那是200萬年前, 我們那時還只是猿人, 開始發育出碩大的前額。 而其他靈長類動物的前額向後傾斜, 因為牠們沒有額葉皮層。 但是,額葉皮層並不意味著質的變化; 而是新皮層量的提升, 帶來了額外的思考能力, 最終促成了質的飛躍。 我們因而能夠發明語言, 創造藝術,發展科技, 並舉辦TED演講, 這都是其他物種難以完成的創舉。
And so, over the next few decades, we're going to do it again. We're going to again expand our neocortex, only this time we won't be limited by a fixed architecture of enclosure. It'll be expanded without limit. That additional quantity will again be the enabling factor for another qualitative leap in culture and technology.
我相信未來數十年, 我們將再次創造偉大的奇蹟。 我們將借助科技,再次擴張新皮層, 不同之處在於, 我們將不再受到頭顱空間的局限, 意味著擴張並無止境。 隨之而來的量的增加 在人文和科技領域, 將再次引發一輪質的飛躍。
Thank you very much.
謝謝大家!
(Applause)
(掌聲)