從技術到科學，中國AI向何處去？

{"type":"doc","content":[{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"引言"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果從達特茅斯會議起算，AI已經走過65年曆程，尤其是近些年深度學習興起後，AI迎來了空前未有的繁榮。不過，最近兩年中國AI熱潮似乎有所回落，在理論突破和落地應用上都遇到了挑戰，外界不乏批評質疑的聲音，甚至連一些AI從業者也有些沮喪。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從90年代到美國卡耐基梅隆大學讀博開始，我有幸成爲一名AI研究者，見證了這個領域的一些起伏。通過這篇文章，我將試圖通過個人視角回顧AI的發展，審視我們當下所處的歷史階段，以及探索AI的未來究竟在哪裏。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"本文部分觀點："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"1、 AI時代序幕剛拉開，AI目前還處於初級階段，猶如法拉第剛剛發現了交流電，還未能從技術上升爲科學。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2、以深度學習爲代表的AI研究這幾年取得了諸多令人讚歎的進步，但部分也是運氣的結果，其真正原理迄今無人知曉。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"3、在遇到瓶頸後，深度學習有三個可能突破方向：深度學習的根本理解、自監督學習和小樣本學習、知識與數據的有機融合。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"4、AI在當下最大的機會：用AI解決科學重要難題（AI for Science）。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"一、AI的歷史階段：手工作坊"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雖然有人把當下歸爲第三波甚至是第四波AI浪潮，樂觀地認爲AI時代已經到來，但我的看法要謹慎一些：AI無疑具有巨大潛力，但就目前我們的能力，AI尚處於比較初級的階段，是技術而非科學。這不僅是中國AI的問題，也是全球AI共同面臨的難題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這幾年深度學習的快速發展，極大改變了AI行業的面貌，讓AI成爲公衆日常使用的技術，甚至還出現了一些令公衆驚奇的AI應用案例，讓人誤以爲科幻電影即將變成現實。但實際上，技術發展需要長期積累，目前只是AI的初級階段，AI時代纔剛開始。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果將AI時代和電氣時代類比，今天我們的AI技術還是法拉第時代的電。法拉第通過發現電磁感應現象，從而研製出人類第一臺交流電發電機原型，不可謂不偉大。法拉第這批先行者，實踐經驗豐富，通過大量觀察和反覆實驗，手工做出了各種新產品，但他們只是拉開了電氣時代的序幕。電氣時代的真正大發展，很大程度上受益於電磁場理論的提出。麥克斯維爾把實踐的經驗變成科學的理論，提出和證明了具有跨時代意義的麥克斯維爾方程。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果人們對電磁的理解停留在法拉第的層次，電氣革命是不可能發生的。試想一下，如果颳風下雨打雷甚至連溫度變化都會導致斷電，電怎麼可能變成一個普惠性的產品，怎麼可能變成社會基礎設施？又怎麼可能出現各種各樣的電氣產品、電子產品、通訊產品，徹底改變我們的生活方式？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這也是AI目前面臨的問題，侷限於特定的場景、特定的數據。AI模型一旦走出實驗室，受到現實世界的干擾和挑戰就時常失效，魯棒性不夠；一旦換一個場景，我們就需要重新深度定製算法進行適配，費時費力，難以規模化推廣，泛化能力較爲有限。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這是因爲今天的AI很大程度上是基於經驗。AI工程師就像當年的法拉第，能夠做出一些AI產品，但都是知其然，不知其所以然，還未能掌握其中的核心原理。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"那爲何AI迄今未能成爲一門科學？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"答案是，技術發展之緩慢遠超我們的想象。回顧90年代至今這二十多年來，我們看到的更多是AI應用工程上的快速進步，核心技術和核心問題的突破相對有限。一些技術看起來是這幾年興起的，實際上早已存在。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"以自動駕駛爲例，美國卡耐基梅隆大學的研究人員進行的Alvinn項目，在80年代末已經開始用神經網絡來實現自動駕駛，1995年成功自東向西穿越美國，歷時7天，行駛近3000英里。在下棋方面，1992年IBM研究人員開發的TD-Gammon，和AlphaZero相似，能夠自我學習和強化，達到了雙陸棋領域的大師水平。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/0a\/0a5f57fccfdc6cba1ad9f98f35d7c8c0.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"（1995年穿越美國項目開始之前的團隊合照）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"不過，由於數據和算力的限制，這些研究只是點狀發生，沒有形成規模，自然也沒有引起大衆的廣泛討論。今天由於商業的普及、算力的增強、數據的方便獲取、應用門檻的降低，AI開始觸手可及。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但核心思想並沒有根本性的變化。我們都是試圖用有限樣本來實現函數近似從而描述這個世界，有一個input，再有一個output，我們把AI的學習過程想象成一個函數的近似過程，包括我們的整個算法及訓練過程，如梯度下降、梯度回傳等。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"同樣的，核心問題也沒有得到有效解決。90年代學界就在問的核心問題，迄今都未得到回答，他們都和神經網絡、深度學習密切相關。比如非凸函數的優化問題，它得到的解很可能是局部最優解，並非全局最優，訓練時可能都無法收斂，有限數據還會帶來泛化不足的問題。我們會不會被這個解帶偏了，忽視了更多的可能性？"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"二、深度學習：大繁榮後遭遇發展瓶頸"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"毋庸諱言，以深度學習爲代表的AI研究這幾年取得了諸多令人讚歎的進步，比如在複雜網絡的訓練方面，產生了兩個特別成功的網絡結構，CNN和transformer。基於深度學習，AI研究者在語音、語義、視覺等各個領域都實現了快速的發展，解決了諸多現實難題，實現了巨大的社會價值。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但回過頭來看深度學習的發展，不得不感慨AI從業者非常幸運。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"首先是隨機梯度下降（SGD），極大推動了深度學習的發展。隨機梯度下降其實是一個很簡單的方法，具有較大侷限性，在優化裏面屬於收斂較慢的方法，但它偏偏在深度網絡中表現很好，而且還是出奇的好。爲什麼會這麼好？迄今研究者都沒有完美的答案。類似這樣難以理解的好運氣還包括殘差網絡、知識蒸餾、Batch Normalization、Warmup、Label Smoothing、Gradient Clip、Layer Scaling…尤其是有些還具有超強的泛化能力，能用在多個場景中。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"再者，在機器學習裏，研究者一直在警惕過擬合（overfitting）的問題。當參數特別多時，一條曲線能夠把所有的點都擬合得特別好，它大概率存在問題，但在深度學習裏面這似乎不再成爲一個問題…雖然有很多研究者對此進行了探討，但目前還有沒有明確答案。更加令人驚訝的是，我們即使給數據一個隨機的標籤，它也可以完美擬合（請見下圖紅色曲線），最後得出擬合誤差爲0。如果按照標準理論來說，這意味着這個模型沒有任何偏差（bias），能幫我們解釋任何結果。請想想看，任何東西都能解釋的模型，真的可靠嗎，包治百病的良藥可信嗎？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/84\/84c2c9f094373eefa46e2340fd033687.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"（Understanding deep learning requires rethinking generalization. ICLR, 2017.）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"說到這裏，讓我們整體回顧下機器學習的發展歷程，才能更好理解當下的深度學習。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"機器學習有幾波發展浪潮，在上世紀80年代到90年代，首先是基於規則（rule based）。從90年代到2000年代，以神經網絡爲主，大家發現神經網絡可以做一些不錯的事情，但是它有許多基礎的問題沒回答。所以2000年代以後，有一批人嘗試去解決這些基礎問題，最有名的叫SVM（support vector machine），一批數學背景出身的研究者集中去理解機器學習的過程，學習最基礎的數學問題,如何更好實現函數的近似，如何保證快速收斂，如何保證它的泛化性？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"那時候，研究者非常強調理解，好的結果應該是來自於我們對它的深刻理解。研究者會非常在乎有沒有好的理論基礎，因爲要對算法做好的分析，需要先對泛函分析、優化理論有深刻的理解，接着還要再做泛化理論…大概這幾項都得非常好了，纔可能在機器學習領域有發言權，否則連文章都看不懂。如果研究者自己要做一個大規模實驗系統，特別是分佈式的，還需要有工程的豐富經驗，否則根本做不了，那時候沒有太多現成的東西，更多隻是理論，多數工程實現需要靠自己去跑。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但是深度學習時代，有人做出了非常好的框架，便利了所有的研究者，降低了門檻，這真是非常了不起的事情，促進了行業的快速發展。今天去做深度學習，有個好想法就可以幹，只要寫上幾十行、甚至十幾行代碼就可以跑起來。成千上萬人在實驗各種各樣的新項目，驗證各種各樣新想法，經常會冒出來非常讓人驚喜的結果。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但我們可能需要意識到，時至今日，深度學習已遇到了很大的瓶頸。那些曾經幫助深度學習成功的好運氣，那些無法理解的黑盒效應，今天已成爲它進一步發展的桎梏。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"三、下一代AI的三個可能方向"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"AI的未來究竟在哪裏？下一代AI將是什麼？目前很難給出明確答案，但我認爲，至少有三個方向值得重點探索和突破。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"第一個方向是尋求對深度學習的根本理解"},{"type":"text","text":"，破除目前的黑盒狀態，只有這樣AI纔有可能成爲一門科學。具體來說，應該包括對以下關鍵問題的突破："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對基於DNN函數空間的更全面刻畫；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對SGD（或更廣義的一階優化算法）的理解；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"重新考慮泛化理論的基礎。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"第二個方向是知識和數據的有機融合。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人類在做大量決定時，不僅使用數據，而且大量使用知識。如果我們的AI能夠把知識結構有機融入，成爲重要組成部分，AI勢必有突破性的發展。研究者已經在做知識圖譜等工作，但需要進一步解決知識和數據的有機結合，探索出可用的框架。之前曾有些創新性的嘗試，比如Markov Logic，就是把邏輯和基礎理論結合起來，形成了一些有趣的結構。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"第三個重要方向是自監督學習和小樣本學習。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我雖然列將這個列在第三，但卻是目前值得重點推進的方向，它可以彌補AI和人類智能之間的差距。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"今天我們經常聽說AI在一些能力上可以超越人類，比如語音識別、圖像識別，最近達摩院AliceMind在視覺問答上的得分也首次超過人類，但這並不意味着AI比人類更智能。谷歌2019年有篇論文"},{"type":"text","marks":[{"type":"italic"}],"text":"on the Measure of intelligence"},{"type":"text","text":"非常有洞察力，核心觀點是說，真正的智能不僅要具有高超的技能，更重要的是能否快速學習、快速適應或者快速通用？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"按照這個觀點，目前AI是遠不如人類的，雖然它可能在一些方面的精度超越人類，但可用範圍非常有限。這裏的根本原因在於：人類只需要很小的學習成本就能快速達到結果，聰明的人更是如此——這也是我認爲目前AI和人類的主要區別之一。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有一個很簡單的事實證明AI不如人類智能，以翻譯爲例，現在好的翻譯模型至少要億級的數據。如果一本書大概是十幾萬字，AI大概要讀上萬本書。我們很難想象一個人爲了學習一門語言需要讀上萬本書。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"另外有意思的對比是神經網絡結構和人腦。目前AI非常強調深度，神經網絡經常幾十層甚至上百層，但我們看人類，以視覺爲例，視覺神經網絡總共就四層，非常高效。而且人腦還非常低功耗，只有20瓦左右，但今天GPU基本都是數百瓦，差了一個數量級。著名的GPT-3跑一次，碳排放相當於一架747飛機從美國東海岸到西海岸往返三次。再看信息編碼，人腦是以時間序列來編，AI是用張量和向量來表達。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"也許有人說，AI發展不必一定向人腦智能的方向發展。我也認爲這個觀點不無道理，但在AI遇到瓶頸，也找不到其他參照物時，參考人腦智能可能會給我們一些啓發。比如，拿人腦智能來做對比，今天的深度神經網絡是不是最合理的方向？今天的編碼方式是不是最合理的？這些都是我們今天AI的基礎，但它們是好的基礎嗎？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"應該說，以GPT-3爲代表的大模型，可能也是深度學習的一個突破方向，能夠在一定程度上實現自學習。大模型有些像之前惡補了所有能看到的東西，碰到一個新場景，就不需要太多新數據。但這是一個最好的解決辦法嗎？我們目前還不知道。還是以翻譯爲例，很難想象一個人需要裝這麼多東西才能掌握一門外語。大模型現在都是百億、千億參數規模起步，沒有一個人類會帶着這麼多數據。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"所以，也許我們還需要繼續探索。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"四、AI的機會：AI for Science"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"說到這裏，也許有些人會失望。既然我們AI還未解決上面的三個難題，AI還未成爲科學，那AI還有什麼價值？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"技術本身就擁有巨大價值，像互聯網就徹底重塑了我們的工作和生活。AI作爲一門技術，當下一個巨大的機會就是幫助解決科學重點難題（AI for Science）。AlphaFold已經給了我們一個很好的示範，AI解決了生物學裏困擾半個世紀的蛋白質摺疊難題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們要學習AlphaFold，但沒必要崇拜。AlphaFold的示範意義在於，DeepMind在選題上真是非常厲害，他們選擇了一些今天已經有足夠的基礎和數據積累、有可能突破的難題，然後建設一個當下最好的團隊，下決心去攻克。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們有可能創造比AlphaFold更重要的成果，因爲在自然科學領域，有着很多重要的open questions，AI還有更大的機會，可以去發掘新材料、發現晶體結構，甚至去證明或發現定理…AI可顛覆傳統的研究方法，甚至改寫歷史。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"比如現在一些物理學家正在思考，能否用AI重新發現物理定律？過去數百年來，物理學定律的發現都是依賴天才，愛因斯坦發現了廣義相對論和狹義相對論，海森堡、薛定諤等人開創了量子力學，這些都是個人行爲。如果沒有這些天才，很多領域的發展會推遲幾十年甚至上百年。但今天，隨着數據越來越多，科學規律越來越複雜，我們是不是可以依靠AI來推導出物理定律，而不再依賴一兩個天才？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"以量子力學爲例，最核心的是薛定諤方程，它是由天才物理學家推導出來的。但現在，已有物理學家通過收集到的大量數據，用AI自動推導出其中規律，甚至還發現了薛定諤方程的另外一個寫法。這真的是一件非常了不起、有可能改變物理學甚至人類未來的事情。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們正在推進的AI EARTH項目，是將AI引入氣象領域。天氣預報已有上百年曆史，是一個非常重大和複雜的科學問題，需要超級計算機才能完成複雜計算，不僅消耗大量資源而且還不是特別準確。我們今天是不是可以用AI來解決這個問題，讓天氣預報變得既高效又準確？如果能成功，將是一件非常振奮人心的事情。當然，這注定是一個非常艱難的過程，需要時間和決心。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"五、AI從業者：多一點興趣，少一點功利"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"AI的當下局面，是對我們所有AI研究者的考驗。不管是AI的基礎理論突破，還是AI去解決科學問題，都不是一蹴而就的事情，需要研究者們既聰明又堅定。如果不聰明，不可能在不確定的未來抓住機會；如果不堅定，很可能就被嚇倒了。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但更關鍵的是興趣驅動，而不是利益驅動，不能急功近利，這些年深度學習的繁榮，使得中國大量人才和資金湧入AI領域，快速推動了行業發展，但也催生了一些不切實際的期待。像DeepMind做了AlphaGo之後，中國一些人跟進複製，但對於核心基礎創新進步來說意義相對有限。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"既然AI還不是一門科學，我們要去探索沒人做過的事情，很有可能失敗。這意味着我們必須有真正的興趣，靠興趣和好奇心去驅動自己前行，才能扛過無數的失敗。我們也許看到了DeepMind做成了AlphaGo和AlphaFold兩個項目，但可能還有更多失敗的、無人聽聞的項目。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在興趣驅動方面，國外研究人員值得我們學習。像一些獲得圖靈獎的頂級科學家，天天還在一線做研究，親自推導理論。還記得在CMU讀書的時候，當時學校有多個圖靈獎得主，他們平常基本都穿梭在各種seminar（研討班）。我認識其中一個叫Manuel Blum，因爲密碼學研究獲得圖靈獎，有一次我參加一個seminar，發現Manuel Blum沒有座位，就坐在教室的臺階上。他自己也不介意坐哪裏，感興趣就來了，沒有座位就擠一擠。我曾有幸遇到過諾貝爾經濟學獎得主托馬斯·薩金特，作爲經濟學者，他早已功成名就，但他60歲開始學習廣義相對論，70歲開始學習深度學習，76歲還和我們這些晚輩討論深度學習的進展…也許這就是對研究的真正熱愛吧。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"說回國內，我們也不必妄自菲薄，中國AI在工程方面擁有全球領先的實力，承認AI還比較初級並非否定從業者的努力，而是提醒我們需要更堅定地長期努力，不必急於一時。電氣時代如果沒有法拉第這些先行者，沒有一個又一個的點狀發現，不可能總結出理論，讓人類邁入電氣時代。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"同樣，AI發展有賴於我們以重大創新爲憧憬，一天天努力，不斷嘗試新想法，然後纔會有一些小突破。當一些聰明的腦袋，能夠將這些點狀的突破聯結起來，總結出來理論，AI纔會產生重大突破，最終上升爲一門科學。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們已經半隻腳踏入AI時代的大門，這注定是一個比電氣時代更加輝煌、激動人心的時代，但這一切的前提，都有賴於所有研究者的堅定不移的努力。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"作者介紹："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"金榕 阿里巴巴達摩院副院長、原密歇根州立大學終身教授"}]}]}