從感知智能邁向認知智能，百分點科技找到了一條跑得通的AI落地新範式

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"“隨着應用場景的複雜化以及需求越來越要求落地，所有的技術，包括大數據、人工智能、數字孿生等等，最終都需要深度融合，才能更好的滿足應用場景的實際需求。而與此同時，數據從生產、治理到形成知識的轉換過程及應用中也在發生多重‘智變’”，百分點科技CTO劉譯璟表示，業界已經感受到，大數據與人工智能技術已經有了非常強烈的融合需求。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/13\/22\/13fe7352c6c76360757710c848c12f22.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"size","attrs":{"size":10}}],"text":"百分點科技CTO劉譯璟"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"近日，在由百分點科技舉辦的數據智能技術實踐論壇上，百度視覺技術部主任架構師劉經拓、百分點科技數據開發部負責人馬偉凱、復旦大學教授肖仰華和百分點科技首席算法科學家蘇海波等專家就這一趨勢展開了探討，他們分別從感知、數據治理、認知各個層面詳細解讀了數據到知識是如何一步步地發生變化並支持智能應用的。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"從數據“治理”到數據“智理”"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"要從單點算法創新轉向全流程、全棧式創新"}]},{"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":"近年來，伴隨數字政府的快速落地，無論是數據的生產環節還是治理環節，都對數據智能技術應用提出了越來越高的需求。"}]},{"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":"計算機視覺作爲在感知層的核心技術，在結合城市感知、工業能源、人機交互、視頻生產等業務場景中得到了不斷髮展。"}]},{"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":"同時，數字政府項目中還存在衆多的文本、語音等形式複雜的非結構化數據，如何基於視覺技術及政府已經有的業務數據進行多模態數據的融合，以及運用智能化技術的實現高效的“智”理，對支撐數字政府智能應用、提升治理水平起着至關重要的作用。"}]},{"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":"百度視覺技術部主任架構師劉經拓表示，隨着行業應用的逐步深入，通過某個單點算法創新就能大大提升應用效果的方式已不復存在，需要用全流程、全棧的方式。"}]},{"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":"因此，爲了實現更好的技術發展和落地效果，百度在技術和產業兩個維度上進行了融合創新，提出了計算機視覺領域從預訓練、定製化到小型化，以及平臺化的一體化的研發方案。"}]},{"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":"具體來說，首先通過用超大規模非結構性的數據做預訓練，能夠在同樣少量數據的情況下取得更好的效果；定製化是指結合特定場景進行自適應遷移學習，根據場景數據進行有監督定製化調優；小型化主要是基於模型剪枝、蒸餾、量化等技術，對模型進行裁剪。"}]},{"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":"他表示，大規模有監督數據的技術紅利逐漸減弱，AI新基建需要更低的研發與部署成本，通過預訓練與自訓練平臺，最終還要沉澱成標準化、低成本複製的模型，並與產業進行更深度的融合，挖掘出更多降低人工成本的新應用點。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"數據治理的“PAI”實施方法論"}]},{"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":"百分點科技數據開發部負責人馬偉凱在分享中表示，數據治理中的技術挑戰除了數據孤島、數據質量等問題，還存在多模態下如何解決數據不可用、不能用及不好用等挑戰。"}]},{"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":"他介紹，基於多源異構數據源的複雜場景下數字政府數據治理項目，百分點科技提出了一套數據治理“PAI”實施方法論，即流程化（process-oriented）、自動化（automation）、智能化（intelligence）。通過引入機器學習算法、NLP等數據智能技術，可以更好地開展數據治理工作，建立全域數據標準、提升數據質量、盤活數據資產，從而支撐數據融通，最終釋放數據價值指導業務創新。"}]},{"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":"馬偉凱認爲，未來數據治理領域會有四大應用趨勢："}]},{"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":"一是智能化建模和數據加工優化，需要不斷地提升建模的能力沉澱到知識庫；"}]},{"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":"二是智能化完善數據安全管理，包括智能化控制數據權限分配、智能化數據審計並制定數據加密脫敏策略；"}]},{"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":"三是智能化設計並維護數據生命週期管理，由於數據量大有時候不一定是好事，對於很多單位來說可能是負擔，因此智能化識別數據、智能化制定數據保留策略非常重要；"}]},{"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":"四是與Data Fabric更好的融合，增強數據目錄，實現動態的獲取數據，保證數據的安全。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"數據到知識的“知”變與“智”變"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"復旦大學肖仰華：認知智能將是知識圖譜等知識工程技術發展的必然歸宿"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在實踐中，行業客戶對知識圖譜的應用訴求愈發強烈，核心需求是將行業數據知識化，並通過搜索、推薦、問答，以及用知識輔助進行更加智能的決策。"}]},{"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":"因此，將結構化和非結構化的數據有效的“智”理起來後，能夠進行數據和知識挖掘，提取當中有價值的信息，並以可視化的分析，爲政府和企業決策提供支持成爲當今亟待解決的問題。"}]},{"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":"復旦大學肖仰華教授在主題分享中表示，作爲大數據知識工程的典型代表，知識圖譜技術近年來取得了長足進步，並在一系列實際應用中取得了顯著效果。知識圖譜之所以備受關注是因爲業界普遍認爲知識圖譜是實現機器認知智能的基礎。"}]},{"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":"但隨着應用的深化，知識圖譜的落地過程單靠其所代表的知識智能本身這套技術體系和範式已經難以解決很多問題：一是數據獲取和治理困難；二是在知識層面，小樣本、低資源情況下知識的表示和獲取代價仍然非常大；此外，獲取知識之後，在應用、服務能力方面也存在很多挑戰。"}]},{"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":"因此，未來破題的關鍵在於要突破以知識圖譜爲代表的知識智能的邊界，向認知智能這樣的智能新形態發展。他認爲，認知智能作爲數據智能、知識智能融合創新產物，將是知識圖譜等知識工程技術發展的必然歸宿。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"百分點首席算法科學家蘇海波："},{"type":"text","marks":[{"type":"strong"}],"text":"未來知識圖譜一定會深入到各行各業"}]},{"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":"百分點科技首席算法科學家蘇海波認爲，近些年，人工智能逐漸從感知智能向認知智能發展，知識圖譜則是實現認知智能的關鍵技術方法，在構建出知識圖譜後，可以實現各種智能場景應用。"}]},{"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":"蘇海波以此爲背景介紹了百分點科技知識圖譜抽取、知識圖譜問答等技術的算法原理和實踐，以及在多個行業場景中的具體應用。"}]},{"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":"他強調，"},{"type":"text","marks":[{"type":"strong"}],"text":"未來知識圖譜一定會深入到各行各業，只有掌握通用的人工智能技術，並將技術和業務需求對應起來，才能真正發揮出知識圖譜的價值，解決行業問題。"}]},{"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":"目前，半自動化結合人工是業內構建知識圖譜所採用的主流方式。蘇海波對InfoQ表示，從長遠來看，完全靠機器自動化，一點都不投入人工，目前不現實，也不可能存在。“現在有很多知識圖譜構建工程化的工具，在解決如何高效地抽取實體關係，如何做出映射、如何融合，以及如何通過預訓練模型減少需要標註數據的數量等問題方面，只能說隨着技術的發展和工具的發展，人工的工作量會逐漸降低，人工的效率會越來越高。但到什麼時候，採用機器構建的比例比人工構建更多，我覺得現在還不好衡量，這是一個逐漸發展的過程”。"}]},{"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":"劉譯璟認爲，“人在閉環”是認知智能行業落地的必由之路，即在知識圖譜構建和應用的過程中，人必須參與。必須要有人在，這是一個責任問題。機器適合做數據密集型和經驗密集型的工作。而人適合做價值判斷型或情感密集型的工作。“我們任何一個在現實中有意義的業務，它的價值一定來自於人。如果沒有人的話，這個東西是沒有價值的，所以不可能離開人”。"}]},{"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":"蘇海波表示，當前，已經進入到一個從數據到知識的“智變”時代，隨着大數據、知識圖譜、NLP等數據智能技術的進一步成熟，數據中的價值將不斷被挖掘利用，幫助人們做出合理決策。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"認知智能落地的新範式"}]},{"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":"人工智能從感知智能向認知智能發展的趨勢日益明顯。從感知智能到認知智能怎麼跨越，認知智能如何應用，怎樣更好的落地....都是需要解決的問題。"}]},{"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":"劉譯璟表示，早在2016年，百分點科技就開始打造從感知、認知、決策到行動的端到端的智能決策閉環，如今經過了5年的探索和實踐，在一些行業，整個閉環已經運行起來了。"}]},{"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":"例如在應急管理行業，百分點科技已經驗證了一種認知智能行業落地的新範式。"}]},{"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":"劉譯璟將這種範式稱之爲“符號主義引導下的多方法融合技術”。它不是純粹的用深度學習，採集大量的數據訓練模型，而是先構建出某一領域裏業務的本體框架，而後再結合深度學習算法和樣例數據細化知識圖譜的本體和事實數據，並根據知識圖譜應用中的反饋信息不斷調整和優化。"}]},{"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":"“AI如何在行業中落地，目前業界還沒有探索出一條特別靠譜的路徑，我們覺得自己找到了這條路徑：基於符號主義引導，先定義一個行業框架；再融合連接主義和行爲主義，在應用中不斷完善行業知識”，劉譯璟進一步解釋說，由於“AI的三大流派 —符號主義、連接主義、行爲主義，用這三種方法去實現通用AI，能組合出25種方法。而百分點團隊發現，基於符號主義做引導，再結合連接主義和行爲主義的方法，是比較好的能在行業裏落地，且能夠真正產生應用價值的方案。”"}]},{"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":"目前，這一新範式正在推廣過程中，百分點科技希望複製到更多行業，實現規模化應用。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"AI行業發展應該克服浮躁心態"}]},{"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":"在接受InfoQ等少數媒體採訪時，劉譯璟對AI這幾年的發展，尤其針對一些亂象發表了自己的看法。"}]},{"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":"“我總體覺得AI產業略顯浮躁，大家似乎總想着去造新概念、捧新概念，而很少關注這些概念（技術）已經用到什麼程度了，應該如何把物質生產力提高上來。很多新概念被過分炒作後，最後就變成了泡沫”。"}]},{"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":"“正兒八經地把它用起來纔是第一位的”，劉譯璟認爲，憑空炒概念沒有意義，如何把這些技術真正應用到實際的場景裏，真正產生價值，纔是最重要的事。"}]},{"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":"從另一個角度看，這種現象也說明，這幾年，AI技術在理論上少有大的突破。深度學習理論早在80年代就有了，發展至今並沒有本質的改變，不過是數據多了，算力強了，更容易爲人們所應用了。"}]},{"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":"劉譯璟覺得，通用人工智能的突破，需要在理論、技術和數據三方面展開。理論方面，腦科學、可解釋AI、大模型小數據方法、量子機器學習等研究內容，會是未來的探索重點；技術方面，三大流派的技術融合和應用會是5年內的熱點。長期來看，AI能力的提升需要量子計算的應用，量子機器學習算法和量子搜索算法將是算力的新增長點；數據方面，ImageNet數據集曾加速了整個CV的發展速度，未來類似的數據集會越來越多，尤其是常識類的知識會越來越標準和易得，這一定能推動通用人工智能的發展。"}]}]}