曠視十年，回答AI技術價值躍遷的“靈魂”三問

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2021年，AI獨角獸 — 曠視科技迎來了成立10週年。"}]},{"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":"7月16日，在2021年曠視技術開放日（MegTech 2021）上，曠視首席科學家、曠視研究院院長孫劍，曠視研究院張祥雨、範浩強、周而進等多位研究員，分享了曠視十年來在AI技術上的實踐和思考，通過一系列技術演示，系統展示了曠視最新的技術成果。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"基礎科研是AI創新突破的基石"}]},{"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":"曠視首席科學家、曠視研究院院長孫劍總結了過去十年中國AI技術價值的發展歷程。"}]},{"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":"AI有沒有用？AI在哪裏用？AI易不易用？這是AI行業的“靈魂”三問，是所有AI企業需要回答的問題。"}]},{"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":"孫劍表示，人工智能最核心的動力是深度學習。在2011年 - 2012年，深度學習剛剛嶄露頭角，展示出它比傳統的方法有一定優勢。在這個時期，一個主要回答的問題是，以機器學習爲代表的AI到底是不是比傳統方法好很多？到底能不能帶來很大的價值？到底是不是有用？"}]},{"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":"因此從2011年開始，可以看到全球各個企業、高校都在投入大量的資金和人力研究AI在不同的應用上到底是不是有用。以曠視爲例，曠視在2012年曠視推出了Face++開放平臺，並在接下來的一兩年裏，使用深度學習第一次顯著超越了傳統的人臉檢測和識別方法。"}]},{"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":"有了AI 技術能解決很多問題，但AI應該用在哪裏？孫劍表示，從2014年-2015年開始，各行各業開始廣泛應用AI技術。自2015年開始，曠視陸續推出了在線身份認證、智能攝像頭、手機人臉解鎖、物流操作系統河圖、屏下指紋和攝影技術。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/ed\/f9\/edc4cc2399bf41efc013b0b04303dcf9.png","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","text":"曠視首席科學家、曠視研究院院長孫劍"}]},{"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發展的第三個階段是現在時。現在AI的應用可以說是無處不在，但在有些行業應用較深，有些行業應用較淺。下一個階段，AI如何有更深入、更廣泛的應用？它是不是好用，是不是易用的？門檻能否降低？"}]},{"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":"在第一階段、第二階段，把AI用得好，主要是靠經驗豐富的算法工程師深入理解問題，一個問題一個問題地去攻克，並沒有解決廣泛行業應用者不懂深度學習的技術，或者對怎麼用好沒有掌握得很精深的問題。"}]},{"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":"孫劍認爲，降低算法門檻，讓非資深的算法工程師也可以使用AI工具解決各行各業的問題。這樣才能回答AI好不好用、易不易用、是不是低成本的問題，讓AI廣泛得到應用，再向前推進。這也曠視推出其 AI 生產力平臺 Brain++的初衷。"}]},{"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有沒有用，歸根到底是看AI基礎科研有沒有突破。基礎科研是AI創新突破的基石，需要長期主義的堅持”，孫劍認爲。"}]},{"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":"曠視研究院張祥雨研究員詳細介紹了曠視在基礎科研領域的佈局。張祥雨介紹，曠視基礎科研主要包括三類：基礎模型、基礎算法，基礎應用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/b0\/e4\/b06e63bf63fe4ebd99fda82e6a51cde4.png","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","text":"曠視研究院張祥雨研究員"}]},{"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":"在基礎科研創新方面，成立10年來，曠視在在學術上發表85篇頂級會議論文，在競賽上獲40項冠軍，在實用上研發了包括ShuffleNets系列在內的衆多基礎模型，已經廣泛應用在手機等智能設備上，推動軟硬協同發展。"}]},{"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":"曠視現場展示了最新的基礎科研成果，包括面向服務器和下一代AI芯片的通用模型RepVGG系列，其在開源後Github star已超過1800；目前性能最強的端到端實時全卷積全景分割算法；以及挑戰經典，目標檢測架構創新的YOLOF。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"AI在哪裏用：從科研到落地"}]},{"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在哪裏用？"}]},{"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":"孫劍表示，要回答這個問題的關鍵詞是“行業落地”，“推動算法邊界拓展”，不斷推動各種各樣的算法在行業中得到深入運用。"}]},{"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":"曠視的技術主要有三大應用場景 —— 消費物聯網、城市物聯網和供應鏈物聯網。圖：曠視研究員張祥雨在曠視技術開放日上的分享AI是核心能力，IoT是落地場景，在數字世界與物理世界融合的智能化時代，AI算法正在爲產品帶來更多的價值提升。從明星產品神行面板機系列，到超越人眼的暗光拍照和暗光視頻，再到突破物理限制的屏下AI光學指紋和屏下攝像，以及追求全局最優解的河圖系統和機器人設備。"}]},{"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":"曠視研究院範浩強研究員認爲，業內存在的一些對於“人工智能和人工智障”的討論說明，一些算法並沒有達到用戶對智能產品的預期。只有生產優質的算法，不斷追求算法品質，才能夠實現物理世界中真實的產品的價值。這些價值體現更加魯棒、適應多環境、更加符合用戶預期等多方面。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/43\/c6\/4316d04dc6003c9d0e3aaacc0fc48bc6.png","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":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"曠視研究院範浩強研究員"}]},{"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":"曠視認爲，算法不只讓很多產品從好變更好，更能讓產品從不可能變可能。"}]},{"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":"“傳統意義上，很多做產品的人將算法定義成軟件的一部分，在硬件設計好之後再思考軟件該怎麼構建，在這個過程中，算法可能只佔最末端、最小的環節”，範浩強表示，“但在AI落地實踐中，曠視發現，傳統的產品算法思維是時候被顛覆了。現在，產品的易用度、特性很大程度上是由其算法所決定，需要在產品最早的定義、設計、研發全生命週期納入對算法的考量。也就是說，"},{"type":"text","marks":[{"type":"strong"}],"text":"算法不再僅僅是軟件物料的一部分，它正在成爲我們日常所使用產品中的核心組件"},{"type":"text","text":"”。"}]},{"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":"在曠視十年的AI實踐落地中發現，從科研到落地，算法正在創造越來越大的價值，也在開始重新定義軟硬件。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"算法生產能力決定算法價值上限"}]},{"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":"算法生產能力決定算法價值上限，也決定AI是不是真的更易用。"}]},{"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":"傳統的算法生產的過程很複雜，如果想在真實的生產環境中生產出有效的算法，往往要經過需求分析、數據處理、模型訓練、上線部署等環節。此外，一次訓練的算法並不能一次性解決實際場景中各種各樣的問題，需要進行算法或模型改進環節。在經歷多輪的常規算法生產過程中的迭代後，最終才能夠得到一個實際落地應用過程中符合期望和體驗的好模型或好算法。"}]},{"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","marks":[{"type":"strong"}],"text":"而且，當前，大量算法生產過程還是非標準化的，因爲非標準化，所以算法生產過程充滿不確定性。"},{"type":"text","text":" 曠視研究院周而進研究員表示，算法生產的過程就像“鍊金術”，需要付出大量時間試錯。這個非標準化的生產過程，一方面需要靠全能型的人才從頭到尾解決這個問題，這對於人才的儲備、素質的要求非常高。另一方面如果算法流程仍然是這樣一個高度耦合化的過程，這也說明整個算法生產在業界處於初級的階段。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/0a\/6f\/0ac6288e7c8923a7aa06c26b8a96796f.png","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","text":"曠視研究院周而進研究員"}]},{"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":"曠視認爲，落地實用是檢驗算法的最高標準之一，而生產落地實用算法的過程只有先標準化才能自動化，才能更進一步實現規模化的普惠易用。"}]},{"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":"從2014年至今，持續打磨升級的AI生產力平臺Brain++，就是曠視推動算法生產邁向自動化時代的AI基礎設施。"}]},{"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":"周而進表示，曠視十年的的算法探索和應用落地，就是希望把整個算法生產的過程標準化。只有標準化了，才把之前積累下來的經驗、模型、算法工程化起來，成爲組件和工具。工程化之後，纔有可能做到整個生產流程的自動化，從而大大擴大算法生產的能力。“我們希望AI生產力平臺能夠把整個算法生產過程，通過標準化和工程化的方式最終走向自動化的平臺”。"}]},{"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":"在現場，曠視介紹了Brain++的最新功能，包括自動化的數據管理和質檢、自動推薦合適的訓練算法和模型、自動檢查算法準確率並給出優化建議等。此外，Brain++具有三大特點，一是生產真正有效算法；二是算法能力複用；三是公式化管理各類算法模型和策略，便捷用戶調用和使用。"}]},{"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":"從基礎科研，到行業落地，再到基礎設施，這是人工智能技術價值躍遷的三個關鍵。預研一代、孵化一代、交付一代，是曠視技術創新的研發模式。"},{"type":"text","marks":[{"type":"strong"}],"text":"在技術開放日現場，曠視將這一套創新研發模式首次對外完整的呈現出來。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/dc\/1b\/dc7c237129117acyy5d81dbcd387eb1b.png","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","text":"曠視技術開放日的Demo展示現場"}]}]}