1、淺談人工智能的下個(gè)十年1淺談人工智能的下個(gè)十年1人工智能的第三次浪潮2人工智能的第三次浪潮2人工智能歷史3人工智能歷史3人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byClaude ShannonShannon, Claude E. XXII. Programming a computer for playing chess. Philosophical magazine 41.314 (1950): 256-275.1950計(jì)算機(jī)象棋博弈1954圖靈測(cè)試Alan TuringTuring, Alan M. Solvable and unsolvable problems. Science News-

2、ens. fr 39 (1954).4人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byClaude Sh人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by1956達(dá)特茅斯會(huì)議McCarthy, J., et al. Dartmouth Conference. Dartmouth Summer Research Conference on Artificial Intelligence. 19561959一般問(wèn)題解決器JohnMarvin McCarthy MinskyNathan RochesterClaudeShannonHerbertJ.C. ShawAllen NewellSimonNewell, A.;

3、 Shaw, J.C.; Simon, H.A. (1959). Report on a general problem-solving program. Proceedings of the International Conference on Information Processing. pp. 256264.5人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by1956達(dá)特茅斯會(huì)人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byDaniel BobrowBobrow, Daniel G.Natural language input for a computer problem solving system

4、. (1964)1964 理解自然語(yǔ)言輸入1966 ELIZA人機(jī)對(duì)話Joseph Weizenbaum Weizenbaum, Joseph. ELIZAa computerprogram for the study of natural language communication between man and machine. Communications of the ACM 9.1 (1966): 36-45.6人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byDaniel Bo人工智能領(lǐng)域發(fā)展趨勢(shì)owered by1968 世界首個(gè)專家系 統(tǒng) DENDRALEdward Feigenbau

5、mBuchanan, Bruce, Georgia Sutherland, and Edward A. Feigenbaum. Heuristic DENDRAL: a program for generating explanatory hypotheses in organic chemistry. Defense Technical Information Center, 1968.7人工智能領(lǐng)域發(fā)展趨勢(shì)owered by1968 世界首個(gè)專家人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byRandall DavisDrew McDermott, Jon DoyleMcDermott D, Do

6、yle J. Non-monotonic logic IJ. Artificial intelligence, 1980, 13(1): 41-72.1976 大規(guī)模知識(shí)庫(kù)構(gòu)建與維護(hù)Applications of meta level knowledge to the construction, maintenance and use of large knowledge basesM. Stanford University, Computer Science Department, AI Laboratory, 1976.1980 非單調(diào)邏輯8人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byRan

7、dall D人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by1980 計(jì)算機(jī)戰(zhàn)勝雙陸棋世界冠軍Hans BerlinerBerliner H J. Backgammon computer program beats world championJ. Artificial Intelligence, 1980, 14(2): 205-220.9人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by1980 計(jì)算機(jī)戰(zhàn)人工智能領(lǐng)域發(fā)展趨勢(shì)owered by1987 基于行為的機(jī)器P人學(xué)Rodney BrooksBrooks R. A robust layered control system for a mobile r

8、obotJ.Robotics and Automation, IEEEJournal of, 1986, 2(1): 14-2310人工智能領(lǐng)域發(fā)展趨勢(shì)owered by1987 基于行為的機(jī)人工智能領(lǐng)域發(fā)展趨勢(shì)owered by1987 自我學(xué)習(xí)雙陸棋P程序Gerry TesauroTesauro G. TD-Gammon, a self-teaching backgammon program, achieves master-level playJ. Neural computation, 1994, 6(2): 215-219.11人工智能領(lǐng)域發(fā)展趨勢(shì)owered by1987 自我學(xué)習(xí)

9、雙陸人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byTim Berners-LeeBerners-Lee, Tim. Semantic web road map. (1998).McGuinness, Deborah L., and Frank Van Harmelen. OWL web ontology language overview. W3C recommendation 10.2004-03 (2004): 10.1998 語(yǔ)義互聯(lián)網(wǎng)路線圖2004 OWL語(yǔ)言12人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byTim Berner人工智能領(lǐng)域發(fā)展趨勢(shì)Powered byGeoffrey Hinton

10、Le, Quoc V., et al. Building high- level features using large scale unsupervised learning. arXiv preprint arXiv:1112.6209 (2011).2006 深度學(xué)習(xí)Hinton, Geoffrey E., Simon Osindero, and Yee-Whye Teh. A fast learning algorithm fordeep belief nets. Neural computation 18.7(2006): 1527-1554.2011 高層抽象特征構(gòu)建13人工智能

11、領(lǐng)域發(fā)展趨勢(shì)Powered byGeoffrey H人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by2009 谷歌自動(dòng)駕駛汽車(chē)Sebastian ThrunMarkoff, John. Google cars drive themselves, in traffic. The New York Times 10 (2010): A1.14人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by2009 谷歌自動(dòng)駕人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by2011 沃森獲得Jeopardy冠軍IBMs WatsonMarkoff, John. Computer program to take on Jeopardy!.

12、The New York Times (2009).2011 自然語(yǔ)言問(wèn)答Apples Siri Sadun, Erica, and Steve Sande. Talking to Siri: Learning the Language of Apples Intelligent Assistant. Que Publishing, 2013.15人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by2011 沃森獲得J人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by16人工智能領(lǐng)域發(fā)展趨勢(shì)Powered by16人工智能近10 年17人工智能近10 年17AI趨勢(shì)：從感知到認(rèn)知From perceptron to

13、 cognitionComputingPerceptionCognitionStorage & ComputingRecognize text, images, objects, voicesOrganize and generate knowledge, reasoning18AI趨勢(shì)：從感知到認(rèn)知From perceptron to Artificial IntelligenceAlphaGoImage recognitionSelf-driving19Artificial IntelligenceAlphaGo20DDPG(2015)A3C(2016)Perceptron(1958)Fr

14、ank Rosenblatt Cornell University psychologistBPNN/MLP(1986)Hopfield Network(1982) recurrent & feedbackGeoffery Hinton University of Toronto deep learningNeocognitron(1980) convolution & poolingLeNet/CNN(1998)Yann LecunNew York University deep learningAlexNet(2012)Relu, dropout & biggerVGG(2014)Goog

15、LeNet(2015)ResNet(2016)Kaiming He MSRA = FAIRcomputer visionDenseNet(2017)RBM(1986/2006)stackDeep Belief Nets(2006)VAE(2013)AutoEncoder(1989/2006) Denosing Autoencoder(2008)Variational InferenceMax WellingUniversity of Amsterdam statistical learningGAN(2014)DCGAN(2014) WGAN(2017) PGGAN(2017)Ian Good

16、fellow Google Braindeep adversarial learningRNN/LSTM(1997)Jrgen SchmidhuberIDSIAUniversal AISeq2Seq(2014)RNN in Speech Recognition(2013)Yoshua Bengio University of Montreal Deep learningNeural Probabilistic Language Model(2003)word2Vec(2013)SeqGAN(2017) LeakGAN(2018)Character CNN(2015) self-attentio

17、n(2017)Deep Q- learning(2013)AlphaGo(2016) Double DQN(2015)Dueling Net(2016)David Silver DeepMindReinforcement learningAlpha Zero(2017)Capsule Nets(2017)算法20DDPG(2015)A3C(2016)PerceptroBERTPre-trainFine tuneBeat all state-of-the-arts on 11 NLP tasks in 201821BERTPre-train21XLNetAutoregressive ModelB

18、eat BERT in 201922XLNetAutoregressive Model22ALBERTA Lite BERTParameter-reduction techniquesBeat XLNet and all the others23ALBERTA Lite BERT23Video-to-Video SynthesisThe best video synthesis performance24Video-to-Video SynthesisThe begraph_netBy DeepMind25graph_netBy DeepMind25MoCoUnsupervised visua

19、l representation learningMomentum contrastive learningOutperform its supervised pre-training counterparts26MoCoUnsupervised visual repres27SimCLRSimplified contrastive learning frameworkOutperform previous self-supervised and semi- supervised methods on ImageNet27SimCLRSimplified contrastive人工智能未來(lái)28

20、人工智能未來(lái)28第三代人工智能的理論體系早在2015年，張鈸老師就提出第三代人工智能體系的雛形； 2017年DARPA發(fā)起XAI項(xiàng)目，從可解釋的機(jī)器學(xué)習(xí)系統(tǒng)、人機(jī) 交互技術(shù)以及可解釋的心理學(xué)理論三個(gè)方面，全面開(kāi)展可解釋性 AI系統(tǒng)的研究2018年底，正式公開(kāi)提出第三代人工智能的理論框架體系建立可解釋、魯棒性的人工智能理論和方法發(fā)展安全、可靠、可信及可擴(kuò)展的人工智能技術(shù)推動(dòng)人工智能創(chuàng)新應(yīng)用具體實(shí)施路線圖與腦科學(xué)融合，發(fā)展腦啟發(fā)的人工智能理論數(shù)據(jù)與知識(shí)融合的人工智能理論與方法第三代人工智能的理念在國(guó)內(nèi)外 獲得廣泛影響力29第三代人工智能的理論體系早在2015年，張鈸老師就提出第三代認(rèn)知圖譜 (Co

21、gnitive Graph)知識(shí)圖譜, 認(rèn)知推理, 邏輯表達(dá)30認(rèn)知圖譜 (Cognitive Graph)30知識(shí)圖譜“Knowledge graph”由Google于2012年提出知識(shí)工程，專家系統(tǒng)CYC: 世界上歷史最長(zhǎng)的AI項(xiàng)目 (1985)Tim Berners Lee Father of WWW Turing AwardEdward Feigenbaum Father of KB Turing Award31知識(shí)圖譜“Knowledge graph”由Google于2認(rèn)知圖譜：算法與認(rèn)知的結(jié)合Quality CafThe Quality Cafe is a now-defunct

22、diner in Los Angeles, California. The restaurant has appeared as a location featured in a number of Hollywood films, including Old School, Gone in 60 Seconds, .Los AngelesLos Angeles is the most populous city in California, the second most populous city in the United States, after New York City, and

23、 the third most populous city in North America.Old SchoolOld School is a 2003 American comedy film released by DreamWorks Pictures andThe Montecito Picture Company and directed by Todd Phillips.Todd PhillipsTodd Phillips is an American director, producer, screenwriter, and actor. He is best known fo

24、r writing and directing films, including Road Trip (2000), Old School (2003), Starsky & Hutch (2004), and The Hangover Trilogy.Alessandro Moschitti is a professor of the CS Department of the University of Trento, Italy. He is currently a Principal Research Scientist of the Qatar Computing Research I

25、nstitute (QCRI)Alessandro MoschittiTsinghua UniversityTsinghua University is a major research university in Beijing and dedicated to academic excellence and global development.Tsinghua is perenniallyranked as one of the topacademic institutions inChina, Asia, and worldwide.32認(rèn)知圖譜：算法與認(rèn)知的結(jié)合Quality Caf

26、Los A算法: BIDAF, BERT, XLNet目標(biāo)：理解整個(gè)文檔，而不僅僅是局部片段但仍然缺乏在知識(shí)層面上的推理能力BiDAFBERTXLNet33算法: BIDAF, BERT, XLNet目標(biāo)：理解整個(gè)文挑戰(zhàn) : 可解釋性34大部分閱讀理解方法都只能看做黑盒:輸入: 問(wèn)題和文檔輸出: 答案文本塊 (在文檔中的起止位置)如何讓用戶可以驗(yàn)證答案的對(duì)錯(cuò):推理路徑或者子圖每個(gè)推理節(jié)點(diǎn)上的支撐事實(shí)用于對(duì)比的其他可能答案和推理路徑挑戰(zhàn) : 可解釋性34大部分閱讀理解方法都只能看做黑盒:認(rèn)知圖譜: 知識(shí)表示, 推理和決策35認(rèn)知圖譜: 知識(shí)表示, 推理和決策35和認(rèn)知科學(xué)的結(jié)合Dual Proc

27、ess Theory (Cognitive Science)System 1 IntuitiveSystem 2 Analytic36和認(rèn)知科學(xué)的結(jié)合Dual Process Theory (C3737Reasoning w/ Cognitive GraphSystem 1:Knowledge expansion by association in text when readingSystem 2:Decision making w/ all the informationSystem 1IntuitiveSystem 2Analytic38Reasoning w/ Cognitive Gr

28、aphSyCogQA: Cognitive Graph for QAAn iterative framework corresponding to dual process theorySystem 1extract entities to build the cognitive graphgenerate semantic vectors for each nodeSystem 2Do reasoning based on semantic vectors and graphFeed clues to System 1 to extract next-hop entitiesQuestionQuality cafTodd PhillipsGone in 60 secondsOld schoolLos AngelesDominic SenaSystem 2Cognitive Graphlocation featured in a number ofSyste