IBM WatsonQuestionanswering:input,query,output.2011, IBM Watson win in the Jeopardy!.It has a 4TB knowledge database, including encyclopedias,dictionaries, thesauri, newswire articles, and literaryworksItused90IBMPower750 servers,720DENRcores, 16TBRAM.N$5.800$21,440$18,200Watsonhealthcare,financialservices,HBAVATEON97%education...contempt108%contemnADespised loon
▪ Question answering: input, query, output. ▪ 2011, IBM Watson win in the Jeopardy! . It has a 4TB knowledge database, including encyclopedias, dictionaries, thesauri, newswire articles, and literaryworks. It used 90 IBM Power 750 servers, 720 cores, 16TB RAM. ▪ Watson healthcare, financial services, education. IBM Watson
强化学习AlphaGOAlphaGO,madereinforcementlearningbecomepopularovernightRLmimicshowanimalsadapttotheenvironment.internal statereward..00:00:4800:01:00environmentAlphaGoactionleaming ratenversetemperaturediscount rateobservation
▪ AlphaGO, made reinforcement learning become popularovernight. ▪ RL mimics how animals adaptto the environment. 强化学习 AlphaGO
13Alphafold2的整体架构：组合创新HIghconfidenceoleerMSALowconoOFMSAStructureEvoformerOIIttmodule(48blocks)(ablocks)inputsequence+r43Dstructure(xe)eretemsplatesRecycling (3 times)End-to-end架构1D与2D信息之间使用了Attention3DEquivariant（等变）StructureModule架构上采用了Recycling·Attention模块的细节：EvoformerJumper,J,etal.Nature, 2021Structuremodule的多个创新：IPA,residuegas
Alphafold2 的整体架构：组合创新 Jumper, J, et al. Nature, 2021 13 • End-to-end架构 • 1D与2D信息之间使用了Attention • 3D Equivariant（等变） Structure Module • 架构上采用了 Recycling • Attention 模块的细节：Evoformer • Structure module的多个创新：IPA, residue gas
MythFacSuporintelliaoncemayhappeninby2100sinevitabledacaoscenturiesornever:AlexpertsMyth7disagree&weSuporintelligencesimpiydon'tknowby210oisimpossibleMythFactOnly LudditesManytopAiworryaboutA!researchersare concernedMythicalworryActuiaiworryAl turning evilAiturningcompetent,AIwith goalsusMythical warry:misalignedwithoursAlturning conscious9Myihs3Faet对AI的担心oRobotsaretheMisaligned intelligencemainconcernis themainconcern:Ditneedsnobody.onlyaninternetconnectionMythFactAi can'tcontrolintelligenceenablos control:humanscontroltigersWEby being smarterMythaFactMachinesAheat-seekingcan'thave goalsmissilehasa goalMythieaiWoryActualworryPLANPANIC!It's at leastSuperintelligenceAHEAD!isjustyears awaydecades away.butitmaytakethatlong tomake itsafehttps://futureoflife.org/background/benefits-risks-of-artificial-intelligence/?cnreloaded=1
https://futureoflife.org/background/benefits-risks-of-artificial-intelligence/?cnreloaded=1 对AI的担心
人工智能流派JomAI开始阶段的两大流派：Neuroscienceinspired模拟神经系统目前是概率模型加上神经网络占了上风neuralnetslearningfromexamplesforartificialperceptionClassicalsymbolicAl符号逻辑学派：Primacyoflogicalreasoningcapabilities两尾神经网络 No learning (humans coding rules)单层神经网络poor handling ofuncertaintyS2012Got eventually fixed (Bayes Nets...)2020CNN神经元OWintert期二次AeBO1986BPAI现有的几大类学派第一次19821995HopfieldSVM·符号学派：知识库，逻辑推断19581969Perceptron"AWinter"第一次·联结学派：Perceptron，神经网络1943E1949MPHebb诞生·进化学派：GeneticAlgorithm194019501960197019801990200020102020·贝叶斯学派：先验概率：类推学派：KNN
人工智能流派 AI 开始阶段的两大流派: Neuroscience inspired 模拟神经系统: ▪ ➪ neural nets learning from examples for artificial perception Classical symbolic AI 符号逻辑学派: ▪ Primacy of logical reasoning capabilities ▪ ➪ No learning (humans coding rules) ▪ ➪ poor handling of uncertainty ▪ Got eventually fixed (Bayes Nets.) AI 现有的几大类学派 • 符号学派：知识库，逻辑推断 • 联结学派：Perceptron，神经网络 • 进化学派：Genetic Algorithm • 贝叶斯学派: 先验概率 • 类推学派：KNN 目前是概率模型加上神经网络占了上风