2023中国人工智能系列白皮书（深度学习）.pdf

!#$%&()*+!#$%&()*+,-./,-./!#$%&(!#$%&()*+,-!1#$%&()*+.1 1.1$%&.1 1.1.1$%&.1 1.1.2$%&.2 1.2$%&k.3 1.2.1 .4 1.2.2 .4 1.2.3 .6 1.2.4$.7 1.2.5 .8 1.2.6 .8 1.3$%&e.9 1.4$%&.12 1.4.1 .12 1.4.2 .14 1.4.3 .15 1.4.4 .15 1.4.5 .16 1.5$%&.17 1.5.1 .17 1.5.2 B.18 1.5.3.19 1.5.4 .19 1.5.5 .20 1.5.6 .21 1.5.7 AlphaGO.22 1.5.8 AlphaFold.23 1.5.9 ChatGPT.23 1.6$%&!.25 1.6.1#$%&.25 1.6.2%&()*.25 1.6.3%3+*.26 1.6.4,$%&-.26 1.6.5/012.26 1.6.6$34567.27 1.6.7 89:;:;.27 1.6.8$%&?.28 1.6.9$%&A4.28!2 ,-./01.29 2.1 .29 2.1.1 B.30 2.1.2 CDE:.32 2.1.3 FB.33 2.1.4.GHB.33 2.2 I*.33 2.2.1 LeNet*.33 2.2.2 AlexNet*.34 2.2.3 VGG*.36 2.2.4 GoogLeNet*.37 2.2.5 ResNet*.39 2.2.6 ShuffleNet*.40 2.2.7 Vision Transformer*.41 2.2.8 JK.42 2.3$LMN5*.46 2.3.1$LMN.46 2.3.2$LMNO*.47 2.4 PQRS5*.49 2.4.1 PQRSTUVW.49 2.4.2 PQR*,XYZ.51 2.5 U*.57 2.5.1 U*.57 2.5.2 _V U*.59 2.6 X.63!3#$%&201.116 3.1 .116 3.2$%&9*a.119 3.3 bcde.121 3.4 fghijkTU.124 3.5$%&%&9*.126 3.6$%&9*lmnhiopq.133 3.7 derstuvj.135 3.8!WvoWw.138!4 ,-./)3456789:;?ABCD.173 5.1 .173 5.2&x.173 5.2.1.tYZ.174 5.2.2.tde.180 5.3&.183 5.3.1,.YZ.184 5.3.2,.YZ.186 5.3.3 L*Z.190 5.3.4.de.193 5.4:;|.195 5.4.1 .195 5.4.2%&:;.197 5.4.3 o$2&.198 5.5 X.199!6 ,-./BEF901GH.66 6.1“”.66 6.2I*.66 6.2.1 .67 6.2.2 .69 6.2.3 .71 6.3I*BM.86 6.3.1,x/.tYZBM.90 6.3.2,MBM.95 6.3.3,ZBM.97 6.3.4 BMLv5.99 6.3.5 5Ww.102 6.4 X.104!7 ,-./JK.LMABCD.213 7.1 leJBxfg.215 7.2 lfg.216 7.3 pUeJS.219 7.4 lpTUfg.220 7.5 lp-fg.225 7.5.1 p.225 7.5.2 N.227 7.5.3-.228 7.5.4 M.231 7.5.5-.234 7.6 8xB&*.239 7.7 gpD.243 7.8 gpx.247 7.9,-.249 7.9.1 .249 7.9.2%.249 7.9.3 N.251 7.9.4-.253 !8 NO%&PQRS.273 8.1 .273 8.2,$.273 8.2.1 5S.273 8.2.2$y.277 8.2.3 .279 8.2.4)$.282 8.3,x.285 8.3.1 x.285 8.3.2 x.291 8.4&$W|.299 8.4.1=X.299 8.4.2 .306 8.4.3%M2.309 8.4.4#t$%&.310!9 ,-./)%TUVWXCD.322 9.1 7x.322 9.2 7$.326 9.3 .330!10 ,-./)Y$ZWXCD.340 10.1 x.340 10.1.1 xL*.340 10.1.2 x3+.341 10.2 .343 10.3 .352!11 ,-./2RSABCD.362 11.1 eJ9:;.362 11.1.1 eJ9:;a.363 11.1.2 eJ9:;rs.365 11.1.3 eJ9:;de.366 11.1.4 fgk.368 11.2 9:;.374 11.2.1 9:;a.374 11.2.2 9:;rs.379 11.2.3 9:;de.382 11.2.4 fgk.384 11.3 9:;.398 11.3.1 9:;a.398 11.3.2 9:;rs.401 11.3.3 9:;de.402 11.3.4 fgk.406 11.4 X.412!12 ,-./_abcABCD.417 12.1 a.417 12.2 .419 12.2.1 a.419 12.2.2 .420 12.2.3 Na.425 12.2.4 N.426 12.3 .428 12.3.1#TUVW.428 12.3.2,.431 12.3.3 .433 12.4 y.435 12.4.1 .435 12.4.2.442 12.4.3 o=.445 12.5 deWw.446 !1!1#$%&()*+1.1!#$%&()*1.1.1 人工智能的概念人工智能的概念$%&Artificial IntelligenceAI!TU?#$%$L&9&nde l$t(u)tu*+&,$%&1-./012v4 1$%&%$L&9&,$%&3R#$x$L=45t$V9u)6789hiTUfg,2$%&R,$%&:$8:;?N,$%&-ABCD%*9,3$%�tu*+,$%&21-#$tuEF*+.Gx89,9de#089:;Hg$8IJKL*+tuv,4$%&1N5%=N,M$%&(NfO$PQRV9u)9IS-TU$LV,$%&BC%2Whi.GXY(N&ZS9,WhiHTUvde_B$%&%=N,$%&WaBCTU bS-9cfg,dR:;ye$%&lE2hifHfg.G&8g(h(NiZ9,$%&WjkR$Ll0vvmn!2()?l0*+S-$o_ tu,1.1.2 人工智能的起源人工智能的起源$%�pqCrnst$%&uvwx yz|$yvK0 1956 M0$%&,科幻小说的人工智能（弗兰肯斯坦）$%&rlJHJ kW 1818 4!uvwx-,tynY%uvwxTU“”2$”$,;JQo$%&t$o$%“”m0n5,uvwx?“”m ZS3N=&$%&?*+2r,阿西莫夫的机器人三定律 l 20 4#4Yyz|&nst“”$,-Jl 1942 4!$L93$LC,42$8$Lmm5,42$?=5,$o$%&?,yz|4p$5$Lb(0?$%&-.2,!3 阿兰图灵的计算机博弈论和图灵测试 yvn$%&,$,l 1950 J!5&-p$J&;2c5l22$V9=2$2#0&,B&$%&hi-v mnYoH&5#NbP5TU&$%&W,人工智能概念的提出 l$%&hiTUD,1956eM-T0-0&$%&n,0nY.GJ,?Mz9,o_?#g%$L&9&;$%&!WV9p,lM0o$%&4$%&!l$%$L&9&n,0&$%&TUW,;CnY$%&pqW,d($%&CWB&nhi2-TUvo,1.2!#$%+,-.$%&Wk 20 4 60-70,l$opq#g%$L&#N,Xj”O$%&W2-v4$,!4 1.2.1 感知机感知机$%r*r#uvM 1957 WTUl1960!1970,l#&FH$R&(N$%&W%&,4&m&de-(),*F*B+,X*,$%,m-&,H./:01%/,1-2L/:;,l22?L,l19601970r TU34%PQR&J$%&-56,Id($o78C9l:;N N#*+=N#:;V9L,Q?N*+HLR,J3TUV2$oLg%A,1969#J,zBCMlJDtuE*+m-&Minsky*+,2=VFeG$o<uH*+H;,Iltu*+o9l:;NJ&$%&W,d($o78J$RMKL#K%pq,1.2.2 贝叶斯网络贝叶斯网络 4&F*$%&W%&,m-&$*O!5 r#:YMFb(7NOb(*,HoPDirected Acyclic Graph,DAG NOb(;#gyaNOb(,5N gv(:;HKQtN(R,1-l#0RC:;K 8IJ!(,1-M.G4 1 N,$HLRN*+,#gNOb(#*$S0:;!V9(,2HoP,HoPQ?NOb(HvoN9lTU,=#*HiV=b(;WXt,3:;B,#0B:;2V9(,F:;BR#*Ke.9,4#tN,YZ*HKQ#tN,$RNOb($ot*u)2,5f=X:;,l=X:;J,v#0(_:;*+al:;H;bJ$HLRV9(,oaMlx9hicfg,d$%&deWjdUe-4g&:;u)AT9hiHL(v,!6 1.2.3 模式识别模式识别 ddeW$%&hi CK%bPTU,l 20 4 60 70 B&n$%&hi-TUvo,D%,Xvml2W k-fg&h0&$%&W,K-fgK-Nearest NeighborKNNMnuijM 1968 FkNv,KNN L2oBlvm*,Ov U4=2=XlMm k 2nMmgf=X9%:op2LO=Xmo2L,KNN qrgsftLu,OvlLu)o=N;gf212=XLuv=XoL,KNN vI8womNO;lLu)=5xg=XHb,KNN v1-yNzH;gf=X IyLivoLaLi6|%v=XtKNN vvK&2,&h0Dynamic Time Warping,DTWFg)-v,rX 1975 gtu8$*+,lDTW 1-g8$&hv2)U6#)nNC9,dTUDTW CW_V,TU$Qjfgg9&$9hiB=,NDTW fgz9B&)hi-%,al!7 eTUCcfgbfg%ZS8q9Hfg,d$%&deWDTW jdUe-4g;&Kchituvj,1.2.4 人机对话人机对话$Human-Computer DialogueB&$%&hi-pq,$F$tg5$V9$deJ$%&5$L6bc,r$1-l,h*(N,|“H”1966 ELIZA,ELIZA *Tfg/,ELIZA;HtRF%$LRl#9bP,ELIZA$0FLnv5gV9 Tfo*+;p*#g%Fl5$L,#g/Ln C ELIZA 0TLn&C*+8IgVF?,ELIZAbHr$(N,k1968M SHRDLU$l%v45g#0$V9fA4,1971J&PARRY e$B2H-H?&,I(NfH:;No&KVTU,l#de r$(NtWJltv$R,Io5#&(Nb*o&$%&WO,ddeVF$TUml&(Nt#0JKe&67,!8 1.2.5 知识表示知识表示 l 20 4 60-70 nYoTU?HLR90#$(8ITU,2TUBn$%&hi;2hi_9l2rs,lnB$L?9$,Il$%&hi1-j9#2$;JLn$L&9&,IhiJH2Stuvj,yvn9$l1969Prolog.2F,#gh#$(tu*+,Prolog l$%&hi(.2W;&TU-U,l Prolog 0-h,ko#0LH FL90=H;W=9h,#0hV9(T*+tu*+,gYZJ,9lhiV9TU,l2WbC;S2&$%&!W,;&lTUhi9VW,ddeW$%&lvVF#$%&fgKec,1.2.6 计算机视觉计算机视觉 l 20 4 60-70 Computer VisionmB&$%&TU-vo,I#Nx;STUVW,#r!9 Blocks Worldn9l%v45V9(N,I9lD%*+&TU,2-VW9h1970 ,j(N&2B-4w2.5D 3D u,lw(Nx,XM/0*x,l 2.5D(NxV9KB-B$,l 3D u(N02%2 2.5D$0,&W,#ltxv9VF,oa 20 4 60-70$%&hi-bc,$9B&OTU&$%&W,I#deC;S2&e&(N 9deBl-O,1.3!#$%/0 20 4 60$%&J&$%&TUfg9dIl 20 4 80,1-&l$%&WD%!J$%&TUOhiVWCw,3D%$%&TUOhi,NNv9l3(NJ,2R,$%&TUS-9I#N;S*J,3:;,$%&S-9:;V9()l 80 9:;BX3:;hH;,4HLfg,$%&delTUhiB=lHfgJ;tu45*+#ATC,2m,l 60 70$%&lD%$!$%&“v,I#80$%&(NC&,3,$%&deVW78xTU,)S$%&“CA o&$%&W#S-KS-K%,4XY(N:;,$%&r1-TUvoXY(NI(NlHfgmC:;,IolpMhiB(RH;tuKec*+,5rs,80$%&de,D%$lfghi;!Qg$%&de 3,oa20 4 80$%&hi*+1-!11 de:;mXY(N:;0rs9_”B=,$%&rsOhi;!S.,TUladU3R$%&;-VW,20 4 80$%&W934 1980 4WABOT-2 lr9B,WABOT J#2$5$oV96;lWio8,1982 4X#HX%!8.5#,g$Ol#06%&t(;JLn$L(,1984 4)*|+?8!W0-o,z-2.$/$2&EdgarHR2$01b(W,1984 4l$%&30o$%&Y2nYJ,$%&$%&TU30 x2k,olB&J,1986 4 4-56lHM78”;9:x;$?$LQUo0 55 r9,1988 4nYAYB7CCD!&(N(-,CbICEF F*HoPNOb(,1988 4 2Q2e$Jabberwacky Cleverbot 1990 GHM Jabberwacky 00H3IJKv%$Le,2e$5$oV96$%&kW,!12 Lt20 4 80$%&$%&W)o2-2W$%&TUWrs:;N,mB&)SM oCz-$%&N4H#O9P IOhiVWS-TU,1.4!#$%123 20 4 80 R:;S$%&deQ,dW_V$%&RV,la$%&l9hiHVF,1.4.1 机器学习机器学习$%&Sde#(N$:;_TN,lD%IAToSupport Vector MachineSVM dUConditional Random FieldCRF Boosting 9;cfg,支持向量机 SVM 29V#:;W98IJLTK3+,lWSVM 1-fgL*+,TUo SVM V9W#0g%LTK*+,SVM WNVWMSE:#SVM$N*+,SVM HQMqrRN$XJXM#HLRJXV9LPacfgJXLm93+,aSVM$()=XCL 8I$xV9La SVM mgxL$Y9!13 3+,条件随机场 CRF Fg)P*$Z)NOb(V9#0oJ$;():;bM/)Z,ru*9$,CRF 1-.()*,CRF#0RjoJZ1(*8IDRM ImhigmJQ*,lHfgCRF 1-g)P*+mNP_9N*$KQRcgoJ$PN,Boosting 算法 Boosting FBv2%229,M 1990 Boosting,OS-c0$3aQ,la,oM|uGH 1996 _V AdaBoost Adaptive Boosting,AdaBoost Boosting :02=X17FyLJ,aS 10 9abHL;W_(fgc%L*3+,u8 1999 GBMGradient Boosting MachinesFdeE:fd()LVF=,AdaBoostGBM$KgeE:8IKQR%2:;,GBM 5 AdaBoost=h&J Boosting,ocfgLVW;&U9de-Uv,a!14 SVMCRF Boosting Bfgm#$o*#tN-NHK&UTUO,1.4.2 自然语言处理自然语言处理 dR09*:;iQ hij-W,TU$QLJXv;k9:;R()%deQN*,*#gJX;W9h:;NlN,N&m*Bag of Words,BAG,1990 Cn9W2&WordNetQ29hro:;i,WordNet j*Bya*b(&p93+-,lJX3+m*hiFgJXv,j2JXq42*B2rs*t aN2*lJXJ9l5,tm*jJX&2oo2f2*uO*lJXJ-:9l5,m*lJXL JXB93+#04&/MtJX:;,m*&$qvqwW,jxyJXJz&m*#0on;JbN8Ivq=NL,Ltm*W$%&hi,&Ude-ATO,*IHLv&JX$q,X%U,WmLVJXMvTU(deW“!15,1.4.3 神经网络神经网络|OMdeV2,%VSvqt 7F2;A4 qr!16 6E%t,V 1990 afghiH9 WcgV9hRm,vlfgTUWm&yDR,-W#)&4 1 55;78 5(9&v72 21 4&deWH-Q,2;9TUbDtC&(B;B&-P,35$%m*TUhi2,fV9Ju$%m-TUul#e4g,4V2,V=m%5VS&vHmHEMj2B&#,1.4.5 深蓝超级计算机深蓝超级计算机 1997$%&hi24#IBM de$QTSlhZS0 3.5:2.52 1f3 s#45“#Hu”:-|,mj)o2Bl#Hu”445_(N l#$o5$%&Wt,;H#gNQo$%&de k JgXY(NAT”u),NO,h(N#g9:i,vqF#”:;vql!17:b#2#RCF,abcg;9R%2*,F0L,F;9del#tV&lV99hAT,cU$%&deW%-1-Jl2v4 1$ol3+oR,s#Hu”45_jl&R3+oRW.45e$%&deTUfgV2,2(vu)deTU,gvqlH;:#2#R;Cu)bc,L#TUlvu)hiV9KTU&hi*+tuO,3(;9deW,U$%&deWC(4g,2.GHBfGHReLU E:Dropout 2.GHBGH softmax E:V9L,AlexNet#g GPU()jB.GHBC GPU oV9;9(),d AlexNet ILSVRC 49Y1f#gIK ZFNetGoogleNetVGG ResNet,lTU$oJ=*RyeB:()0JJudye()()078|#-yeB:0()y9,tu*+2-v,HzQNhi2-CocbP,X*,1-BKB ReLU CDE:|IB,YZ&4/:;j/:;N-/CBReLU KBj=C%2*,-jC=/CKB%2.GHBC/=,TUom_V1-Jl*,YZo,1.5.2 生成对抗网络生成对抗网络 BGenerative Adversarial NetGAN-*,GAN./2*2B*generative model 2*discriminative model,B*Bqw:;Hfk:;,*kb$&”,l()B*3+B(*3+B*B(!19(,7()#CKhJKQL=*/:;b”,GAN cfglE2hixyi):;BB9,GAN 1-*+4QB 8I#Bl2i3B=%=N,1.5.3 变分自动编码器变分自动编码器 Q-()bMLH3M,GAN4H3mH“=B*Y,21B*.(Variational AutoEncoderVAE),VAE*F./B*$mR=yacg()C2E:m-&(B VIB():;./:;,VAE*lB%F:;vt9$R.Gg:zx$Yx!xCIFAR xU“*x08&xV99,VAE*HGUm=#0#d=uKev,VAE#0EFL*:;)GUH#oB&9B%,VAE 9l21-oB/VAE*H0%J=X,1.5.4 迁移学习迁移学习 VAE*tQBL=4&vU9l:;N ao”F#gx():;V9*4,Hi8;BXaowgtHbRmJH,2A9%#gv0()*&1*!20:ux*(),v#0&L4(1,=X,=X1=Xn1,1=XKSB(2,M,ji():;i,CmfWSB(3,*,*=X2:0*:SB(4,b(,nU“b(=H():;iib(#0Cn:;o,gU“4(19:;=XP-dUyeK%:;PS-a9BX,acg M#0:;nIWtHP:;cg:;nNV9:;L=P-,(2*M-HKcR,:;=Xx,#BfgNO9H:;#gN,IH:;Lf,N&:;I%2*+I4QRcCU“,(3:;,N():;:;,IlD%b:;a0dmIm0dbIao#g:;v():;,l():;U“t()SB*S-l#gV9:0,1.5.5 知识图谱知识图谱 al5VFml%5CcbPK%9U“fg!21 hi,Xo1-ya459lEF0ob(a#0&F,8W2ql,oWI,H1(,52#08BtgQ_Vvq,k&vqwK-vq=4#0Qb(vq b1#0&vqwKKFvq=gKRCS$,#0*(NV9tg*V98IoNb(9$,Ft#0*(NKQRtgQ8IKej,1.5.6 自动驾驶自动驾驶 deVF1-$%&VF,l#0gJ=9u)ZS9,l#gV9 S-#g9Px:;()*,*#0J=9kU=99$?9,ab#0gJ=9u)ZSk=99U6vo9,lR4#0=9V9-R8IAT=9U6hu),lFfg/#=9:;:x;C GPS 9/#=9lRo-vo9$,l9&hu)4#0=9V99&!22 hu)8IJ=9ZS,lFfg/#=9:;R$/#=9l$R&hi AI fgO,$%&lBm&$oW AI l45*+fg“,AlphaGo c$24-N,5NQo$4AlphaGo o45W AI 5$L”g734#N#AI lD%hie4gI$L,LtAlphaGo WO$%&hi-9,Blz”hi%&;WCK%$%&TUpq,AlphaGo cm&$o$%<u*+(ndVF9$R,!23 1.5.8 AlphaFold AlphaFold DeepMind*(N,#0pqC 2016#DeepMind B AlphaGolz”hiB$%&2,AlphaGo ODeepMind ujLndefgtu*rs,%TUDeepMind l 2020(AlphaFold 2.0 DX2,*|(N,AlphaFold Wj%2DX_V*N,AlphaFold BnhiH012v41e*TU,AlphaFold N#nY$KR*e*TUVW,2(T,*Tb-,AlphaFold R&K%*nYKHL,3tu*m*+,l*3*i9l9m,AlphaFold J&nYK.*$,4OK%AI lfg,AlphaFold BDlhi$R&K%AI delTUfgR,LtAlphaFold Wfg*hi%9,&*nNVW&nY9%e*TUTV2,1.5.9 ChatGPT ChatGPT OpenAI$%&TU*,GPTGenerative Pre-trained TransformerZF,Transformer 9h()*,GPT()!24 GPT-1W&GPT-2 GPT-3I ChatGPT GPT-3 MfgDXXP67,GPT-3 ChatGPT,l()2#g9h71JX:;,()9*GPT-3$CFt,d OpenAI GPT-3 V9:j”Bg67ChatGPT,ChatGPT&2(N#05gV9I6,l%2fghiWJcg5.G4 1*+t$q,ChatGPT#0Tg*+$q;&gj,2.2,ChatGPT#0&2Q tu*+v,3 4“Q,ChatGPT#0“4BJX.GJM9,4,ChatGPT#04&l&t,5t,ChatGPT H3+JXL9,ChatGPT Jlhi%9,ChatGPT“KK$67v#$%&(N$Kef$LS,ChatGPT#0fK2N+,ChatGPTcfgW$%&lE2hifg#N&9hiK%“tuvj,ChatGPT(tBhiTU&deVFW-R,ChatGPT lD%vVWJ9lrsN#9,Idde_V ChatGPT jdU&$%&hiWn5#N,!25 1.6!#$%67 d$LQ$2%$%&lE2hiCbP;78fg$S”P99,$%&3RN#B&129TU|,adndeW$%&fgjJl$oD%v,:T$%&!WH0|,1.6.1 可信人工智能可信人工智能#$%&$%&del-.#tNvC#$OR$L$3tZS$%&(N8I$%&l0L3,I#$%&121-dei4 1-.i,Q#3$%&(NL_#yNN,2i,():;.#g_#g2$,3i,:;9lp#VF93&u)Jp,4#ti,NN0*91N#3$%&(NJ Y*+,_V e-.!gde:;0*#tN#07FJ#$%&,1.6.2 多模态预训练模型多模态预训练模型%&()*$%&8*&o%&2g&WV2 ly*g$%&RE&$?7Fh*9&,CLIP BEiT 9l%&()*deoVW,2%hiNU“%3+%&,*tB&$%&-Wvo,9*jB&!,J!26 xJX8N8IJ(T*+L“”R&V,1.6.3 多任务模型多任务模型$%&*Ma*#3-=,%3+*#0%2hi3+kxq*t9,%3+*lcg3+bN7N*cRLx:#JKgD$%&fg,%3+9*KQR%$L2Jhi&2,a b&U“$&9K9tR,NZ:EF/%3+*b$%&(N$,1.6.4 基于人工智能的网络安全基于人工智能的网络安全$%&delf-.%&vbc,l-.Zfg$%&aL:;()Nb&_Vg,Q*cfgM1B%&y,1-.gm4xbP-.N#B&q,l-.()v#0-.BX,%&rsD%3+S-0KQRfbc3+,h3+a$RblKK=,1.6.5 边缘计算和联合学习边缘计算和联合学习 /0!$%&W9|,j&:;R(o/0Hf:;,Q?l/0o#0JK&u)(IS-NO+,/0!27 lnK,-K N-:;.U“=9&Ys9*Ij:;l2Rv,12a:;bL*cR,lhi1P/&u)9U“H9$R&$%&Wn#N,1.6.6 人机协作与交互人机协作与交互 !$%&(NjK%R5$LJ34,$34yfJ#$8$L+;9&,F34b(jH$%&(N2N+g#KQRfgS,&jB&!$LD-01,$67de&$KQRtgQ;&g2N$,&tcfg&g&89!jVFE2hi 2 9&$LK&3+,1.6.7 从大数据转向多样化小数据从大数据转向多样化小数据$%&*l()2S-9:;I#3j):;#y,&KHLRcg#g:;%=:;B&F-v,l:;U“g2B:;x=X9degx:;f,%=:;3cgEF*9*:;.GJXx889;:;&b1N%=N,Fvfx9:;NO$cgtR,o#fg%2hikS-.f1(N9,F0:;&,!28%=vjl!$%&W-&:;S0U“KHHfgutuvj,1.6.8 人工智能伦理人工智能伦理 d$%&decfg$%&?l0*+4,lW$%&o?B,.G?$%&(Nu)?g:;-.0?$%&lfg”Bf,!$%&?TUcjB&$%&W-P$%&defgKe2f,$%&cfgjl0%&,.G5U6(29vfS-pq?fbrs 92Rcg$%&den$lQ,1.6.9 人工智能平台的操作化人工智能平台的操作化$%&A4(+*I-&$j$%&8&%#$NO$%&tuvjtuEF*+,TU=H7$%&$BR8V%IWBV%hS-82,I“$%&.0*A4de$%&tuvj#gN#WNN8Ie$%&gy,A4%#$KeLR$%&#$KRlH+%uJ+*,!29!2 ,-./01 4&$%&1delxJX9:;b3+9B=,X”L1-JK-”%FI95$LMN1(aPQRS5 U*2$#KHLR:$LMN,2.1 89:;BPooling.GHBFully Connected/BOutput,2-1 W CNN,X,jx4&/BlB./gMMSCDE:CDE:geNRg ReLUSigmoid#gx&Sigmoid 0Jd*+FB4gl*:8IR.GHB4glBFB/MV9jCM,C=XZm*+/0#gE:softmax E:gL*+NE:gTK*+,!30 !2-1 CNN#$%&2.1.1卷积层卷积层 l3+0$WJaEN1-KB,BH9MR 8/xM$VIx,/xmHHeightH?WidthW:ChanelC2,MBSJSl?of&/xlo5/x,M2Sl/xoA;SB,aBH:_MN#0:8I(B:%2mxC,!2-2()*!31 2.1,Sl/xoAx:-&FStride,2 2-2,N;S5/x2#&F FullSameValid,Full S5/xD2Same S5/x/2Valid S5/x.2,F 2-3,&2B/x/0$l%B#0E,0l/xV9:;_Padding0(B/0$E,:;_lxN 0_x#0#g Full Same FV90 x/0$,B/./8/xCM$aB/-&MFeature map,M95S9F:;_B:Hb,/x9&nm#gF&fSV9A4lV9B:;_ pSF&s/M9#2.2,Moxl/xo,i-&GReceptive field,G 2-4,S(i,j)=(I K)(i,j)I(i+m,j+n)K(m,n)!2.1 Size#$%&($=3!)*+,#-+15 3)*+,#-+15 2.2 !2-3 Full,Same,Valid+,(-.!32 !2-4/01 2.1.2 激活函数激活函数 lB0e2CDE:N,R,&XoFe1N,N&,*RH;CDE:N8I*R,Nb(K,#E:/0:9ld*+,#/f:ReLU E:0f:SNR&03MEmcgdK1u,a#ReLU E:L#0qg Leaky ReLU E:2.4,Leaky ReLU E:ME1uK9*+,f(x)=$0,x 0 x,x 0 （2.3）f(x)=(,x A4jpg/LNM4&a/,FA4H9FMax-PoolinghuFMean-Pooling,FBH#():S-FS9AF0FL*,FS9g&22,AF&2FFv=2-5,F 2Xo2M=2,lMx:(B%2,!2-5 234567845 2.1.4 全连接层 4&B oB/MV9,BFB.GCDE:jw:;,CBMmI.GHB25oBHGH jCM,C=XZmgLTK,a.GHB:9,2.2;%B.GHB,l3+WJNN5ATo!34 Support Vector MachinesSVMI,LeNet lg:z3+$0 1%SB3+,LqLeNet&.GH,Z2-6,2B2Bsigmiod CDE:huFBB,B#g 5 5 Sj/x,C%2M,2B/./6 2 I2B/16 2,B/M0FS9&2 2DF&2 FBV9=,0 x:,/M04&.GH/lM/S-V9Bm_&FS&33AF&3 9F5 LeNet 2-1,AlexNet 5 LeNet n ReLU CDE:9F#AlexNet Q;Wd*+,K&-AlexNet#g GPU Graphics Processing Unite(),$K9Mm:2LR()AlexNet NE-w,!2-7 AlexNet:;(?A !36 B 2-1 AlexNet 6 LeNet 9&!2-8 VGG CD-E#$%&2.2.3 VGG网络模型网络模型 AlexNet l 2012 ILSVRC 9END 3+HBL,H2gv,VGGNet 3*PQ91?Visual Geometry Group,VGGlP2,2 VGG 2-8fBCDE:FBB,lB#g.33 S 1111 9SHKx:%2S2 ReLU CDE:#0HKNR,Bb./_Padding0T,&#NVGGNet#0B-H!37 VGG-16 VGG-19,2-2 WB-VGGNet Z,VGGNet ld*ye:mdye,B 2-2 VGG-11,VGG-13,VGG-16,VGG-19 9&2.2.4 GoogLeNet网络网络模型模型 gt NHye?mQ?:,:X%2m099ye,B:#0*N0:3%2Ju,GoogLeNet4 Inception Zj.GHgGHvV9_V&XGHDXNm#0#g GPU V9RN,!38 GoogLeNet _ 22 B,Inception 2-9 B/#g9SM GMjMV9o2C%F4g,&vM2MS9&13 5 V9 012 9SA4DM93A4#0HloMV9|e,5 5 9a#g 11 SV9,_V Inception 2-10,Inception Dl2%29V9qrj|lT#gFU&29FBj7,GoogLeNet,Inception V9&d 2 2softmax LBgl.d;TdLBl0,!2-9 Inception#$-F !39 !2-10 GHI Inception-F 2.2.5 ResNet网络模型网络模型 d*ye oV29ldd*+3*0(),0Jd*ye*l()20J-oJu,&tu0o*+?W9$l2015ResNet*5,ResNet H ResNet-18ResNet-34ResNet-50ResNet-101 ResNet-152,ResNet-18 ResNet-34onHBIXoB,ResNet VGGNet=B ResNet l,oeYGH-&Residual Block 2-11,!40 !2-11 JKL%&MNOPKL%&MQO 2.2.6 ShuffleNet网络模型网络模型 ShuffleNet61-tu9*+,2-11 aH*,bShuffleNet*,e7ZA4c F&2 ShuffleNet*,!2-12 ShuffleNet RS !41 l*&a:H;03-,XGHv2=lf/o9 99BX,I=%2|lH2/:;2-12b/jS.b,toBBM#0j 2&12WoB2W,#0ZA4HLIRJ 2-12c 42Bg 2/Hgn 2j/&g,n j:OA4J0L,a Zm#Q?#0CV9+C+(),2.2.7 Vision Transformer 模型模型 lNatural Language ProcessingNLP3+B,Vision TransformerViT*l3+9 xL,klxLN CNN*VGGNet ResNetlImageNet o&71.62%78.63%I,Transformer*ViT-LargeViT-L#0C 85.63%,CNN bcBB1u0S,Vision Transformer 1-.EncoderBI.%;PQRS Multi-Head Self Attention MSA%B Multi-Layer PerceptronMLPB1u,XoR,!42 Vision Transformer/nHZ),&xx RHWCVision Transformer j/x&G_v xp R(Np2,C)HHight WWidthwxCChannel:P,P2v N=HWP2 v:u Vision Transformer/),Vision Transformer B/#2.54 x0=7xclass;xp1W;xp2W;.;xpNW;:+Epos （2.5）MSA MLP#02.62.7MLP2NB2 GELU BB./:W1 RDDh W*R11!b RD,MSA(xl)=softmax(QlKlTD)VWlmsa （2.6）Ql=xlWlq,Kl=xlWlk,Vl=