上海交通大学神经网络原理与应用作业1.pdf

1、Neural Network Theory and ApplicationsHomework Assignment 1oxstarSJTUJanuary 19,2012Problem oneOne variation of the perceptron learning rule isWnew=Wold+epTbnew=bold+ewhere is called the learning rate.Prove convergence of this algorithm.Does the proofrequire a limit on the learning rate?Explain.Proo

2、f.We will combine the presentation of weight matrix and the bias into a single vector:x=?Wb?,zq=?pq1?.(1)So the net input and the perceptron learning rule can be written as:WTp+b=xTz,(2)xnew=xold+ez.(3)We only take those iterations for which the weight vector is changed in to account,so thelearning

3、rule becomes(WLOG,assume that x(0)=0)x(k)=x(k 1)+dz(k 1)(4)=dz(0)+dz(1)+.+dz(k 1)(5)where dz zQ,.,z1,z1,.,zQ.Assume that the correct weight vector is x,wecan say,xTdz 0(6)From Equation 5 and Equation 6 we can show thatxTx(k)=xTdz(0)+xTdz(1)+.+xTdz(k 1)(7)k(8)From the Cauchy-Schwartz inequality we ha

4、ve|x|2|x(k)|2(xTx(k)2(k)2(9)1The weights will be updated if the previous weights were incorrect,so we have xTdz 0and|x(k)|2=xT(k)x(k)(10)=x(k 1)+dz(k 1)Tx(k 1)+dz(k 1)(11)=|x(k 1)|2+2xT(k 1)dz(k 1)+2|dz(k 1)|2(12)|x(k 1)|2+2|dz(k 1)|2(13)2|dz(0)|2+.+2|dz(k 1)|2(14)2kmax(|dz|2)(15)From Equation 9 and

5、 Equation 15 we have(k)2|x|2|x(k)|2|x|22kmax(|dz|2)(16)k|x|2max(|dz|2)2(17)Now we have found that the proof does not require a limit on the learning rate for the reasonthat k has no relations with the learning rate.Intuitively,this problem equals to all inputsbeing multiplied by the learning rate(ez

6、=e(z).The cost will not change a lot to findthe correct boundary to proportional data.Problem twoWe have a classification problem with three classes of input vectors.The three classes areclass1:?p1=?11?,p2=?02?,p3=?31?class2:?p4=?21?,p5=?20?,p6=?12?class3:?p7=?12?,p8=?21?,p9=?11?Implement the percep

7、tron network based on the learning rule of problem one to solvethis problem.Run your problem at different learning rate(=1,0:8,0:5,0:3,0:1),compareand discuss the results.Ans.In my experiment,the learning rates are set form 0.1 to 1 with step=0.05.Here Ishow the times of iteration at different learn

8、ing rates(Figure 1).Note that this perceptronalgorithm can only handle two-class problems,so I use two phases of classification to solvethe three-class problem.I can hardly find the relationships between the learning rates and times of iteration.Justas what have been proved above,the cost will not b

9、e influenced a lot to find the correctboundary at different learning rates.I also proved the results for classification at learningrate=1(Figure 2).Problem threeFor the problem XOR as follows,class1:?p1=?10?,p2=?01?class2:?p3=?00?,p4=?11?20.10.20.30.40.50.60.70.80.91051015202530Learning ratesTimes o

10、f learning Classification 1Classification 2Figure 1:Times of Iteration atDifferent Learning Rates21.510.500.511.522.53302520151050510 Class 1Class 2Class 3Figure 2:Result for Classification(P2)0.500.511.53.532.521.510.500.51 Class 1Class 2Figure 3:Result for Classification(XOR Problem)43210123454321

11、01234 Class 1Class 2Figure 4:Result for Classification(Two Spiral Problem)and the two spiral problem delivered as the material,could the perceptron algorithm cor-rectly classify these two problems?If not,explain why.Ans.Just as the results(Figure 3 and Figure 4)show,the perceptron algorithm cannot correctly classify these two problems.As we know,the single-layer perceptrons can onlyclassify linearly separable vectors while the vectors in these two problems are just linearlyinseparable.3