Multilayer neural network only slightly better than linear neural network

2 次查看（过去 30 天）

Triple G 2020-6-22

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/552757-multilayer-neural-network-only-slightly-better-than-linear-neural-network

评论： Aditya Patil 2020-9-22

I have made a linear neural network with a hidden layer and a multilayer non-linear neural network with 3 hidden layers. Both of them get as input a matrix of this kind:

7300    1.9500    2.3800    1.4400    5.0000    4.7500    2.1000    3.2000    2.2500    1.3000
4000    3.2000    3.3000    3.7500    3.5000    3.4000    3.2000    3.4000    3.2500    5.2500
0000    3.0000    2.0000    7.0000    1.0000    1.0000    3.0000    2.0000    2.0000    9.0000

and have this kind of matrix as target:

   0     0     1     0     0     0     0     1     1
   1     0     0     0     1     1     0     0     0
   0     1     0     1     0     0     1     0     0

The code for the first neural network is the following

function [trainClassificationRatio, testClassificationRatio] = feedForwardLinearNet(xTrain,tTrain, xTest,tTest)
rng ('default'); 
net = feedforwardnet(3);
net.layers{1}.transferFcn = 'purelin'; 
[net, tr, Ytrain] = train(net,xTrain,tTrain); 
trueTrainClass = vec2ind(tTrain);
assignedTrainClass = vec2ind(Ytrain);
trainError = assignedTrainClass ~= trueTrainClass;
trainNumError = sum(trainError);
trainPctError = 100 * trainNumError / size(xTrain,2);
trainClassificationRatio = 100 - trainPctError;
Ytest = net(xTest); 
trueTestClass = vec2ind(tTest);
assignedTestClass = vec2ind(Ytest);
testError = assignedTestClass ~= trueTestClass;
testNumError = sum(testError);
testPctError = 100 * testNumError / size(xTest,2);
testClassificationRatio = 100 - testPctError;
         
end

The code for the second non-linear multilayer neural network differs from the first only in this part:

net = feedforwardnet([6 6 6]); 
net.layers{3}.transferFcn = 'softmax';

I have also used patternnet either with the default settings or with setting the output transfer function to logsig with no improvement. Why the accuracy of the 2 different NN's is almost the same?

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Aditya Patil 2020-9-22

Neural network accuracy can be affected by many factors. Can you elaborate more on what the accuracies are currently, so that better suggestions could be provided? Also, any information on the data size and features would be helpful.

请先登录，再进行评论。

请先登录，再回答此问题。