BP Neural Network

Hi, I want to implement some custom BP Neural Networks with MATLAB! I have an aging database that contains face features (68 pairs of face points), gender and age. so I have a [1002x138] matrix as input datas for NN. 1002 face features and 138 value for each face...

face features must classify to 4 groups of ages [1-12,13-25,26-45,46-63] by NN.

Back-propagation network features : 136(68x2) Input values, 2 hidden layers, 4 output values (boolean type).

To achieve my target, I need to construct five NNs. The first NN is the main neural network. The main NN is concern with the main age classification with four outputs as I mentioned before. The other four NNs are concern with secondary age classification. Each neural network (the five networks) has 68 pairs of inputs representing the face features in addition to the gender of the person. It is also to be noted that we need two hidden layers, one hidden layer is to correlate each pair in one meaningful unit and the second is consider to be the real hidden layer after organizing the input data in the first hidden layer. The gender is connected directly to the second hidden layer. The number of outputs for each NN related to secondary classification is two representing the more specific age range in its domain.

I did read this article : (<http://web.eecs.umich.edu/~someshs/nn/matlab_nn_starter.htm>) but finally I've never been success in nnetwork creation.

It's my network definition :

net =network;

net.numInputs=138;

for i=1:136

      net.inputs{i}.size = 1;

end

net.numLayers = 3;

net.layers{1}.size = 69;

net.layers{2}.size = 69;

net.layers{3}.size = 4;

net.inputConnect(1) = 1;

net.layerConnect(2, 1) = 1;

net.layerConnect(3, 2) = 1;

net.outputConnect(3) = 1;

net.targetConnect(3) = 1;

net.layers{1}.transferFcn = 'logsig';

net.layers{2}.transferFcn = 'purelin';

net.layers{3}.transferFcn = 'purelin';

net.biasConnect(1) = 1;

net.biasConnect(2) = 1;    %how can I connect this bias to gender?

net.biasConnect(3) = 1;

net.initFcn = 'initlay';

net = init(net);

net.performFcn = 'mse';

net.trainFcn = 'trainlm';

I want to know how can I define, connect, train and test these networks using matlab?

Thanks guys.

[Greg Heath]

There is no such thing as a BP Neural Network. There are a number of NN training algorithms. BP is just the initials of the most common one. I assume you mean a feedforward multilayer perceptron (FFMLP or just MLP). Given a trained FFMLP you cannot tell if it was trained by BP or another training algorithm.

What version of the NNToolbox do you have?

Do you have access to source code via the function TYPE?

type feedforward

Hope this helps.

Greg

[Greg Heath]

Have you considered k-means clustering, using a radial basis function net or using a LVQ net? These aproaches tend to allow a better visualization and understanding of the class division.

Hope this helps.

Greg

[Younes Jafari]

Hi Greg,

thanks for your comment.

I'm using Matlab 2011 NNToolbox. Put that nnetwork source code here please!

Answer 1

[Greg Heath]

If the net

1. MUST be a 2-hidden-layer FFMLP

2. MUST use the 1st hidden layer to recognizably separate the 4 classes

3. MUST be trained via BackProp

Try versions of the following:

1. For each class, design a 136-1-1 two-class classifier using patternnet(1). Since the initial weights are random, design many ( 10 each?) and choose the best.

2. Use the weights of the 4 selected classifiers for the first layer weights of a 136-4-4 four-class classifier using patternnet([4 H]).

3. Choose a good value for H by trial and error.

4. If you cannot freeze the 1st layer weights with learning rates of zero, store the hidden layer outputs of the two-class classifiers to train the last two layers of the four-class classifier.

Hope this helps.

Greg

Answer 2

[Younes Jafari]

Thanks Greg! Can you leave some code here? I'd edit my first post, please check it again. I want to know how can I implement your suggest?

please help me.