No validation and test set in feedforwardnet

Hello. I'm training a feedforwardnet to approximate y = sin(x^2) and I've noticed something strange going on. When I use con2seq, the data is not automatically split in train, validation and test sets, so the network overfits easily (see pictures below). This happens for all training algorithms. Without con2seq there are 3 sets, but the training accuracy is always much worse. Any ideas as to what may be causing this? Thanks!

Code with con2seq:

    x = 0:0.1:3*pi;
    y = sin(x.^2);
    x = con2seq(x);
    y = con2seq(y);
    net = feedforwardnet(30,'trainlm');
    net.trainParam.epochs = 200;
    [net, tr] = train(net, x, y);
    targets_hat = sim(net, inputs);

Same, but without con2seq:

    x = 0:0.1:3*pi;
    y = sin(x.^2);

    net = feedforwardnet(30,'trainlm');
    net.trainParam.epochs = 200;
    [net, tr] = train(net, x, y);
    targets_hat = sim(net, inputs);

[Greg Heath]

1. Always begin with the function examples

 help feedforwardnet
 doc feedforwardnet

2. Always initialize the RNG at the beginning of the code.

 rng('default')
   [x,t] = simplefit_dataset;
     net = feedforwardnet(10);
     net = train(net,x,t);
     view(net)
     y = net(x);
     perf = perform(net,t,y)
 perf =
          1.4639e-04

3. Check to make sure you know what the performance function is

 e   = t-y ;    % error
 MSE = mse(e)  % 1.4639e-04

4. Normalize the mse by comparing with the simple model where the output is assumed to be the average of the target rows and the corresponding mse is just the mean of the the target row variances:

 y0   = mean(t,2)               % 5.7254
 MSE0 = mse(t-y0)               % 8.3378
 check = MSE0 - mean(var(t',1)) % 0
 NMSE = MSE/MSE0                % 1.7558e-05

5. NOTE: Rsquare = 1 - NMSE (See Wikipedia)

6. Now repeat with your own data.

7. In general, I use a double for loop over a. number of hidden nodes (outer loop) b. random initial weights (inner loop)

The goal is to minimize the number of hidden nodes subject to an upper bound NMSE <= 0.01 (Rsquare >= 0.99)

I have posted zillions of examples in both the NEWSGROUP (comp.soft-sys.matlab) and ANSWERS.

Hope this helps.

Greg

[Olga Kravchenko]

This is not my question was about though. I know how to train a neural network in general, I'm wondering where the test and train sets disappear depending on the format of the data I feed in.

[Osama Tabbakh]

I have the same problem, there is no test and validation, when I use con2seq.

p %% input

p = con2seq (p);

t %% target or output

t = con2seq (t);

net = newff(p,t,[2 2]);

net.divideParam.trainRatio=0.7;

net.divideParam.testRatio=0.15;

net.divideParam.valRatio=0.15;

net.trainParam.lr=0.01;

net.trainParam.min_grad=1e-20;

net.trainParam.goal=1e-30;

net.trainParam.epochs=200;

net = train(net,p,t);

Unfortunately, I can't train without con2seq, because the Input and the target have different size.