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
Best Answer