Patch for /trunk/core/learn/train_bp_netlab.m

[GoogleCodeExporter]

The number of hidden neurons should be defined. Changing the activation 
functions to linear. A 3-layer NN with linear activation units is equivalent to 
a 2-layer NN with nonlinear activation units. My tests has shown better 
accuracy with linear activation units than logistic activation units. 
Another addition is a random seed so that different NN weights are generated at 
each run. I am not sure if netlab handles this, but, it seems that netlab3.3 
does not have this feature.
Finally, we need few hidden neurons/units. The value nHidden=10 is a good 
choice. 

I would also suggest to reduce the defaults.epochs from 500 to 200 to speedup 
the training. I am using 100 now and it works pretty fine.

With very best regards,
Mohammed S. Al-Rawi,
Visual neuroscience lab, IBILI, University of Coimbra, Portugal
al-rawi(aaattt)uc(dddooottt)pt


With very best regards,
Mohammed S. Al-Rawi

Original issue reported on code.google.com by [email protected] on 23 Jan 2014 at 10:59

Attachments:

• train_bp_netlab.m.patch