Choosing training algorithm and performance function for neural network

I am trying to do my first research task with artificial neural networks (ANNs) task. The task is pattern recognition for biometrics. I have just basic knowledge and background in ANNs. I have some books about this topic, but most of them contain information about just different types of ANNs, and not about training algorithms and/or performance functions. MATLAB help does not have information in detail about differences between training algorithms. I have some basic questions about ANNs in MATLAB and will appreciate any help from the community.

1. Where I can find information about differences between training algorithms and performance functions and their applications in practice?

2. I am trying all possible combinations of training algorithms and performance functions. I have very good recognition results for some combinations, others have very bad recognition results. Is it possible to improve bad results by changing such training parameters like maximum number of epochs, validations number, steps etc.?

3. Is there any strategy to find optimal training parameters for ANNs for particular task?

4. Is it possible to use ANNs for recognition (classification - I hope these task are the same) task with many (thousands) classes like for biometric task? If not, why?

Thanks in advance.

Answer 1

[Greg Heath]

Use patternnet for classification.

 help patternnet

 doc patternnet

Search on

 greg patternnet

Practice on a MATLAB classification dataset

 help nndatasets

 doc nndatasets

Hope this helps.

Thank ou for formally accepting my answer

Greg

[Evgeny]

This answers to just a small part of my questions, but thank you, Greg!

[Greg Heath]

Except for algorithm comparisons, what info could you not find?