Dear fellow programmers..

I have been working with an quite complex OCR project, with some heavy image preprocessing and character separation. And now im looking for a simple and quick way to do the actual comparing. Some friends suggested Neural networks (NN) to handle that for me, and after running some demo programs i must say im quite impressed. This will most likely save me tons of time/work...

... but the problem is that neither me or any of my friends really have any proir NN experience, and its quite hard to come by some userfriendly tutorials on the subject. Tried both google and IRC without any good results.

So now im left clueless with no where to turn really..

This is how far ive gotten with the OCR comparing procedure so far:

http://cpp.sourceforge.net/?show=21304

Actually works quite well, but once i feed it more then ~25-35 patterns to train on, it goes bananas. Gives really wierd comparing results and takes about 100 times longer to train.

Please help me, growing quite desperate here...

EDIT:

A breif terminology lesson could also be in place, eg what an "epoch" is...

Thanks in advance, Alex

From what I've read, ANNs has difficulties to generalize characters
that has been rotated, stretched - locally or globally, noise and
personal styles. The mainstream I believe is to use Kernels. It is
highly(!) recommendable that you google for a few articles that are
similar to your approach, read what they did, mimic it and improve
with that experience behind you.

ANNs are a lot simpler than many Kernels however. If you haven't had
any experience of the kind before, ANNs is a much better start. You
should to begin with have them working in stand-alone tests, i.e.
not OCR. Go from simple to complex, linear threshold units to
sigmoid backpropagation, from single layers to multiple layers. It
is a difficult task - especially to verify that it Does work, not
just that it seems to - but keep at it and it should fall into place
eventually. If you would get stuck, you can always ask us follow
online companions.
To learn how ANNs work, I'd suggest reading the chapter in
Mitchell's "Machine Learning" aside from reading the first few
chapters of "Machine Learning: A probabilistic approach". Virtually
everything it says up to the neural nets chapter is good - and
covers OCR with k-nearest neighbours, an even easier approach. "The
nearest neighbour method makes 18 errors out of the 600 test
examples."
http://www.idiap.ch/~barber/mlgm.pdf

Many of your questions should be answered in them. The weight update
ensued by a sample is called an epoch. If you have a training set of
ten items and do five passes, there's a total of 50 epochs.

Can you give specific examples or cite references to your assertion that kernel methods are superior to neural networks?

-Will Dwinnell
http://will.dwinnell.com

I said they were mainstream, as far as I know, neither dominates the
other (i.e. in all respects). There seems to be ways to cope with
what I mentioned above and I don't know which one would be the more
complicated for high levels of accuracy. By Kernels, I particularly
referred to SVMs.