Recommendations for Neural Network Modeling Software?

iconian asks: "I'm looking for a freeware or a low-cost neural network modeling software to do some academic research and perhaps also to teach an undergraduate course. I need something that is not difficult to use and doesn't require an extensive programming background. I've tried PDP++ and I've found it to be very powerful but a bit overwhelming because of the programming needed to perform seemingly minor tasks. Right now, I'm trying out SNNS with the Java GUI but the interface is not as intuitive as PDP++ and documentation is a bit sparse. Do you guys have any suggestions?"

Since processing power is now cheap...

[Anonymous Coward]

To compute f(x), just search the training set for the x0 closest to x (in some reasonable metric), and return f(x0) as your answer. You'll get much better results than with neural nets.

PDP++

[IAmATuringMachine!]

While slightly biased (as a PDP++ user) I find that it really is the best tool for the job, though for doing small networks it can be a pain, for the "real thing" it can't be beat, both in speed and in refinement of the interface. Randall O'Reilly, the author, has released a new book which would serve as an undergraduate/graduate text that has experiments in it which revolve around using PDP++ - they provide good practice. Currently, though, I don't think that the documentation is too stellar, so I agree it might be a bit tough. SNNS, the Stuttgart Neural Network Simulator is pretty okay too, and like PDP++, is available for several platforms - http://www-ra.informatik.uni-tuebingen.de/SNNS/ Other than those, my mind is drawing a blank. Good luck!

--
"Computer Science is no more about computers than astronomy is about telescopes."

A few suggestions...

[grammar nazi]

Hi,

We use NN's for similar purposes at my work. The particular ones that we use are Neuralware's NGO, SNNS (without the JavaGui) and the Nueral Network Toolbox for Matlab.

Since we distribute the code for the final NN, we've found that writing up our own C NN to be ideal. There are many books that tell you how to do this. A feedforward linear NN is very easy to program yourself. The trainer is what get's complicated. The best/cheapest solution that we've found is to use trainers from COTS packages, such as those that I mentioned, and then implement the trained NN with our own code.

As we evolve, we are slowly developing our own trainer with features that are specific to my area of work.

The comp.ai.neural-nets FAQ

[hengist]

at ftp://ftp.sas.com/pub/neural/FAQ.html [sas.com] lists several free packages. Source code packages are listed in this [sas.com] section.

That said, if you have the time, implementing it yourself is a great way of learning the algorithms. My PhD topic is neural networks, and I 've always found that implementing a particular architecture myself gives me the highest level of understanding and satisfaction.

Re:Since processing power is now cheap...

[hengist]

Sounds a lot like Kohonen's Self Organising Map. SOMs would potentially be more efficient.