[SciPy-User] Artificial Neural Network thoughts

[Gael Varoquaux]

On Sat, Oct 30, 2010 at 12:43:18PM +0200, Gerrit Holl wrote:
> There are others. That is in fact a disadvantage.

Welcome to a world of choice and meritocracy. Choice is hard, for the
user, but also for the developers. Different packages reflect different
choices from the developers. Who can tell which choice is best without
time for things to settle down? In my opinion, finding the right approach
to neural-network-related studies is still very much an open research
question.

> What experiences do people have in using Artificial Neural Networks in
> Python+friends? What packages are suitable for production code and can
> be more or less relied on without having to understand all the
> details? Is any package much more widely used than the others?

I don't have experience. I don't do ANN (I tend to prefer other
approaches to machine learning, and you have plenty of choice in Python
for these too). But from what I hear I would suggest that you look at:

* pybrain (http://pybrain.org/). It's a well-polished library with good
  documentation done by a research community that focuses on neural
  networks. My impression is that it will not scale on large datasets,
  but then some might say that neural networks won't scale on large
  datasets without relying on specificities of the learning problem :).

* pylearn (http://deeplearning.net/software/pylearn/) is probably promising 
  as these guys are leading the research field in modern approaches to ANN,
  and have shown very interesting approaches to programming (theano). It is
  probably a bit too early to use it as an end product, though. These
  guys however want to scale on large problems.

HTH,

Gaël