Welcome to the HNN development efforts

[Jorden Mauro]

On Thu, Jan 28, 2010 at 10:32 AM, Alp Mestan <alpmestan at gmail.com> wrote:
> Hi,
> I think all the people interested in HNN so far have subscribed to this
> mailing list so here we go !
> First, thank you all for your interest and particular thanks go to Thomas
> who joined the development and is experimenting together with me some design
> ideas for the next version.
> About that next version... There are *many* goals to achieve, and I think we
> should discuss them and establish a feature/goal list we must achieve for
> 0.2 (not sure we'll name it 0.2 if we really go the way Thomas and I are
> thinking about).
> A primary goal is to make the design much more general. The current version
> is very weak on that point and thus we are rethinking the structural and
> computational aspects of HNN. Our current experiments lean on a graph
> structure that we traverse during the evaluation. We're also trying to find
> a way to let GHC some place to do some optimizations regarding the
> computations and compositions of bunch-of-neurons computations. We're not
> there yet but already thinking about that.
> An orthogonal goal is to provide nice tools to work with neural networks.
> I'm particularly thinking about Jorden's work on Genetic Algorithms for
> altering weights, a tool for visualizing neural networks (it'll be very easy
> if we stick with the graph-based approach), and so on.
> And, finally, a reasonably-sized neural network does a lot of computations,
> and that's why we also will have to introduce optimizations to make HNN a
> serious candidate for real-world neural network usage. However, until we
> have the basis "ready", we won't be able to think about them, since it
> depends quite much on what will be there.
> Do you see any other goal we may have to achieve for the next version of HNN
> ?

On the scalability front: anything that can be done to facilitate
parallelization and distribution of computations might be important.

Eventually, people will be asking for bells and whistles like Hopfield
neural networks and PCNNs.

> And an auxiliary question : what module hierarchy can you think of for HNN ?
> Thank you all for your time and interest.
> --
> Alp Mestanogullari
> http://alpmestan.wordpress.com/
> http://alp.developpez.com/
>
> _______________________________________________
> Hnn mailing list
> Hnn at projects.haskell.org
> http://projects.haskell.org/cgi-bin/mailman/listinfo/hnn
>
>