|From:||Jean Michel Sellier|
|Subject:||Re: [Gneuralnetwork] Joining the project.|
|Date:||Wed, 24 Aug 2016 09:57:17 +0200|
I've been writing neural network software in C, with a commitment to
having no dependencies beyond system libraries, and I've gotten
fairly far myself. But I've decided, for several reasons, that I
want to be part of a project with more people. Gneuralnetwork
seemed like the right one technically.
I have kind of a lot of stuff I'm ready to do more-or-less immediately.
Some of it will be imports from my personal project and some will be
mostly about removing limitations (like number of incoming connections
to a node) from the system. How much infrastructure depends on
gneuralnetwork's current configuration (ie, how much damage would
be done by a change affecting how people use and configure it)? I
think I can be compatible with extant configuration methods, but
it's important to know....
My primary interests at this time (and the bulk of stuff I'd want to
contribute in the near term) are about large, deep, recurrent, and
topologically "weird" networks. It is likely that some of the things
I'd want to add will not be amenable to vectorization; I think that
I can add them without altering the degree to which extant code can
be used and could be vectorized, but it would mean that there are
multiple execution paths; one (easily parallelizable) that gets used
in layer-structured feedforward networks and one (less so) that gets
used for topologically strange networks where there are lots of
skip-layer connections or in some cases where there is no layer
structure at all.
PS. I think the project needs some benchmark data sets and
sample configurations. Both for new users and for testing
to make sure when we submit code that we didn't break stuff.
pps. Public Key Attached.
|[Prev in Thread]||Current Thread||[Next in Thread]|