Funn: Functional Neural Networks in Haskell
This is an experimental library exploring a combinator approach for building and training neural networks in haskell.
Traditional (eg. in C) libraries construct neural networks monolithically, by providing a comprehensive list of the layers' topologies to a function. The approach used in this library is an attempt at a composable system, in which networks are built by connecting smaller units together:
As let one = fcLayer :: Network m (Blob 10) (Blob 20) is a fully connected 10x20 layer, and let two = sigmoidLayer :: Network m (Blob 20) (Blob 20) is a sigmoid activation layer, we can compose them directly by feeding the output of the first into the second.
one >>> two :: Network m (Blob 10) (Blob 20)
Parts of the interface are still quite ad-hoc and subject to change.
MHUG Talk
The slides in /mhug-talk-15 describe a mini talk I presented at the Melbourne haskell user group.