Self-referential neural networks; first program in the new experimental programming language Fluid

It turns out that the best way to create self-referential generalized recurrent neural networks (dataflow matrix machines) is to allow the network to work with streams of matrices describing connectivity patterns and weights of the networks of this kind:

http://arxiv.org/abs/1605.05296

Today we mark the emergence of the first simple program written in the experimental dialect of the new Fluid programming language (a language for specifying dataflow matrix machines).