Dataflow matrix machines (by Anhinga anhinga)

We know two things about machine learning schemes for "simple DAG images-as-matrices DMMs".

1) If we want to implement interactive evolution, we can follow the "population of directions' scheme, but because these are DAGs, it becomes much closer to CPPNs (still want to have 2D interactive sliders between 2 (or 3) points, and might add occasional backprop runs later).

2) If we want to implement sparsity-inducing L1-regularization, we can just run something like a novelty search (i.e. don't impose much of a goal function besides L1 itself + some behavior and/or weight novelty term).