I shared this AI code with nVidia and I suppose someone over here might be interested too:
https://devtalk.nvidia.com/default/topic/1023786/cuda-programming-and-performance/walsh-hadamard-tra...
The overall idea I want to try is to combine a deep neural network with "soft" associative memory (AM) to make an aLife. A deep network would find it extremely difficult to learn how to use digital random access memory like DRAM. However AM allows sloppy inputs to produce sloppy outputs, unlike digital memory which would give completely the wrong result with any error in the input.
Using evolution such a system should be able to learn any algorithms it needs to survive assuming the combination is Turing complete.