Here are some small ideas. Some guessing on how to improve machine learning.
Its somewhat hard to make a function that creates noise. Therefore I wonder if you can improve it by letting the machine learning function just sample some already created noise. That is. You add a noise sample vector to the input vectors. Making the vector size larger.
The idea is that if the output data has data that looks random. Then it could be an idea to input some similar random looking data. For the transfer function to sample from. I mean.
Below is a test