My idea is simple.
If you train the parameters of the machine learning setup using a genetic algorithm you can put in almost anything in the system.
From my trial with smooth gaussian blur filters I realised that I could use the weights as parameters for a 2D spline function. From this I get pretty much any size of a weight matrix for free.
Because I believe you need to noise filter the weights. Its ridiculus to have so much gradient supporting weights. Just use a spline function or 2D spline surface.