I was thinking that by choosing a minimum weight value, a maximum weight value and a step value, one could iterate over all weight value combinations in the range to find the best fitting weights. This could be repeated after the initial 'best' weights were found by searching a smaller range around each weight again with a smaller step value.
On a GPU, it should be possible to calculate the results in parallel for many weight sets by computing one layer at the time.
[link][8 comments]