I'm hoping to use a simpler version of this neural net, which should reduce computation time by half. There are 8 weight matrices that are dynamically blended every frame ( the bulk of the computation ). This number can supposedly be reduced to 4 (according to original research).