Now that we know about this, that the calculations in the trained models follow some particular forms, is there an approximation algorithm to run the models without GPUs?