Doesn’t training need higher precision to avoid getting stuck at local minima, at least with back propagation style learning?
Maybe something alternate like evolutionary algorithms could work in a domain like this, but so far those have proven to be less efficient.