As a matter of short term practicality currently we don’t have the hardware for GPU acceleration. This limits the things we can do, but at this stage of learning most of the time spent is on understanding and implementing the basic concepts anyway.
For what you’re doing, GPU stuff probably doesn’t make that big of a difference. Convolutional networks will train and run faster, but a digit recognition network should be tiny and fast anyway.
For what you’re doing, GPU stuff probably doesn’t make that big of a difference. Convolutional networks will train and run faster, but a digit recognition network should be tiny and fast anyway.