I wonder if one could store only the binary representation at training and sample a floating point representation (both weights and gradient) during backprop.