Probabilistic rounding in neural network learning with limited precision