1: \begin{abstract}

2: Conventional stochastic rounding (CSR) is widely employed in the training of neural networks (NNs), showing promising training results even in low-precision computations. We introduce an improved stochastic rounding method, that is simple and efficient.

3: %However, the gradient information is limited by a low rounding probability. The present study introduces an improved stochastic rounding method, that is tailored to training NNs for binary classification problems. 

4: The proposed method succeeds in training NNs with 16-bit fixed-point numbers and provides faster convergence and higher classification accuracy than both CSR and deterministic rounding-to-the-nearest method. 

5: \end{abstract}

6: