Dynamically Adapting Floating-Point Precision to Accelerate Deep Neural Network Training | Publicación