Deep Neural Network Training without Multiplications
This work addresses the computational cost of multiplications in deep neural network training and inference, potentially benefiting hardware designers and practitioners seeking more efficient models.
This paper explores replacing floating-point multiplications with integer-add instructions for training deep neural networks. They demonstrate that ResNet can be trained with competitive classification accuracy, sometimes matching FP32 baseline results, without requiring additional stabilization methods.
Is multiplication really necessary for deep neural networks? Here we propose just adding two IEEE754 floating-point numbers with an integer-add instruction in place of a floating-point multiplication instruction. We show that ResNet can be trained using this operation with competitive classification accuracy. Our proposal did not require any methods to solve instability and decrease in accuracy, which is common in low-precision training. In some settings, we may obtain equal accuracy to the baseline FP32 result. This method will enable eliminating the multiplications in deep neural-network training and inference.