Neural network precision tuning

Abstract

Minimizing the precision in which the neurons of a neural network compute is a desirable objective to limit the resources needed to execute it. This is specially important for neural networks used in embedded systems. Unfortunately, neural networks are very sensitive to the precision in which they have been trained and changing this precision generally degrades the quality of their answers. In this article, we introduce a new technique to tune the precision of neural networks in such a way that the optimized network computes in a lower precision without modifying the quality of the outputs of more than a percentage chosen by the user. From a technical point of view, we generate a system of linear constraints among integer variables that we can solve by linear programming. The solution to this system is the new precision of the neurons. We present experimental results obtained by using our method.