Code Generation for Neural Networks Based on Fixed-point Arithmetic

Abstract

Over the past few years, neural networks have started penetrating safety critical systems to make decisions as, for example, in robots, rockets, and autonomous driving cars. Neural networks based on floating-point arithmetic are very time and memory consuming, which are not compatible with embedded systems known to have limited resources. They are also very sensitive to the precision in which they have been trained, so changing this precision generally degrades the quality of their answers. To deal with that, we introduce a new technique to generate a fixed-point code for a trained neural network. This technique is based on fixed-point arithmetic with mixed-precision. This arithmetic is based on integer operations only, which are compatible with small memory devices. The obtained neural network has the same behavior as the initial one (based on the floating-point arithmetic) up to an error threshold defined by the user. The experimental results show the efficiency of our tool SyFix in terms of memory saved and the accuracy of the computations.