Optimizing Neural Networks for Embedded Hardware

Abstract

Neural networks are a pervasive technology, which is, however, still held back in the area of embedded systems by the high resource requirements, especially memory size, memory access time and power dissipation. In recent years, several different methods have been proposed to transform given neural networks in such a way that they can get by with much fewer resources while maintaining almost the same accuracy. This work reviews, categorizes and describes the state of the art in adapting and simplifying neural networks to make them better applicable to embedded systems. Even though we developed this study from a purely automotive context, the techniques described are also valid in other areas.