A Systematic Literature Review of Hardware Neural Networks

Although Neural Networks (NN) are extremely useful for the solution of several problems such as object recognition and semantic segmentation, the NN libraries usually target devices which face several drawbacks such as memory bottlenecks and limited efficiency (e.g. GPUs, multi-core processors). Fortunately, the recent implementation of Hardware Neural Networks aims to tackle down this problem and for that reason several researchers had turn back their attention to them. This paper presents the Systematic Literature Review (SLR) of the most relevant HNN works presented in the last few years. The main sources chosen for the SLR were the IEEE Computer Society Digital Library and the SCOPUS indexing system, from which 61 papers were reviewed according to the inclusion and exclusion criteria, and of which after a detail assessment, only 20 papers remained. Moreover, the increase in the number of papers per year reflects that the interest in HNN had been growing up. Finally, the results show that the most popular NN hardware platforms are the FPGAs-based.