FAPNN: An FPGA based Approximate Probabilistic Neural Network Library

Kizheppatt Vipin, Yerbol Akhmetov, Serikbolsyn Myrzakhme, Alex Pappachen James

2018

Closed access Metadata (CSL-JSON)
DOI 10.1109/coconet.2018.8476889
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Due to their flexible architecture and inherent parallelism, FPGAs are ideal candidates for neural network implementations. Still they have not achieved wide-spread acceptance in this regard. One of the major roadblocks for FPGAs is the implementation of complex mathematical functions encountered in neural networks. Exact implementation of these functions consume large number of resources. In this paper we discuss an FPGA-based neural network prototyping platform and the approximate implementation of a probabilistic neural network (PNN) on a Xilinx 7-Series FPGA. The complex mathematical functions as replaced by approximations. Analysis shows that hardware performance is much higher than that of software counter parts and the error induced due to approximations is within tolerable limit.