Design methodology of sigmoid functions for Neural Networks using lookup tables on FPGAs

EasyChair Preprint 583

Abstract

Sigmoid functions are used as transfer functions in artificial Neural Networks (NN). Normally, NNs are implemented in processors systems, both in training and testing phases. But in some scenarios these systems do not reach real time operation. In these cases, the NNs can be implemented in specific digital devices. For prototypes design it is convenient to use Field Programmable Gate Arrays (FPGA). The sigmoid functions are non-linear systems; therefore, they are not directly implementable in fixed point format, and some approximations are used. A very used one is the lookup table technique. In this paper, an advanced design method based on Matlab and Simulink is presented. It allows scan the number of samples and the number of fractional bits in input and output. The Signal to Noise Relation power (SNR) is used to measure the approximation functionality. This allows to observe linearities in physical performances against the number of bits address bus or the number of words in the lookup table. The automatic generation code to a Hardware Description Language (HDL) is possible. The HDLs can be Very High Speed Integrated Circuit Hardware Description Language (VHDL) or Verilog.

Keyphrases: FPGA, MATLAB, Sigmoid function, Simulink, VHDL, Verilog, design methodology, fixed point, floating-point, lookup table, neural network

@booklet{EasyChair:583,
  author    = {Santiago Tomás Pérez Suárez},
  title     = {Design methodology of sigmoid functions for Neural Networks using lookup tables on FPGAs},
  doi       = {10.29007/xg75},
  howpublished = {EasyChair Preprint 583},
  year      = {EasyChair, 2018}}