Home         Authors   Papers   Year of conference   Themes   Organizations        To MES conference

The Technique for Implementing a Neural Network for Recognizing Handwritten Digits in FPGAs Based on Fixed Point Calculations  

Authors
 Solovyev R.A.
 Kustov A.G.
 Rukhlov V.S.
Date of publication
 2018
DOI
 10.31114/2078-7707-2018-3-126-131

Abstract
 The article proposes the implementation of a convolutional neural network in a FPGA for the solution of a practical problem, namely the recognition of handwritten figures from a video stream. The first part of the article is devoted to the adaptation of the MNIST dataset to the problem of recognition of numbers from a real camera. A new proposed training method allowed to avoid the long process of collecting data and to confine only to a small set in addition to the modified MNIST. The second part of the article is devoted directly to the method of implementing a neural network in a hardware device. Requirements for real-time video processing on a neural network imposed greater restrictions on the performance of device.
Thanks to the proposed approaches, it became possible to achieve the required performance characteristics and even exceed them by several times. Such indicators were achieved due to a number of optimizations. A neural network with a small number of weights and architectural features aimed at reducing consumed resources was designed. Calculations by a special algorithm are transferred from a floating point to a fixed point. It turned out that only 12 bits are sufficient to store the weights, so that there are no differences from the mathematical model. The proposed device was implemented and tested on a De0Nano debug card based on the Cyclone IV FPGA. The processing speed of the video stream exceeded 150 frames per second. The code for training the neural network and for generating the Verilog code of the target device was published in Open Source.
Keywords
 convolutional neural nets, FPGA, fixed point calculations, 2D convolution
Library reference
 Solovyev R.A., Kustov A.G., Rukhlov V.S. The Technique for Implementing a Neural Network for Recognizing Handwritten Digits in FPGAs Based on Fixed Point Calculations // Problems of Perspective Micro- and Nanoelectronic Systems Development - 2018. Issue 3. P. 126-131.
URL of paper
 http://www.mes-conference.ru/data/year2018/pdf/D035.pdf

Copyright © 2009-2019 IPPM RAS. All Rights Reserved.

Design of site: IPPM RAS