Akademik Veri Yönetim Sistemi
Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi, cilt.9, sa.1, ss.39-49, 2017 (TRDizin)
Artificial Neural Networks (ANNs) are used in many
applications ranging from computer vision to speech
recognition. These algorithms give superior results
than conventional calculation methods. However, in
applications where the number of data is large, it
requires a lot of computation. For this reason, the
use of Graphics Processing Units (GPUs), which
speed up data processing by parallel processing of
data, has become mandatory. In recent years, deep
learning algorithms, a kind of ANNs, have been
successfully implemented in many real life
applications thanks to GPUs, and demand for
embedded systems containing GPUs has increased.
In this study, Deep Convolutional Neural Networks
(DCNNs), which are one of the deep learning
methods, were briefly introduced and the low cost
Nvidia Jetson TK1/TX1 embedded systems were
examined in general way. The Caffe program was
used to implement the deep learning algorithms.
Finally, the LeNet network was trained using MNIST
data on Nvidia Jetson TK1/TX1 boards with parallel
processing power. CPU and GPU modes of the
boards were used for training. Moreover, two
computers with i7 processor were used for
computing with the Nvidia Jetson TK1/TX1 boards.
The results were evaluated in terms of time and
accuracy.
Caffe is an open source platform developed by the
University of California, Berkeley to apply deep
learning algorithms. Codes are written in C ++ with
the CUDA library used for GPU computation. It
also supports the CUDA library, Python/Numpy,
and MATLAB.
This study used two computers with Nvidia GTX550
and GTX960 graphics cards, and Nvidia Jetson TK1
/ TX1 development cards. Handwriting recognition
was done as an application. The MNIST handwritten
dataset and LeNet network were used.
The MNIST dataset contains 60000 training images
and 10000 test images in 28x28 handwritten
numbers. MNIST dataset consisting of 0, 1, 2, 3, 4,
5, 6, 7, 8, and 9 handwritten digits.
LeNet is the first network structure in which
convolutional networks are successfully applied.
LeNet was developed by Yann LeCun in 1990. LeNet
architecture is used for other applications, but it is
generally used for numbers.
In the experiments, the handwritten numbers were
recognized by using the LeNet network. The
performances of both two boards, and the computer
were evaluated in terms of speed and accuracy. The
results showed that GPUs were faster than CPUs
although the accuracy rates were close to each
other.