International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869 (O) 2454-4698 (P) Volume-9, Issue-12, December 2019
5 www.erpublication.org
Abstract—This paper presents a brief theoretical review on
deep neural network architectures, deep learning procedures,
as well assome of its possible applications. The paper focuses on
the most common networks structures: Convolutional Neural
Network (CNN), Deep Belief Network (DBN) and Stacked
Auto-encoders (SA). The building blocks which enabled the
construction of deeper networkssuch as Rectified Linear Unit
(ReLU) and softmax activation functions, convolution filters,
restricted Boltzmann machines and autoencoders, are
explained in the beginning and middle sections of the paper. A
few examples of hybrid systems are also presented at the last
sections of the paper. The paper concludes with some
considerations on the state-of-art work and on the possible
future applications enabled by deep neural networks.
Index Terms— Autoencoder, Boltzmann Machine,
Convolutional Neural Network, Deep Leaning Review.
I. INTRODUCTION
Neural networks algorithms have been applied to a wide
variety of complex tasks, inareas ranging from computer
vision, speech recognition, text translation, system
identification and control, among others.
The greatest advantage of this algorithm lies on their
ability to learn from a set of examples, without the need for
defining a set of explicit rules for a given task. After learning
how to solve a given problem, an artificial neural network
would generally perform in the same level or better than a
rule-based algorithm for the same task, especially for very
abstract problems such as in computer vision.
While neural networks were shown to theoretically be able
to represent any nonlinear function [1], in practice neural
networks were limited in depth and by long training times.
What allowedneural networks to achieve the high level of
performance seen today was the development of a series of
techniques for training deep networks in the past decade. This
set of techniques is what is now known as deep learning.
This paper presents a brief theoretical review on deep
Rômulo Fernandes da Costa, GraduateProgram in
ElectronicEngineeringAnd Computer Science, ITA, São José dos Campos
Sarasuaty Megume Hayashi Yelisetty, Graduate Program in Electronic
Engineering And Computer Science, ITA, São José dos Campos, Brazil
Johnny Cardoso Marques, Computer Science Division, ITA, São José
dos Campos, Brazil.
Paulo Marcelo Tasinaffo, Computer Science Division, ITA, São José
dos Campos, Brazil.
This work was funded by the Brazilian National Council for
Scientific and Technological Development (CnPq), in the form of
funding for the first author.
neural network’s structures, training procedures, and
enumerates some of its possible applications. The paper
focuses on presenting a general description on the inner
workings of the most common deep architectures, namely the
Deep Belief Networks (DBN), Stacked Autoencoders (SA)
and Convolutional Neural Networks (CNN).
In terms of structure, these three topologies can be
decomposed in fundamental blocks, such as the ReLU and
softmax activation functions, convolution filters, restricted
Boltzmann machines and autoencoders. These blocks, along
with the associated architectures, are described in the middle
sections of the paper.
A few examples of hybrid systemsare also presented at
later sections of the paper. The paper concludes with some
considerations on the state-of-art work and on the possible
future applications enabled bydeep neural networks.
II. BASIC CONCEPTS
A. Artificial Neuron Structure
An Artificial Neural Network (ANN) is a parallel
computational structure loosely inspired on real neural
networks,capable of learning from large amounts of data. The
network is trained to generate a set of outputs from the inputs
presents on the training data. Thus, an ANN can act as
anuniversal approximator of nonlinear functions [1].
These networks are composed of several small units, called
neurons or nodes, grouped in multiple sequential layers. Each
neuron in a layer receivessignals fromneurons localized in
other layers or from the network’s input itself.The neuron
then responds by emitting a signal of its own, propagating the
information forward to the next layers in the network.
The output signaly
n
fired by a neuron as a response to an
input vector x
n
isdescribed by:
(1)
Here,
and
are the connection weight vector and
activation bias respectively.The mathematical function
is
a nonlinear function called "activation function” and
describes the response of the neuron to its collective input.
Historically,
used to be simple linearfunctions (such as
in the original perceptron[2]) and sigmoid functions, but with
the popularization of deeper networks, less computationally
expensive options such as Rectified Linear Unit (ReLU)
started to be employed. Fig. 1 shows a plot ofsome of the
commonly used activation functions.
A Brief Didactic Theoretical Review on Convolutional
Neural Networks, Deep Belief Networks and Stacked
Auto-Encoders
MSc.Rômulo Fernandes da Costa, MSc.Sarasuaty Megume Hayashi Yelisetty, Dr. Johnny
Cardoso Marques, Dr. Paulo Marcelo Tasinaffo