Você está na página 1de 10

Artificial Neural Networks 1

C ONTENTS

I Introduction 3
I-A History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

II Design of a Neural Network 3

III Comparison of Artificial and Biological Neuron 4

IV Structure of Neural Network 5

V Neural Network Models 6


V-A MLP model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
V-B Feed-forward networks . . . . . . . . . . . . . . . . . . . . . . 6
V-C Feedback networks . . . . . . . . . . . . . . . . . . . . . . . . 6
V-D Network layers . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

VI Applications of Artificial Neural Networks 7

VII Limits to Artificial Neural Networks 8

VIII Future of Artificial Neural Networks 9

IX Conclusion 9

References 9

Information Technology Government Engineering College, Barton Hill


Artificial Neural Networks 2

Artificial Neural Networks

A BSTRACT
An artificial neural network (ANN), usually called "Neural Network" (NN), is a
mathematical model or computational model that is inspired by the structure and/or
functional aspects of biological neural networks. It consists of an interconnected group
of artificial neurons and processes information using a connectionist approach to
computation. In most cases an ANN is an adaptive system that changes its structure
based on external or internal information that flows through the network during the
learning phase. Modern neural networks are non-linear statistical data modeling tools.
They are usually used to model complex relationships between inputs and outputs or
to find patterns in data.

Information Technology Government Engineering College, Barton Hill


3

Artificial Neural Networks


Sreekanth H
Department of Information Technology
Government Engineering College, Barton Hill
h.sreekant@gmail.com

Abstract—An artificial neural network A. History


(ANN), usually called "Neural Network" (NN),
is a mathematical model or computational
Neural network simulations appear to
model that is inspired by the structure be a recent development. However, this
and/or functional aspects of biological neural field was established before the advent
networks. It consists of an interconnected of computers, and has survived at least
group of artificial neurons and processes one major setback and several eras.Many
information using a connectionist approach importand advances have been boosted by
to computation. In most cases an ANN is an
adaptive system that changes its structure
the use of inexpensive computer emula-
based on external or internal information that tions. The earliest work in neural com-
flows through the network during the learning puting goes back to the 1940’s when
phase. Modern neural networks are non-linear McCulloch and Pitts introduced the first
statistical data modeling tools. They are neural network computing model. In the
usually used to model complex relationships 1950’s, Rosenblatt’s work resulted in a
between inputs and outputs or to find patterns
two-layer network, the perceptron, which
in data..
was capable of learning certain classifi-
cations by adjusting connection weights.
I. I NTRODUCTION Although the perceptron was successful
in classifying certain patterns, it had a
A neural network is a powerful data number of limitations. The perceptron
modeling tool that is able to capture and was not able to solve the classic XOR
represent complex input/output relation- (exclusive or) problem. Such limitations
ships .In the broader sense , a neural led to the decline of the field of neural
network is a collection of mathematical networks. However, the perceptron had
models that emulate some of the observed laid foundations for later work in neural
properties of biological nervous systems computing.
and draw on the analogies of adaptive
biological learning. It is composed of
a large number of highly interconnected II. D ESIGN OF A N EURAL N ETWORK
processing elements that are analogous The brain is principally composed of
to neurons and are tied together with about 10 billion neurons, each connected
weighted connections that are analogous to about 10,000 other neurons. Each neu-
to synapses. ron receives electrochemical inputs from
Artificial Neural Networks 4

other neurons at the dendrites. If the sum • Neural networks modify own topol-
of these electrical inputs is sufficiently ogy just as neurons in the brain can
powerful to activate the neuron, it trans- die and new synaptic connections
mits an electrochemical signal along the grow.
axon, and passes this signal to the other
neurons whose dendrites are attached at III. C OMPARISON OF A RTIFICIAL
any of the axon terminals. These at- AND B IOLOGICAL N EURON
tached neurons may then fire. So, our Comparing a artificial neuron and a
entire brain is composed of these inter- neuron of a brain with the help of
connected electro-chemical transmitting figures 1 and 2 given below:
neurons. From a very large number of
extremely simple processing units (each
performing a weighted sum of its inputs,
and then firing a binary signal if the total
input exceeds a certain level) the brain
manages to perform extremely complex
tasks. This is the model on which arti-
ficial neural networks are based. Neural
network is a sequence of neuron layers.
A neuron is a building block of a neural
net. It is very loosely based on the brain’s
Figure 1. Artificial Neuron
nerve cell. Neurons will receive inputs via
weighted links from other neurons. This
An artificial neuron is a mathemat-
inputs will be processed according to the
ical function conceived as a crude
neurons activation function. Signals are
model, or abstraction of biological
then passed on to other neurons. In a more
neurons. Artificial neurons are the
practical way, neural networks are made
constitutive units in an artificial neu-
up of interconnected processing elements
ral network. Depending on the spe-
called units which are equivalent to the
cific model used, it can receive dif-
brains counterpart ,the neurons. Neural
ferent names, such as semi-linear
network can be considered as an artificial
unit, Nv neuron, binary neuron, lin-
system that could perform "intelligent"
ear threshold function or McCulloch-
tasks similar to those performed by the
Pitts neuron. The artificial neuron
human brain. Neural networks resemble
receives one or more inputs (repre-
the human brain in the following ways:
senting the one or more dendrites)
and sums them to produce an output
• A neural network acquires knowl- (synapse). Usually the sums of each
edge through learning node are weighted, and the sum is
• A neural network’s knowledge is passed through a non-linear function
stored within inter-neuron connec- known as an activation function or
tion strengths known as synaptic transfer function. The transfer func-
weights tions usually have a sigmoid shape,

Information Technology Government Engineering College, Barton Hill


Artificial Neural Networks 5

but they may also take the form of


other non-linear functions, piecewise
linear functions, or step functions.
They are also often monotonically
increasing, continuous, differentiable
and bounded.

Figure 3. Perceptron

before it terminates.

IV. S TRUCTURE OF N EURAL


N ETWORK
According to Frank Rosenblatts the-
Figure 2. Biological Neuron
ory in 1958 ,the basic element of
a neural network is the perceptron,
which in turn has 5 basic elements:
A neuron also known as a neurone or an n-vector input, weights, summing
nerve cell is an electrically excitable function, threshold device, and an
cell that processes and transmits in- output. Outputs are in the form of -1
formation by electrical and chemical and/or +1. The threshold has a setting
signaling. Chemical signaling occurs which governs the output based on
via synapses, specialized connections the summation of input vectors. If the
with other cells. Neurons connect summation falls below the threshold
to each other to form networks. A setting, a -1 is the output. If the sum-
typical neuron possesses a cell body mation exceeds the threshold setting,
(often called the soma), dendrites, +1 is the output. Figure 3 depicts the
and an axon. Dendrites are filaments structure of a basic perceptron which
that arise from the cell body, often is also called artificial neuron.
extending for hundreds of microns The perceptron can also be dealt as
and branching multiple times, giving a mathematical model of a biological
rise to a complex "dendritic tree". neuron. While in actual neurons the
An axon is a special cellular filament dendrite receives electrical signals
that arises from the cell body at a site from the axons of other neurons, in
called the axon hillock and travels for the perceptron these electrical signals
a distance, as far as 1 m in humans or are represented as numerical values.
even more in other species. The cell A more technical investigation of a
body of a neuron frequently gives single neuron perceptron shows that
rise to multiple dendrites, but never it can have an input vector X of
to more than one axon, although the N dimensions(as illustrated in fig-
axon may branch hundreds of times ure.4). These inputs go through a

Information Technology Government Engineering College, Barton Hill


Artificial Neural Networks 6

vector W of Weights of N dimension.


Processed by the Summation Node,
"a" is generated where "a" is the
"dot product" of vectors X and W
plus a Bias. "A" is then processed
through an activation function which
compares the value of "a" to a prede-
fined Threshold. If "a" is below the
Threshold, the perceptron will not
fire. If it is above the Threshold, the
Figure 5. Graphical representation of MLP
perceptron will fire one pulse whose
amplitude is predefined
model can then be used to produce
the output when the desired output
is unknown.

B. Feed-forward networks
Feed-forward ANNs allow signals to
travel one way only; from input to
output. There is no feedback (loops)
i.e. the output of any layer does not
affect that same layer. Feedforward
ANNs tend to be straight forward
networks that associate inputs with
outputs. They are extensively used
in pattern recognition. This type of
Figure 4. Mathematical model of a perceptron organisation is also referred to as
bottom-up or top-down.
V. N EURAL N ETWORK M ODELS
A. MLP model C. Feedback networks
The most common neural network Feedback networks can have signals
model is the multilayer perceptron travelling in both directions by intro-
(MLP). It is composed of of hier- ducing loops in the network. Feed-
archical layers of neurons arranged back networks are very powerful and
so that information flows from the can get extremely complicated. Feed-
input layer to the output layer of back networks are dynamic; their
the network. The goal of this type ’state’ is changing continuously un-
of network is to create a model that til they reach an equilibrium point.
correctly maps the input to the out- They remain at the equilibrium point
put using historical data so that the until the input changes and a new

Information Technology Government Engineering College, Barton Hill


Artificial Neural Networks 7

equilibrium needs to be found. Feed- than hierarchically structured multi-


back architectures are also referred to layer organisations. In multi-layer
as interactive or recurrent, although networks, units are often numbered
the latter term is often used to denote by layer, instead of following a
feedback connections in single-layer global numbering.
organisations.

D. Network layers VI. A PPLICATIONS OF A RTIFICIAL


N EURAL N ETWORKS
The commonest type of artificial
neural network consists of three Neural networks have broad applica-
groups, or layers, of units: a layer of bility to real world business prob-
"input" units is connected to a layer lems. In fact, they have already
of "hidden" units, which is connected been successfully applied in many
to a layer of "output" units. industries. Since neural networks are
– The activity of the input units best at identifying patterns or trends
represents the raw information in data, they are well suited for
that is fed into the network. forecasting needs. From the wide
– The activity of each hidden unit range of applications of neural net-
is determined by the activities of works,we deal with few of them as
the input units and the weights follows:
on the connections between the • Character Recognition:- The idea
input and the hidden units. of character recognition has become
– The behaviour of the output very important as handheld devices
units depends on the activity of like the Palm Pilot are becoming in-
the hidden units and the weights creasingly popular. Neural networks
between the hidden and output can be used to recognize handwritten
units characters.
This simple type of network is inter- • Image Compression:- Neural net-
esting because the hidden units are works can receive and process vast
free to construct their own repre- amounts of information at once,
sentations of the input. The weights making them useful in image com-
between the input and hidden units pression. With the Internet explosion
determine when each hidden unit is and more sites using more images on
active, and so by modifying these their sites, using neural networks for
weights, a hidden unit can choose image compression is worth a look.
what it represents. We also distin- • Stock Market:- The day-to-day
guish single-layer and multi-layer ar- business of the stock market is ex-
chitectures. The single-layer organ- tremely complicated. Many factors
isation, in which all units are con- weigh in whether a given stock will
nected to one another, constitutes go up or down on any given day.
the most general case and is of Since neural networks can examine
more potential computational power a lot of information quickly and sort

Information Technology Government Engineering College, Barton Hill


Artificial Neural Networks 8

it all out, they can be used to predict also one of the first to use a neural
stock prices. network chip.
• Food Processing:- The food indus- • Monitoring:- Networks have been
try is perhaps the biggest practical used to monitor the state of aircraft
market for electronic noses, assisting engines. By monitoring vibration
or replacing entirely humans. Inspec- levels and sound, early warning of
tion of food, grading quality of food, engine problems can be given.British
fish inspection, fermentation control, Rail have also been testing a simi-
checking mayonnaise for rancidity, lar application monitoring diesel en-
automated flavor control,monitoring gines.
cheese ripening, verifying if orange
juice is natural, beverage container
inspection, and grading whisky. VII. L IMITS TO A RTIFICIAL N EURAL
• Medicine:- One of the areas that N ETWORKS
has gained attention is in cardiopul- The major issues of concern today are
monary diagnostics. The ways neural the scalability problem, testing, verifica-
networks work in this area or other tion, and integration of neural network
areas of medical diagnosis is by the systems into the modern environment.
comparison of many different mod- Neural network programs sometimes be-
els. They assist doctors with their come unstable when applied to larger
diagnosis by analyzing the reported problems. The defence, nuclear and space
symptoms and image data such as industries are concerned about the issue
MRIs or X-rays. of testing and verification. The mathe-
• Target Recognition:- Neural net- matical theories used to guarantee the
works also play an important part in performance of an applied neural network
the warfield , particularly in fighter are still under development. The solution
jets and tanks.The tanks and planes for the time being may be to train and
are fitted with high resolution dig- test these intelligent systems much as we
ital cameras hooked upto a com- do for humans. Also there are some more
puter which would continually scan practical problems like:
the environment outside for possible
• the operational problem encountered
threats.
when attempting to simulate the par-
• Machine Diagnostics:- Detect when
allelism of neural networks. Since
a machine has failed so that the sys-
the majority of neural networks are
tem can automatically shut down the
simulated on sequential machines,
machine when this occurs.
giving rise to a very rapid increase in
• Signature Analysis:- Neural nets
processing time requirements as size
can be used as a mechanism for
of the problem expands.
comparing signatures made (e.g. in a
• instability to explain any results that
bank) with those stored. This is one
they obtain. Networks function as
of the first large-scale applications of
"black boxes" whose rules of oper-
neural networks in the USA, and is
ation are completely unknown.

Information Technology Government Engineering College, Barton Hill


Artificial Neural Networks 9

VIII. F UTURE OF A RTIFICIAL IX. C ONCLUSION


N EURAL N ETWORKS The computing world has a lot
to gain from neural networks.
Their ability to learn by example
Neural networks are very good at makes them very flexible and
a wide variety of problems, better powerful. Furthermore there is
suited than traditional computer ar- no need to devise an algorithm
chitecture to problems that humans in order to perform a specific
are naturally good at and which com- task; i.e. there is no need to
puters are traditionally bad at ! And understand the internal mecha-
researchers are continually construct- nisms of that task. They are also
ing networks that are better at these very well suited for real time
problems. But will neural networks systems because of their fast re-
ever fully simulate the human brain? sponse and computational times
Will they be as complex and as which are due to their paral-
functional? Will a machine ever be lel architecture. Neural networks
conscious of its own existence? Of also contribute to other areas
course, the whole future of neural of research such as neurology
networks does not reside in attempts and psychology. They are regu-
to simulate consciousness. Indeed, larly used to model parts of liv-
that is of relatively small concern ing organisms and to investigate
at the moment; more pressing are the internal mechanisms of the
issues of how to improve the sys- brain. Perhaps the most exciting
tems we have. All current neural net- aspect of neural networks is the
work technologies will most likely possibility that some day ’con-
be vastly improved upon in the fu- scious’ networks might be pro-
ture.Everything from handwritng and duced. There is a number of sci-
speech recognition to stock market entists arguing that conscious-
prediction will become more sophis- ness is a ’mechanical’ property
ticated as researchers develop better and that ’conscious’ neural net-
training methods and network archi- works are a realistic possibility.
tectures. Neural Networks might, in Finally, I would like to state
the future, allow: that even though neural net-
– robots that can see, feel, and works have a huge potential we
predict the world around them will only get the best of them
– trends found in the human when they are integrated with
genome to aid in the understand- computing, AI, fuzzy logic and
ing of the data compiled by the related subjects.
Human Genome Project
– self-diagnosis of medical prob- R EFERENCES
lems using neural networks and [1] N.P. Padhy, "Artificial Intelligence and
much more Intelligent Systems", Oxford Univer-

Information Technology Government Engineering College, Barton Hill


Artificial Neural Networks 10

sity Press
[2] A. Joseph, D.B.L. Bong, D.A.A.
Mat,"Application of Neural Network
in User Authentication for Smart
Home System", World Academy of
Science, Engineering and Technology
[3] Bishop,"C.M. Neural Networks for
Pattern Recognition" Oxford Univer-
sity Press. 1995.

Information Technology Government Engineering College, Barton Hill

Você também pode gostar