A survey research summary on neural networks

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Special Issue: 03 | May-2014 | NCRIET-2014, Available @ http://www.ijret.org 385
A SURVEY: RESEARCH SUMMARY ON NEURAL NETWORKS
ShrutiB.Hiregoudar1
, Manjunath.K2
, K.S.Patil3
1
Dept. Computer Science, BasaveshwareEngineering College, Bagalkot, Karnataka, India
2
Dept. Computer Science, Basaveshware Engineering College, Bagalkot, Karnataka, India
3
Prof: Dept. Computer Science, Basaveshware Engineering College, Bagalkot, Karnataka, India
Abstract
Neural Networks are relatively crude electronic models based on the neural structure of the brain. The brain basically learns from
experience. It is natural proof that some problems that are beyond the scope of current computers are indeed solvable by small energy
efficient packages.In this paper we propose the fundamentals of neural network topologies, activation function and learning
algorithms based on the flow of information in bi-direction or uni-directions. We outline themain features of a number of popular
neural networks and provide an overview on their topologies and their learning capabilities.
Keywords: Neural Network, Feed Forward, Recurrent, Radial Basic Function Network (RBFN), Kohonen Self Organizing
Map (KSOM).
-----------------------------------------------------------------------***----------------------------------------------------------------------
1. INTRODUCTION
The human bairn has capabilities in processing information
and marking instantaneous decision. The many researchers
shown that the human brain make computations in a radically
different manner to that done by binary computers. The
neurons is a massive network of parallel and distributed
computing elements, many scientists are working last few
decades to build computational system called neural network,
which is also called as connectionist model.A neural network
is composed of set of parallel and distributed processing units
called nodes or neurons, these neurons are interconnected by
means of unidirectional or bidirectional links by ordering them
in layers.
Fig-1:Basic Structure of Neural Network
The basic unit of neural network is neuron, it consist of N no
of inputs to the network are represented by x(n) and each
input are multiply by a connection weight these weights are
represented by w(n).The product of input and weight are
simply summed and feed through a transfer function
(activation function) to generate the result (output).
2. NEURAL NETWORK DESIGN
Neural network mainly consist of three things
 Network Topology
 Network Transfer Function
 Network Learning Algorithm
2.1 Network Topology
the neural network topologies are classified based upon
interconnection are arranged with in the layer, there two well-
known neural network topologies are.
 Feed Forward Topology
 Recurrent Topology
2.1.1 Feed Forward Topology
In feed forward topology network, the nodes are hierarchically
arranged in layers starting with the input layers and ending
with output layers. In between the input layer and output layer
the number of hidden layers provide most of the network
computational power. The nodes in each layers connect to next
layer through uni-direction paths starting from one layer
(source) and ending at the subsequently layer (sink). The
output of a given layers feed the nodes of the following layers
in a forward directions and does not allow feedback flow of
information in the structure. Application multilayer layer
perception network and radial basic function network.

__________________________________________________________________________________________
Fig-2: Feed Forward Topology
2.1.2 Recurrent Topology
In (RNT) allows flow of information in between the connected
nodes in bi directions.i.e support both feed forward and
feedback. In recurrent network, structure has some sort of
memory which help topermit storageof information in
theiroutput nodes through dynamicstates. The mapping of
inputs and outputs are dynamic in nature.Application:
Hopfield Network and Time Delayed Neural Network
(TDNN).
Fig-3: Recurrent Topology
2.2 Network Transfer Function
The basic unit of neural network is neuron,these are sorts of
simple processors which take the weighted sum of their input
from other node and apply to them non-linear mapping
function called an activation function before delivering the
output to the next to the next neuron.
Fig-4: Transfer Function
2.3Neural Network Learning Algorithm
Learning algorithm are used to update the Weight parameter
of the input connections level of the neuron during the training
processes of the network. There are three types of leaving
algorithms are classified:
 Supervised
 Unsupervised and
 Reinforcement
2.3.1 Supervised Learning
In supervised learning mechanism,the external source provides
the network ith a set of input stimulus for which the output is a
priori known and during the running process the output results
are continuously compared with the desire data. After number
of iterations, the gradient descent rule uses the error between
the actual output and the target data to adjust the connections
weights so as to obtain the closest match between the target
out and the actual out. Application: feed forward network.
Fig-5: Supervised Learning
2.3.2 Unsupervised Learning Algorithm:
It is also called as self-organizing learning algorithm because
there is no any external source to provide the network and
relies instead upon local information and internal control.
Thetraining data and input pattern are presented to the system
and system organization .the data into clusters or categories. A
set of training data is presented to the system at the input layer

__________________________________________________________________________________________
level, the network connection weights are then adjusted
through some sort of competition among the node of the
output layer where the successful candidate will be the node
with the highest value.
Fig-6: Unsupervised learning
2.3.3 Reinforcement Learning:
The reinforcement learning algorithm also called as graded
learning it has mimic in a way the adjusting behaviour of
humans which interacting with a given physical environment
.the network connections are modified according to feedback
information provided to the network by its environment. This
information simply instructs the systems on whether or not a
correct response has been obtained. In case of correct response
the corresponding connections leading to that output are
strengthened otherwise they are weakened. Reinforcement
learning doesn’t get information on what the output should be
when the network is presented with a given input pattern.
Fig-7: Reinforcement learning
3. MAJOR CLASSES OF NEURAL NETWORKS
There are 4 types of neural network classes
 Multilayer Perception
 Radial Basis Function Network
 The Kohonen Self Organizing Map
 Hop Field Network.
3.1 The Multilayer Perception:
Topology: The multilayer perceptron belong to the class of
feed-forward networks, means that the information flows
among the network nodes exclusively in the forward direction.
the number of hidden layers required with in multi-layer
perceptron depends in major part on the type of problem being
addressed .for instances, a system with a single hidden layer
able to solve the problem of XOR function or related problems
in which the separate boundaries are relatively simple.
Activation function: in multilayer perception network with
one single hidden layer composed of an appropriate number of
nodes with sigmoid activation function ,as the activation
function for all the neurons of the network defined
as:Ec=1/2 1𝑛
𝑖=0 𝑡𝑖 𝑘 − 𝑜𝑖 𝑘 2
(𝑞
𝑖=1 .
Learning algorithm: The algorithm is based on the gradient
descent technique for solving an optimization problem which
involves the minimization of the network cumulative error Ec ,
Ec represents the sum of n squared errors
E(K)=I/2 𝑡𝑖 𝑘 − 𝑜𝑖 𝑘 2
𝑞
𝑖=1 , Where the index represents
the i-th neuron of the output composed of a total number of q
neurons. The algorithm is designed in such a way as to update
the weights in the direction of the gradient descent of the
cumulative error. Applications: signal processing, weather
forecasting, financial market prediction, pattern recognition,
signal compression.
3.2 Radial Basis Function Networks:
Topology: radial basis function network represent a special
category of the feed forward neural network architecture. The
basic RBFN structure consists of an input layer, hidden layer
with activation function and output layer. The network
structure use non liner transformations at its hidden layer but
uses linear transformations between the hidden and the output
layers.in general form of RBF function is given by gi(x)=
𝑟𝑖 | 𝑥 − 𝑣𝑖 | 𝜎𝑖 ,where x is the input vector and vi is the
vector denoting the center of the receptive field unit gi with 𝜎i
as its unit width parameter.
Activation function: the logistic function of RBF is given by
gi(x)=1(1+exp( | 𝑥 − 𝑣𝑖 |2 𝜎𝑖2 .
Learning Algorithm: In RBF network is a two stage learning
strategy:

__________________________________________________________________________________________
Step1: train the RBFN layer to get the adaption of centres and
scaling parameters using the unsupervised training.
Step2: adapt the weights of the output layer using the
supervised training algorithm.
Applications: control system,audio and video signal
processing and pattern recognition, weatherand power load
forecasting.
3.3 Kohonen’s Self Organizing Network:
Topology: The kohonen self-organizing network (KSON) also
called self-organizing map(SOM) belongs to the class of
unsupervised learning networks in KSON , the node
distributed themselves across the input space to recognize
groups of similar input vector while the output nodes compute
among themselves to de fired one at a time in response to a
particular input vector this processes is called competitive
learning. The nodes of the KSON can be recognize groups of
similar input vector this generate a topographic mapping of the
input vector to output vector which depends primarily on the
pattern of the input vector and results in dimensionality
reduction of the input space.
Learning algorithm: The learning here permits the clustering
of input data into a smaller set of elements having similar
characteristics. It is based on the competitive learning
technique also known as the winner take all strategy.
Application: speech recognition, vector coding, texture
segmentation, designing nonlinear controllers.
3.4 Hopfield Network
Topology: The hope field is a recurrent topology and working
of Hopfield is based on the associative memory concept that
means the network is able to recognize newly presented
pattern using an already stored complete version of that
pattern. Hopfield defines as any physical system whose
dynamics in phase space is dominated by a substantial number
of locally stable states to which it is attached can therefore be
regarded as a general content addressable memory.
The activation function of each neuron in Hopfield is defined
by the equation:
oi=sign( 𝑤𝑖𝑗𝑜𝑖 − 𝜃𝑖)𝑛
𝑖=1
𝑖!=𝑗
,
Where𝜃𝑖𝑠𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑𝑣𝑎𝑙𝑢𝑒,oi is output of the current
processing unit.
Learning algorithm: The learning algorithm of Hopfield is
based upon the Hebbain learning rule which is based on
supervised learning algorithm. The Hebbian learning rule
applied to a set of q presented patterns
Pk(k=1,2,3,4…….q)each with dimension n, where n is the
numbers of neurons unit in the Hopfield network.
Applications: It has been used to solve optimization problem,
original optimization problem, combination tutorial
optimization problem.
4. APPLICATION OF NEURAL NETWORKS
 Neural network for process monitoring and optimal
control.
 Neural network is a semiconductor manufacturing
process.
 Neuralnetwork is a power systems.
 Neuralnetwork in robotics.
 Network in communications.
 Neuralnetwork in pattern recognition.
5. CONCLUSIONS
In this paper mainly we description of neural network main
features in terms of topology, learning algorithm and
activation functions. These include the feed forward and the
recurrent based topologies, they also include network with
supervised and unsupervised learning algorithm. We provide
detailed descriptions of a numbers of the very often used
neural network models the Multilayer Perceptron, Radial Basis
Function network, Kohonen’s Self-Organizing network, the
Hopfield network along with highlights on their fields of
applications.
REFERENCES
[1]. Hop, good, A. (1993) knowledge –based systems for
Engineer and scientists,Boca Raton, Florida, CRC Press , 159-
85.
[2]. Jang, J, Sun, S, and Mizutani, Etice (1997) Neuro-Fuzzy
and Soft Computing, prentice Hall, Upper Saddle Rver, NJ.uin
[3].Haynkin,S. (1994) Neural Networks, A Comprehensive
Foundation, MacMillan Publishing,EnglewoodCliffs,NJ.
[4]. Werbos,P.(1947) “Beyond Regression: new tools for
predicgions and analysis in the behivoral sciences .” PhD
dissertation, Harvard University .
[5]. Rumelhary, D. Hinton, G., and Williams, R, (1986)
“Learning represantations by backpropagation errors,” Nature,
vol.323,pp. 553-6.
[6]. Haykin, S.(1994) Neural Networks, a Comprehncive
Foundation, MacMillan Publishing,EnglewoodCliffs,NJ.
[7]. Shar, S.andPalmieri, F.(1990) “MEKA. A fast local
algorithm for training feed-forward neural networks,”in
proceedings of the International Joints Conference on Neural
Networks.
[8]. Kattyama, R., Kuwata, K., Kajitani,
Y,.andWatanabe,M.(1995) “Embeding Functions networks,”
Fuzzy Sets and Systems, 72(3):311-27.
[9]. Hopfield, J. (1984) “Neurons with graded response have
collective computational properties like those of two state
neurons ,” in proceedings of the national properties like those
of two state national Acccadamy of science, pp.388-92.

__________________________________________________________________________________________
BIOGRAPHIES
1 ShrutiB.Hiregoudar, Mtech(cse), Basaveshware Engineering
College, Bagalkot. and shrutishining411@ gmail.com
2 Manjunath K, Mtech(cse), Basaveshware Engineering
College, Bagalkot. andmanjuise.026@gmail.com.
3 Prof: K.S.Patil,Mtech(cse), Basaveshware Engineering
College, Bagalkot.and kamalashashibec@rediffmail.com

A survey research summary on neural networks

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to A survey research summary on neural networks (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

A survey research summary on neural networks