Neural Networks

Artificial neural network is an inspirational technologypower, becoming a simple matrix multiplication
of biological neural network. In general a biologicaloperation from linear algebra.
neural network is composed of a group or groups ofAlternative calculation models in neural networks
chemically connected or functionally associatedinclude models with loops, where some kind of time
neurons. A single neuron may be connected to manydelay process must be used, and "winner takes all"
other neurons and the total number of neurons andmodels, where the neuron with the highest value
connections in a network may be extensive.from the calculation fires and takes a value 1, and all
Connections, called synapses, are usually formedother neurons take the value 0.
from axons to dendrites, though other kind ofTypically the weights in a neural network are initially
connections, also possible.set to small random values. This represents the
An Artificial Neural Network is also called as simulatednetwork knowing nothing; its output is essentially a
neural network of many very simple processorsrandom function of its input. As the training process
("units"), each possibly having a (small amount of)proceeds, the connection weights are gradually
local memory. The units are connected bymodified according to computational rules specific to
unidirectional communication channels ("connections"),the learning algorithm being used. Ideally the weights
which carry numeric (as opposed to symbolic) data.eventually converge to values allowing them to
The units operate only on their local data and on theperform a useful computation. Thus it can be said
inputs they receive via the connections. Artificialthat the neural network commences knowing nothing
neural networks are made up of interconnectingand moves on to gain some real knowledge, though
artificial neurons (programming constructs that mimicthe knowledge is sub-symbolic.
the properties of biological neurons). Artificial neuralARCHITECTURE OF NEURAL NETWORKS
networks may either be used to gain anFEED-FORWARD NETWORKS:-
understanding of biological neural networks, or for Feed-forward ANNs allow signals to travel one way
solving artificial intelligence problems withoutonly; from input to output. There is no feedback
necessarily creating a model of a real biological(loops) i.e. the output of any layer does not affect
system. The real, biological nervous system is highlythat same layer. Feed-forward ANNs tend to be
complex and includes some features that may seemstraight forward networks that associate inputs with
superfluous based on an understanding of artificialoutputs. They are extensively used in pattern
networks.recognition. This type of organization is also referred
HISTORY OF NEURAL NETWORKS:-to as bottom-up or top-down. ex.,  Fault Diagnosis
The concept neural networks started in thesystems.
late-1800s as an effort to describe how the humanFEEDBACK NETWORKS:-
mind performed. These ideas started being applied toFeedback networks can have signals traveling in both
computational models with Turing's B-type machinesdirections by introducing loops in the network.
and the perceptron. In 1940  Donald Hebb madeFeedback networks are very powerful and can get
one of the first hypotheses for a mechanism ofextremely complicated. Feedback networks are
neural plasticity  (i.e. learning), Hebbian learning.dynamic; their 'state' is changing continuously until
Hebbian learning is considered to be athey reach an equilibrium point. They remain at the
'typical' unsupervised learning rule. In 1943 McCullochequilibrium point until the input changes and a new
and Pitts proposed the McCulloch-Pitts neuron model.equilibrium needs to be found. Feedback architectures
In 1948 Alan Tuning, first proposed the concept ofare also referred to as interactive or recurrent,
Neural Networks in his paper "Intelligent Machinery". although the latter term is often used to denote
In 1949 Hebb published his book The Organization offeedback connections in single-layer organizations. Ex.,
Behavior, in which the Hebbian  learning rule wasRadiosity for Virtual Reality Systems (ROVER)
proposed. In 1958 Rosenblatt introduced the simpleTRAINING OF ARTIFCIAL NEURAL NETWORKS:-
single layer networks now called Perceptrons. In 1969A neural network has to be configured such that the
Minsky and Papert’s book Perceptronsapplication of a set of inputs produces (either 'direct'
demonstrated the limitation of single. In layeror via a relaxation process) the desired set of
perceptrons, and almost the whole field went intooutputs. Various methods to set the strengths of the
hibernation. In 1982 Hopfield published a series ofconnections exist. One way is to set the weights
papers on Hopfield networks. In 1982 Kohonenexplicitly, using a priori knowledge. Another way is to
developed the Self-Organizing Maps that now bear his'train' the neural network by feeding it teaching
name. In 1986 The Back-Propagation learningpatterns and letting it change its weights according to
algorithm for Multi-Layer Perceptrons wassome learning rule.
rediscovered and the whole field took off again. InWe can categorize the learning situations in three
1990s The sub-field of Radial Basis Function Networksdistinct sorts. They are:
was developed. In 2000s The power of Ensembles- Supervised learning or Associative learning in which
of Neural Networks and Support Vector Machinesthe network is trained by providing it with input and
becomes apparent.matching output patterns. These input-output pairs
NEURAL NETWORKS:-can be provided by an external teacher, or by the
 Neural networks take a different approach tosystem which contains the neural network
problem solving than that of conventional computers.(self-supervised).
Conventional computers use an algorithmic approach 
i.e. the computer follows a set of instructions in orderThis method has two sub varieties called
to solve a problem. Unless the specific steps that theauto-associative and hetero-associative. In
computer needs to follow are known the computerauto-associative learning, the target values are the
cannot solve the problem. That restricts the problemsame as the inputs, whereas in hetero-associative
solving capability of conventional computers tolearning, the targets are generally different from the
problems that we already understand and know howinputs.
to solve. But computers would be so much moreOne of the most commonly used supervised NN
useful if they could do things that we don't exactlymodel is back propagation network that uses back
know how to do.propagation learning algorithm. Back propagation (or
Neural networks process information in a similar wayback prop) algorithm is one of the well-known
the human brain does. The network is composed ofalgorithms in neural networks.
a large number of highly interconnected processing- Unsupervised learning method is not given any
elements(neurons) working in parallel to solve atarget value. Unsupervised learning or
specific problem. Neural networks learn by example.Self-organization in which an (output) unit is trained to
They cannot be programmed to perform a specificrespond to clusters of pattern within the input. In this
task. The examples must be selected carefullyparadigm the system is supposed to discover
otherwise useful time is wasted or even worse thestatistically salient features of the input population.
network might be functioning incorrectly. TheUnlike the supervised learning paradigm, there is no a
disadvantage is that because the network finds outpriori set of categories into which the patterns are to
how to solve the problem by itself, its operation canbe classified; rather the system must develop its
be unpredictable.own representation of the input stimuli. A desired
On the other hand, conventional computers use aoutput of the network is unknown. During training the
cognitive approach to problem solving; the way thenetwork performs some kind of data compression
problem is to solved must be known and stated insuch as dimensionality reduction or clustering. The
small unambiguous instructions. These instructions arenetwork learns the distribution of patterns and
then converted to a high level language program andmakes a classification of that pattern where, similar
then into machine code that the computer canpatterns are assigned to the same output cluster.
understand. These machines are totally predictable; if Kohonen network is the best example of
anything goes wrong is due to a software orunsupervised learning network. According to Sarle
hardware fault.(1997) Kohonen network refers to three types of
Neural networks and conventional algorithmicnetworks that are Vector Quantization,
computers are not in competition but complementSelf-Organizing Map and Learning Vector
each other. There are tasks are more suited to anQuantization. 
algorithmic approach like arithmetic operations and- The third major category of learning that will be
tasks that are more suited to neural networks. Evendiscussed is termed "reinforcement learning". As
more, a large number of tasks, require systems thatmentioned earlier, reinforcement learning lies between
use a combination of the two approaches (normally asupervised and unsupervised learning. It is often called
conventional computer is used to supervise the neurallearning with a critic rather than learning with a
network) in order to perform at maximum efficiency.teacher as the feedback is evaluative (right or
Neural networks, with their remarkable ability towrong) rather than instructive (desired output).
derive meaning from complicated or imprecise data,Reinforcement learning in neural networks has its
can be used to extract patterns and detect trendsroots in behavioral psychology. Here the learning
that are too complex to be noticed by either humansmachine does some action on the environment and
or other computer techniques. A trained neuralgets a feedback response from the environment.
network can be thought of as an "expert" in theThe learning system grades its action good
category of information it has been given to analyze.(rewarding) or bad (punishable) based on the
This expert can then be used to provide projectionsenvironmental response and accordingly adjusts its
given new situations of interest and answer "what if"parameters. Generally, parameter adjustment is
questions.continued until an equilibrium state occurs, following
Other advantages include:which there will be no more changes in its
1. Adaptive learning: An ability to learn how to doparameters. The self organizing neural learning may
tasks based on the data given for training or initialbe categorized under this type of learning.
experience.APPLICATIONS OF NEURAL NETWORK:-
2. Self-Organization: An ANN can create its own- Signal processing: suppress line noise, with adaptive
organization or representation of the information itecho canceling, blind source separation
receives during learning time.- Control: e.g. backing up a truck: cab position, rear
3. Real Time Operation: ANN computations may beposition, and match with the dock get converted to
carried out in parallel, and special hardware devicessteering instructions. Manufacturing plants for
are being designed and manufactured which takecontrolling automated machines.
advantage of this capability.- Siemens successfully uses neural networks for
4. Fault Tolerance via Redundant Information Coding:process automation in basic industries, e.g., in rolling
Partial destruction of a network leads to themill control more than 100 neural networks do their
corresponding degradation of performance. However,job, 24 hours a day
some network capabilities may be retained even with- Robotics - navigation, vision recognition
major network damage.- Pattern recognition, i.e. recognizing handwritten
A SIMPLE NEURON:-characters, e.g. the current version of Apple's
 An artificial neuron is a device with many inputs andNewton uses a neural net
one output. The neuron has two modes of operation;- Medicine, i.e. storing medical records based on case
the training mode and the using mode. In the traininginformation
mode, the neuron can be trained to fire (or not), for- Speech production: reading text aloud (NETtalk)
particular input patterns. In the using mode, when a- Speech recognition
taught input pattern is detected at the input, its- Vision: face recognition , edge detection, visual
associated output becomes the current output. If thesearch engines
input pattern does not belong in the taught list of- Business, eg.. rules for mortgage decisions are
input patterns, the firing rule is used to determineextracted from past decisions made by experienced
whether to fire or not.evaluators, resulting in a network that has a high level
 STRUCTURE:-of agreement with human experts.
Artificial neural networks are quite different from the- Financial Applications: time series analysis, stock
brain in terms of structure. Like the brain, however, amarket prediction
neural net is a massively parallel collection of small- Data Compression: speech signal, image, e.g. faces
and simple processing units where the- Game Playing: backgammon, chess, ROBOCUP.
interconnections form a large part of the network'sRobocup: Robot World Cup:-
intelligence; however, in terms of scale, a brain isThe RoboCup Competition pits robots (real and
massively larger than a neural network, and the unitsvirtual) against each other in a simulated soccer
used in a neural network are typically far simpler thantournament. The aim of the RoboCup competition is
neurons. Nevertheless, certain functions that seemto foster an interdisciplinary approach to robotics and
exclusive to the brain such as learning, have beenagent-based AI by presenting a domain that requires
replicated on a simpler scale, with neural networks.large-scale cooperation and coordination in a dynamic,
MODELS:-noisy, complex environment. RoboCup has three
A typical feed-forward neural network is a set ofdifferent leagues to-date. The Small and Middle-Size
nodes. In computing, feed-forward normally refersLeagues involved physical robots; the Simulation
to a multi-layer perceptron network in which theLeague is for virtual, synthetic teams. This work
outputs from all neurons go to following but notfocus on building softbots for the Simulation League.
preceding layers, so there are no feedback loops.Machine Learning for Robocup involves:
Some of these are designated input nodes,1.The training of player in the process of making the
some output nodes, and in-between are hiddendecision of whether (a) to dribble the ball; (b) to pass
nodes. Each connection between neurons has ait on to another team-mate; (c) to shoot into the
numerical weight. When the network is in operation,net.
a value will be applied to each input node -- the 2. The training of the goalkeeper in process of
values being fed in by a human operator, fromintelligent guessing of how the ball is going to be kick
environmental sensors, or from some externalby the opponents. Complexities arise when one
program. Each node then passes its given value toopponent decides to pass the ball to another player
the connections leading out from it, and on eachinstead of attempting a score.
connection the value is multiplied by the weight3. Evolution of a co-operative and perhaps
associated with that connection. Each node in theunpredictable team.
next layer then receives a value which is the sum ofCONCLUSION:-
the values produced by the connections leading intoThe computing world has a lot to gain from neural
it, and in each node a simple computation isnetworks. Their ability to learn by example makes
performed on the value -- a sigmoid function isthem very flexible and powerful. Furthermore there is
typical. This process is then repeated, with the resultsno need to device an algorithm in order to perform a
being passed through subsequent layers of nodesspecific task; i.e. there is no need to understand the
until the output nodes are reached. Early modelsinternal mechanisms of that task. They are also very
(circa 1970) had a fixed number of layers. Morewell suited for real time systems because of their
recently, genetic algorithms are used to evolve thefast response and computational times which are due
neural structureto their parallel architecture.
CALCULATIONS:-Neural networks also contribute to other areas of
The sigmoid curve is often used as a transferresearch such as neurology and psychology. They are
function because it introduces non-linearity into theregularly used to model parts of living organisms and
network's calculations by "squashing" the neuron'sto investigate the internal mechanisms of the brain.
activation level into the range [0,1]. The sigmoidPerhaps the most exciting aspect of neural networks
function has the additional benefit of having anis the possibility that some day 'conscious' networks
extremely simple derivative function, as required formight be produced. There is a number of scientists
back-propagating errors through a feed-forwardarguing that consciousness is a 'mechanical' property
neural network. Other functions with similar featuresand that 'conscious' neural networks are a realistic
can be used, most commonly tanh which squashespossibility.
activations into the range of [-1,1] instead, orFinally, to state that even though neural networks
occasionally a piece-wise linear function that simplyhave a huge potential we will only get the best of
clips the activation rather than squashing it.them when they are integrated with computing, AI,
If no non-linearity is introduced by squashing orfuzzy logic and related subjects.
clipping, the network loses much of its computational