Você está na página 1de 3

“Necessity is the mother of Invention”

Man created machines to reduce his laborious and


monotonous work and perform it with unparalleled
accuracy. However, until recently machines
couldn‟t do work which required thinking and
analytical skills.
This scenario has begun on the path of change.
History shows that the greatest technologies have
come as a result of inspiration from biology; ANN
(artificial neural network) is one such technology.
A Biological Neuron
Let‟s dive into the world of neural networks.
Example:
Man has always believed that, the human brain is Y = F (5A + 2B)
the greatest machine ever. It consists of a dense In the above equation, A has more „importance‟ in
network of tiny processing elements called deciding the output Y compared to B.
neurons; this network is the force behind the
amazing processing power that the brain has. Let‟s implicate the neuron mathematically,
Wouldn‟t it be wonderful, if machines also have
such power? Synapses will get modeled as multiplying factors
If we are able to implicate a neural network or weights; dendrites as input terminals; soma as
artificially, then, it would also be able to process the processing block; axon as the output terminal.
and analyze like the human brain. To create such a
network, we need to have artificial neurons. For
that, let‟s first have a look at the biological neuron.

The biological neuron has dendrites along which it


receives the inputs. These signals are then fed to
the soma (or nucleus) which processes them. In
effect, it sums the signals and transforms the
resultant by a function. The output is then sent
down the axon. The output signal is then given to
other neurons through links known as synapses.
The strength of the synapses decides how much
the signal gets weakened before reaching the Mathematical Model
next neuron; thus determining, the degree of
„importance‟ this signal has, in deciding the Now, a system can be represented in a block
output of the next neuron. diagram form like this:
Now let us add series connections from left to right
and parallel connections from top to bottom.

An n-input m-output system

As the diagram shows, every desirable system


would have a relationship or formula between its
input and output.
If we connect multiple neurons with each other,
they would be able to establish such a relationship.
What if, the formula for a particular system
consists of functions other than the function
provided in the neurons
i.e. G ( ) is not same as F ( )?
Say the function provided in the neuron is addition; A 3 layer neural network
now, multiplication can be performed using
addition, subtraction can be performed using Hey presto! We have a 3 layer neural network!
addition; integration is finally addition, and so The output of every neuron is connected to every
on… neuron in the corresponding next layer. This will
Thus, by having appropriate multiplying factors or provide the possibility of realizing every probable
weights, we can achieve any kind of a relationship input output relationship, so long as there are
between the input and the output. However, care enough layers of neurons in the hidden layer.
should be taken to choose the function which But Wait! Before we can implement any system
would suit the desired application i.e. F ( ) should using this network, we need to have the
suit G ( ), so as to avoid the unnecessary usage of appropriate weights for that system.
more neurons.
Going back to Mother Nature for inspiration;
Neurons are interconnected in layers. Let‟s create Human beings learn by experience. Similarly, we
one such network. can have a process of training the ANNs, to
Let‟s have an input layer for buffering the inputs; achieve the desired weights.
core (or hidden) and output layers, to give it the This would be similar as solving a puzzle of
ability to establish the complex relationships. „Complete the Series‟

Here, in the first case, the input is 2 and the output


is 4. On observing this, we try to find out the
relationship between input and the output. Now, 2 2
is 4, 2+2 is 4, 2 x 2 is 4, and so on… Then we
move on to the next set of data, 4 and 6. After
looking at such sets of data, it finally becomes
3 layers of Neurons clear that +2 is the relationship which satisfies all
the examples. From this relationship it is Example: In the case of launching a rocket,
established that the output is 10. various laws of physics are involved. These laws
are expressed mathematically, as a result of which
Similarly, while training, ANNs should be the relationship becomes clearly definable. Such
provided with sets of inputs and outputs known as relationships are easily realizable using
“training examples”. Then, an iterative algorithm conventional systems.
be adopted to modify the weights, to obtain values Conversely, in cases like, face detection in an
of weights which satisfy all examples. Thus image, obviously no such mathematical laws are
establishing a generalized relationship which applicable, and thus the relationship can‟t be
would produce the correct output in all possible clearly defined. But, ANNs are able to realize such
conditions which could occur during operation. relationships, due to the method of design used
Because we use practical examples for training, the (involving training).
relationship developed is free from the In conventional systems, we have serial
assumptions and approximations taken to computation i.e.; only 1 thing is processed at a
formulate the mathematical model of a system, as time, but at a high speed. So, even if a single
in conventional systems. Thus the relationship activity gets disturbed the entire output gets
developed is very accurate. degraded.
However, to achieve this we need to cover all types In ANN based systems, we have highly parallel
of data during the training. Thus, training becomes computation i.e.; at a time different processing is
a very complicated and time consuming process. being done by a large number of neurons. So, even
However, training is only a one time issue. if degradation happens, it is graceful, because each
But, how do we realize ANNs? neuron has only a small part to play.
We saw the mathematical model of ANNs. The Like, conventional systems require programming,
mathematical functions can be realized using op- Neural Network based systems require training
amps, and thus the ANNs. We may also map the before they can be put to work.
entire circuitry on a single chip using VLSI
technology. Areas of Application
However, we do not have any way to predict ANNs are used in systems where thinking and
beforehand, what kind of a neural network, how analytical skills are required; from a broader
many layers, what kind of training, etc. will be best perspective, they may be used in any system where
suitable for the desired application. the relationship is implicit.
So, for the purpose of development, testing and Due to the accuracy of ANNs in deciphering the
simulation, tools like MATLAB and LISP can be relationships, they are also quite commonly used to
used to realize ANNs superficially. Though, in devise the formula while designing systems.
some cases these may be used as the mode of
realization for operation too. Conclusion
Artificial Neural Networks are adaptive in nature
Although we‟ve managed to invent ANNs here, it and are hence more versatile.
is not difficult to guess that the ANN technology is With their versatility and the ability to provide
a pre-existing one. Let‟s now take a comparison of control customized for specific application, ANNs
ANNs with conventional systems. are set to break all the barriers in automation; those
which exist due to the inabilities of present day
ANNs v/s Conventional systems conventional systems, to process like the human
ANNs are able to realize systems having implicit brain.
input output relationships, which conventional
systems can‟t.
- Prashant Serai Shuchi Jain
(II Year EXTC)

Você também pode gostar