WHAT IS A NEURAL NETWORK ? (short description)

Biological neurons


Our brain is made of neurons. Each neuron is a cell connected to others
through dendrites and axon. A cell receives information through dendrites
and sends information through axon.
When a cell decides to send information depends on its internal state.
The interconnection of these neurons gives us the means to think, to recognize...

Biological neuron

Artificial neurons and artificial neural networks


The simple model of biological neuron describes above can be modeled with
mathematical functions and implemented on computers. The result is an artificial
neural network.
The goal is to simulate some cerebral functions, such as classification or
pattern recognition.
A classical artificial neural network is made of input neurons that receive
input information (a picture for example) and output neurons that give a
result (may be a particular shape in this picture).
Each neuron in input can be fully or in part, connected to each output neuron.

We call input layer the set of input neurons and output layer the set of output neurons.

There are several different architectures used in artificial neural networks
and sometimes we find one - or more - hidden layers, a set of neurons between
the input and output layers.

The most popular architecture is called backpropagation neural network.

How does it work ?


An artificial neuron also called formal neuron is a threshold robot : It sums all the input values
coming from the links with the others neurons. Each of these links has coefficients. The neuron
gives a value (greater than 0) if the result of summing is over a threshold value.

Formal neuron



The problem is to find methods that adjust the coefficients (or weights) for an appropriate
result.
The learning (or training) phase is used to adjust the weights. During the learning phase we give a pattern
(or scheme) to input neurons. If the result achieved is different than the input, we modify the
weights until we have the desired output. This operation is repeated for all patterns.
The difference between the input and the desired output is called error. The goal in training is to minimize
the error.

We can train - for example - a neural network to recognize shapes (such as handwritten) or to detect viruses
in your computer.

Example of multi-layers neural network


After the learning phase, a neural network can recognize a learned pattern or classified a new
pattern.
The choice of patterns for the training phase is very important. It determines the abilities
of the neural network.

[MAIN PAGE] [WHAT IS NEURAL NETWORK ?] [WHAT IS BRAININABOX ?]