|
|
|
|
|
|
 |
Neural Networks ComponentsArtificial NetworksArtificial neural networks consist of groups of artificial neurones connected together. Some early researchers attempted to simply connect neurones in a random manner, without much success. These failures demonstrated how critical the connectivity of neurones within a neural network is to a networks ability to process data. The biological neural networks that reside in the human brain consist of billions of neurones each connected to thousands of their neighbours. The neurones are organised at both a very gross level into the cerebral cortex, cerebellum and brainstem, and also at a microscopic level into cortical columns. The connections between these neurones are determined during development and modified by experience. Unfortunately, today's artificial neural networks are unable to model the complexity of biological systems. Generally artificial neural networks are organised into different layers of neurones that are then connected to one another. Although there are artificial neural networks that contain only one layer, most applications utilise at least three layers - input, hidden, and output. The neurones within the input layer receive data from either the outside world using sensors or, more commonly, from an input file. The output layer sends information either directly to the outside world via a control system or to an output file. Between the input and output neurones can be many hidden layers of neurones. The inputs and outputs of these hidden neurones simply go to other neurones. In most networks, layers of neurones are connected using
a "feed-forward" structure that allows signals to travel from
input to output only. Neurones within the input layer pass their output
to the first hidden layer, neurones in this layer then pass their output
to the second hidden layer etc. until eventually the output layer is reached.
In some networks neurones within the same layer also connect together.
This is often the case in the output layer where neurones may inhibit
each other. This "lateral inhibition" leads to a sort of competition
between output neurones. These types of network are used extensively in
pattern recognition. Another network structure is "recurrent"
or "feedback" that allows signals to travel both directions
by introducing loops in the network. That is, neurones of one layer are
able to send their output to previous layers. Feedback networks are potentially
very powerful. However, the use of feedback in neural networks is limited
by the degree of complexity they introduce. In summary there are several
different types of network architecture including feed-forward and feedback
networks, and each has their own advantages and disadvantages. Index | Components | Artificial Networks © 2008 Marcus bros
|
|
|
|
