Neural
Networks:
|
Neural
Networks: |
|
Intelligence in a control system is the ability for a control system to adapt to unknown circumstances and then execute the best logical solution statistically based upon the information given in order to accomplish an ultimate goal. The goal then of intelligent systems from an engineering standpoint is to create a system that will rival that of natural intelligence in the performing of meaningful task in the natural world. The meaningful task of a specific system is assigned through human intervention by means of a designer, programmer or operator of the system. This means that the intelligent systems do not learn as human being might learn, but instead learn by way of pattern recognition of information given to the system allowing for the system to respond quickly and successfully to new situations. Neural networks are one form of an intelligent system and the following few paragraphs will give a general overview to the origin, general principle and applications of neural networks.
|
|
THE ORGIN AND FUNDEMENTALS OF NEURAL NETWORKS Although there is a variety of ways that neural networks function the name neural network is derived from that of the nervous system in biological creatures. This is why this discussion will begin with the neuron, the basic building block of the human brain. A neuron in a biological system is made up of several major parts: the cell body that is known as the soma (this is where the nucleus is located), the dendrites, the axon and finally the Synaptic terminals. Each part has a specific task assigned to it that is essential in the overall operation of the nervous system. The dendrites act as a web of nerve fiber connected to the soma that receives the input signal. The cell body, or soma, is the “brains” of the neuron and processes the information received by the dendrites and sends the output signal to the axon. The axon is a single long connection that carries the output signal to the synaptic terminals to be sent out to other neurons through chemicals known as neurotransmitters. |
Image taken from http://ieee-nns.org/
Figure 1-1 shows a simple mathematical model of the biological neuron mentioned in the previous paragraph. McCulloch and Pitts proposed this model in 1943 and is normally referred to as the M-P neuron. In this M-P neuron the weight Wij represents the strength of the synapse that directly links the neuron source j to neuron i destination. Weights are assigned to each input and are summed (minus any offsets) by the processing unit to achieve a desired output signal. Although this representation is crude with respect to an actual biological neuron the mathematical model here does possesses the ability to perform many task and is referred to as the processing unit. The processing unit gives the artificial neuron the ability to perform basic logic operations NOT, OR and AND. By connecting a large number of these M-P neurons together, just as neurons are connected in biological creatures, the processing elements can assign or adjust the weights in a way that allows for the network to learn, recall or generalize sets of data.
Figure 1-1
M-P neuron
NETWORK GEOMETRIES
As mentioned in the previous paragraph a neural network is made up of these artificial neurons that we will now refer to as nodes. As signals pass through the network between each node the networks weighted connections places conditions on the signal before sending it on to the next node. The actual arrangement of the nodes creating the network can take on numerous geometries such as, but not limited to:
feed-forward network: output node never forms part of its own network
feed-back or recurrent network: the output of at least one node is propagated back and a modified version of this signal is used as an input to the same node.
autoassociative network: these networks attempt to reconstruct the true version of the corrupt input signal that is presented to the network.
heteroassociative network: attempts to map the input to an alternative representation.
