Argentum Solutions, Inc.
Sterling guidance on corrosion and materials degradation
| Potential-pH Diagrams |
 |
|
| Intelligent Tools |
 |
|
 |
|
 |
|
| Corrosion Calculator |
 |
|
| Corrosion Economics Estimator |
 |
|
|
David C. Silverman |
|
Artificial Neural Network Background Artificial neural networks can be viewed from two distinctly different standpoints, a computational standpoint and a biological standpoint. Computationally, artificial neural networks are a framework for representing functions through fairly simple computing elements. Their strength lies in their ability to represent complex functions which have large numbers of "noisy" inputs and in which the relationship between the observable inputs and outputs cannot be written as an equation. The relationship instead appears to the viewer as a pattern. One of the attractions of the artificial neural network is that these networks can be trained using an inductive type of learning algorithm. After initialization, the network can be modified to improve performance as new input/output pairs are presented. The type of network most often encountered in practice is the back-propagation network, so-named because the network learns a predefined set of input-output example pairs by a two phase propagate-adapt cycle. Other network structures exist and learning algorithms exist for them. Their use is sometimes more appropriate. A few are explained in more detail in this tutorial.Biologically, the back-propagation neural network and its related cousins (e.g. perceptron and other multilayer feed forward networks) provide a somewhat simple model of the human brain with the network loosely corresponding to the neuron structure of the brain (hence the name "neural network"). The similarity lies in the structure of the node containing a summation of inputs, an activation function determining if that node "turns on", and an output, the magnitude of which is determined by the activation function. The difference lies in all neurons and synapses in the brain being active simultaneously (massively parallel) whereas the computer has only one or several CPU’s. This difference means that the computer may require hundreds of cycles to determine if a single network element should fire while the brain can make this determination in a single step. Other differences exist so caution should be exercised when an attempt is made to equate an artificial neural network to the human brain. An alternative to the above neural networks are those networks that use a probabilistic representation of uncertain knowledge. These networks are in a class called belief networks. They have many similarities to back-propagation neural networks in their ability to learn by local, gradient descent methods. Bayesian (or probabilistic) learning is an example. In a general sense, the idea is to use hypotheses between data and predictions. The probability of the hypothesis is estimated. Predictions are made from the hypotheses as weighted by their probabilities. The probabilistic and general regression networks described in this tutorial are in this category. |
|
Previous Page: Overview of Tutorial |
|
David C. Silverman, Ph.D. - Primary Consultant |