Abstract: This "classic" style of neural network is used for spontaneously generating an algorithm for the analysis of numeric patterns. The goal is to "teach" the network how to analyze a desired set of numeric inputs. For instance, the network might be taught to calculate the AND, OR, and XOR of two input numbers: given the inputs of 0 and 1, the network should output 0, 1, and 1. The trick is that the network is not taught how to analyze the numbers; the network is given several sets of inputs and the correct output for each input set, and attempts to synthesize an algorithm to provide correct outputs. If this process is correctly performed, the network may be able to yield the correct analysis of input sets for which it has not been taught the correct output.

A backpropagation network usually consists of three "layers" of neurons: a layer of "input" neurons, each corresponding to a numeric input; a layer of "output" neurons, each corresponding to an output value; and a "hidden" layer of neurons that adjusts in order to synthesize an algorithm. The user first configures the network by designating the number of neurons in each layer, and then provides a set of training patterns, each consisting of input values and correct output values. The user then trains the network to recognize these input values and the correct outputs. The application features several training models, each of which conducts a different training process for the neural network. Once the network is fully trained, the user may provide new input values, and can watch the network analyze the values and produce output. Both the network and the training patterns may be saved to disk or later reloaded.

Version: 1.0 - last updated Sunday, June 05, 2005

Links: Download (Contents: Executable (Windows), Source Code (C#))

Screenshots: