The two forms of network analysis available within the NeuralWizard are cross validation and sensitivity analysis. Both of these methods are executed in concurrence with the training of the network. Every so often, the network weights are frozen, the testing data is fed through the network, and the results are reported.
The cross validation test set is used to determine the level of generalization produced by the training set. Often a network can be overtrained, such that it will begin to "memorize" the training set at the expense of the test set performance. The Simulation Control panel can be configured to automatically stop the training when the error in the test set begins to rise.
The second form of network analysis computes sensitivities of the network's outputs with respect to each of its inputs. This method generates a test set by dithering the data from each channel of the input file by a small amount. It then feeds the dithered data into the network and computes the percentage change in the output. The set of percentages reveal the input channels that are most significant and the effect that a change in a particular input has on a particular output. The sensitivity analysis feature can be used for "data mining", which will allow you to discard the insignificant portions of your data and focus the training on the most important data.
NeuroSolutions Home Page