The output neurons of most neural-network architectures produce data in the range of -1 to 1 or 0 to 1, depending on the type of neuron. Therefore, the desired outputs must be coded to fit this scale. Also, non-numeric labels, such as those found with classifiers, must be converted to something numeric. This necessitates coding the target outputs. Once the neural network is trained and in operation, the reverse applies and these unitless values must be converted into useful terms. This is done through post-processing of the neural network outputs.
|