As mentioned in the previous section, the output of a neural network is generally a set of unitless numbers on a scale between 0 to 1 or −1 to 1. Therefore, for applications that require data ranges outside of the neuron output, the data must be rescaled to the desired data range.
Classifiers usually have a separate output for each class. In this case, the outputs need to be thresholded so that a value above the threshold indicates that a given input is classified in that class and a value below the threshold indicates that input is not a member of that class. This thresholding is often accomplished by using a step function like those shown in Fig. 6.1, which results in a binary output as shown on the left side of Fig. 6.1. The value used to threshold the output can be adjusted to produce the optimum ratio of detection to false alarms. Sometimes, it is useful to have an upper and lower threshold for a given classifier design, permitting the classifier to have a “not sure” or indeterminate region. If an output falls above the upper threshold, it is marked as part of the class. If it falls below the lower threshold, it is marked as not part of the class. If it falls between the two thresholds, then the class should be considered indeterminate. This results in two binary outputs: one indicating class membership and one indicating no class membership. If both outputs are low, then class membership is possible but not definite, indicating an indeterminate condition, as shown on the right side of Fig. 6.1.
Online access to SPIE eBooks is limited to subscribing institutions.