Delta terminator block
The delta terminator is an error criterion component that compares two signals. It's called a terminator because it terminates the forward information flow in a topology.
It is a two in-port block meaning that it takes two independent signals as input.
Although the delta terminator has an output, it is seldom use. Usually the its operation lies in its propagator being the first block in the backwards propagation pass of a system.
The delta terminator is:
where is the error metric, is the signal from the first input port and from the second.
The settings can be modified using the settings browser.
|Delta terminator settings|
When the component interaction tool is selected the delta terminator toolbar can be used. The following elements are available:
- Visualization selection: Selection of visualization GUI mode (see below for details)
- Legend: Controls if a legend for the variables will be shown (off by default).
- Save image: Saves the plot to an image file.
- Print: Prints the plot (except confusion plot).
- Help: Does nothing for now.
Error vs epoch
This mode plots the error for each epoch. It is used to see how the adaptation of a system progresses. While the error decreases, the system is getting better. Two graphs are plotted for each feature - one for training data and the other for validation data.
If the the graph for the validation data and the training data start diverging significantly it means that the system is over-trained. That means that the system has started learning the training set by heart instead of learning how to generalize.
There is no magic bullet against overtraining. Some algorithms are more suceptible to it than others. Once the curves start diverging your best bet is to stop the training.
Using too large systems (with too many free prameters) with too little data will cause overtraining.
Output vs target
This mode plots the two input ports of the delta terminator. It is usually used to see how the desired output differs from the system output for an adaptive system. A perfectly adapted system will result in a straight line.
Training and validation sets are shown in different colors.
The output vs target visualization is useful for at a quick glance seeing how well a system has adapted to data. It is less useful quantitatively.
This mode is used only in classification where each output feature is a binary class flag. It allows one to see the system performance in terms of misclassification. The rows represent the output variables of the system and the columns present what they have been classified as. In the example image for instance 96.6% of C=Fail have been classified as C=Fail while 3.4% have been classified as C=Pass.
The delta terminator can show a confusion matrix for any number of variables. In the special case of two variables it can be interpreted like this:
|Predicted Negative||Predicted Positive|
|Negative Cases||True Negatives||False Positives|
|Positive Cases||False Negatives||True Positives|
- The two ports on the delta terminator are linked meaning that when you set the number of features on one, the number of features on the other will be set to the same number.
- The delta terminator is not an adaptable component and does not have any update rules compatible with it.
- Multiple delta terminators can be used in a system for multi-objective optimization.