4.5. New from PickSets
The Neural Network plugin for OpendTect supports two types of neural networks based on picks: fully connected Multi-Layer-Perceptions (MLPs) and Unsupervised Vector Quantizers (UVQs). MLPs are used in supervised (learn by example) mode while UVQs are used in unsupervised experiments (segmentation = clustering).
Analysis method. The Supervised method allows the choice of one or more output nodes. The groups of nodes (or PickSets) indicate how the neural network should separate the character found in the input attributes. Unsupervised separates the nodes in the (single!) Pickset based on a clustering in a user defined number of classes (see below).
Input Training data set. Specify whether the Input data set must be extracted on the fly, or retrieve a stored input set. An input training set consists of a range of attributes (names and values) at given example locations. To create a training set you must specify which attributes to use and at which locations these attributes must be calculated.
Select input/output attributes. The Select input attributes list on the left-hand side lists all attributes defined in the current attribute set as well as all data that is stored on disk. Select any or all of these to serve as input to the neural network. The Select output nodes on the right contains all available Pickset groups. Select the Pickset Group containing the locations at which attributes must be extracted to create a training set. Note that for an object probability cube such as TheChimneyCube® you need two Picksets: chimneys and non-chimneys.
Percentage used for test set. It is recommended to create a subset for testing the neural network's performance during training, specify a Percentage. The test set is created by randomly drawing example locations from the selected Picksets. Test set examples will not be used to update the neural network weight set during training. They are merely passed through the network to compare the network's classification with the actual classification.
Number of Classes. In unsupervised mode, attributes at locations in the specified Pickset are clustered (segmented) into the specified Number of classes. During the training phase the UVQ network learns to find the cluster centers. In the application phase the input attributes are compared to each cluster center. The input is assigned to the winning segment, which is a number from 1 to N, where N is the number of clusters. In addition, the network calculates how close the input is to the cluster center of the winning cluster. This measure of confidence is called a match, which can range between 1 (perfect match, i.e. input and cluster center are the same) and 0 (input and cluster center are completely different).
