Exploratory Analysis of Neural Networks for Facies Discrimination from Logs

Geoffrey C. Bohling
Kansas Geological Survey, University of Kansas, Lawrence, KS

Many investigators regard neural networks with some skepticism due to their perceived black box nature. However, with some rudimentary understanding of the internal mechanics, it is not that difficult to extract meaningful interpretations of the weights in a single hidden layer network. If the network is designed for classification, then the set of weights from each hidden layer node to the set of output nodes can be regarded as a particular contrast between the corresponding output classes. This means that the hidden layer nodes can be ranked according to the strength of their contributions to certain desirable or interesting contrasts between classes or groups of classes. The input to hidden layer weights for the highest ranking hidden layer nodes can then be regarded as the linear combinations of input variables that contribute most to those interesting contrasts. Scores on these combinations can then be computed and plotted, leading to a form of non-orthogonal factor analysis. I will examine the utility of these ideas in the context of a neural network trained to discriminate between a number of facies based on geophysical well logs, demonstrating that the same set of hidden layer nodes can be ranked according to different criteria (strengths of contrast between different subsets of facies) to yield multiple sets of interesting factors from the same network.