Learning internal representations by error propagation
Parallel distributed processing: explorations in the microstructure of cognition, vol. 1
Artificial Intelligence
Combining the results of several neural network classifiers
Neural Networks
| Hi-index | 0.00 |
We explore the use of multiple artificial neural networks combined within the framework of the Dempster-Shafer Theory of Evidence to construct a hybrid information processing system for the correlation of tephra layers. The working hypothesis is that the system can correctly correlate tephra layers from one site to another even when data are sparse. The collection and analysis of data appropriate for utilization in standard statistical techniques aiding correlation is costly and time consuming. Given this state of affairs, here we employ a hybrid pattern recognition approach, which allows us to produce a recognition result with a relatively small amount of data. We used the major tephra-fall layers within one eruption sequence, the North Mono eruption, Mono Craters, CA, USA, to determine whether the system can be trained to distinguish layers on a bed-by-bed basis. The beds are distinguished in the field by the fraction of pumice, grading, zoning, thickness, and size of large pumice or lithic fragments. These same features were used to train the hybrid system. In the best case, the hybrid system was able to categorize observations correctly 93% of the time, which was markedly better than using neural networks alone. The average result for all pairwise comparisons of beds by the hybrid system is 76%, with the results for two beds that were not distinct removed. We conclude that it is possible to train the system to discriminate reasonably successfully among tephra layers with limited data. Because the system was not designed with any reliance on features specific to tephra layers, it may be possible to apply the system to the categorization of any stratigraphic units.