International Joint Conference on Neural Networks (IJCNN'05), August 1-4, 2005, Montréal, Québec, Canada
A hybrid stochastic-deterministic approach for solving NDA problems on very high dimensional biological data is investigated. It is based on networks trained with a combination of simulated annealing and conjugate gradient within a broad scale, high throughput computing data mining environment. High quality networks from the point of view of both discrimination and generalization capabilities are discovered. The NDA mappings generated by these networks, together with unsupervised representations of the data, lead to a deeper understanding of complex high dimensional data like Leukemia and Alzheimer gene expression microarray experiments.
Proceedings of the International Joint Conference on Neural Networks (IJCNN'05).