Idea: Trying to get the maximum from datasets where we have very few training examples, but each example has a very large number of features. Examples of such datasets include medical databases where we have gene activation measurements for very few patients but many different genes.
Method: We design a neural network architecture whose number of parameters is constant with respect to the number of features (which is not the case with a typical linear classifier). The basic idea is that we use a linear classifier whose coefficients for each features are generated by a single MLP that takes as an input a representation for this feature, which is basically a transformation of the set of values taken by this feature through all the examples. More complex (deep) architectures are also experimented.
- ICML 2003 feature selection challenge datasets: Arcene, Dorothea
- AML/ALL Leukemia classification dataset
See doc/README.pdf for detailed explanations.