This was a group project I was part of for my Machine Learning class under Dr. Isbell. Each group developed their own semester long project and for the final submission, wrote a NIPS style paper and presented their work.

One of the difficulties of Machine Learning and classification in specific is the curse of dimensionality. This is basically the idea that the more features/dimensions you have to classify on the more computation is needed. To alleviate this a lot of time can be invested in trying to fine tune the features used on a particular learning algorithm.

The idea behind our project was to exploit the internal workings of Neural Networks to automatically perform feature selection and additionally reduce overall computation time.
We had some promising results that gave us better accuracies than a well performing statistical algorithm on our data sets. Our results were not a large improvement though, so this is one area further work would focus on. Even without a huge accuracy gain though, our method provides a mechanism for parallelization of Neural Network learning.

A copy of our final paper is available in pdf.