In this proposal, we introduce a novel model based on a current deep architecture, multi-layer kernel machine. Our model combines deep architectures, non-linear dimension reduction methods and information theory. A deep model usually consists of multiple layers. We feed each layer with new representations discovered by non-linear dimension reduction methods and discriminate the new representations based on their information gain. Best features are selected through the deep architecture pipeline and the final representation is the best representation that helps to increase classification accuracy. We will fine-tune the choice of non-linear dimension reduction and information measuring methods iteratively in the next six months.
Adviser: Randy C. Paffenroth, Ph.D.