Exploiting Geometry in Counting Grids

In this paper we exploit the use of known information about the geometry structure of a recently proposed generative model, namely Counting Grid (CG) [1] to improve the performance of classification accuracy. Once the generative model is trained, the geometric structure of the model introduces a natural spatial relations among the estimated latent variables. Such relation is generally ignored when standard maximum likelihood approach (or classical hybrid generative-discriminative approach) is employed for classification purpose. In this work, we propose to take into account the geometric relations of the generative model by proposing an ad hoc similarity measure for CG. In particular, the values relative to each point of the grid is spread around its neighborhood by using information coming from the CG training phase. The proposed approach is succesfully applied in two applicative scenarios: expression microarray classification and MRI brain classification. Experiments show a drastic improvement over standard schemes when our approach is employed.