Abstract:

In classification methods that explicitly model class-conditional probability distributions, the true distributions are often not known. These are estimated from the data available, to approximate the true distributions. Errors in classificaiton that arise due to this approximation can be reduced to some extent if the estimated distributions are used merely to project data into a space of likelihoods and classification is performed in that space suing discriminant functions. In this article, we discuss the rationale behind this, and also the general properties of likelihood projections. We demonstrate the utility of likelihood projections in improving classification performance through experiments carried out on a standard image database and a standard speech database.