The Gaussian process classifier (GPC) is a very promising machine learning concept that is based on a probabilistic model of the underlying class probabilities. The model is based on a Bayesian formulation that combines the observed class memberships with an assumed prior distribution. While its counterpart, the Gaussian process regression (GPR), has a simple closed-form solution and therefore enjoys wide applicability, the GPC has some computational issues to overcome. It is given in terms of a prohibitive multi-integral formula, that can only be solved through some approximations or using lengthy algorithms. In this talk I will review the theory behind GPC's. I will review its Bayesian formulation, and its parameter estimation methods. In addition, I will propose a new algorithm that leads to an exact evaluation of the classification.
Monday, February 14, 2011
Free and open to the public