Title: Bayesian inference for shape and high-dimensional inputs with a Gaussian process prior
Authors: Chafik Samir - UCA-LMBP/CNRS (France) [presenting]
Anis Fradi - UCA (France)
Abstract: A Bayesian point of view is adopted for studying shapes and high-dimensional data. Since computing the exact marginal likelihood for a Bayesian model remains difficult if not impossible in high-dimensional and manifolds data, we introduce two types of methods to improve the efficiency and the scalability of Gaussian processes. We to handle multimodal posteriors and the prediction quality, jointly. We furthermore show some asymptotic properties such as non-unbiasedness, prediction sufficiency, etc. The proposed approaches are shown to provide computational advantages with respect to some existing methods that rely on modified covariance function. Various tests on real and simulated data will be discussed as well as the efficiency of Bayesian and non-Bayesian predictors.