Machine learning and statistical models are broadly used in artificial intelligence. Inference refers to the task of estimating the distribution of plausible parameter values conditional on the observed data available to fit the model, and it is typically carried out by advanced Markov Chain Monte Carlo (MCMC) algorithms. Even though the algorithms used in modern probabilistic programming and deep learning systems are relatively efficient and accurate, they have problems exploring the whole posterior distribution. More accurate samplers have been developed based on differential geometry, conducting inference on a Riemannian manifold, but they are computationally inefficient and hence not used in practice. This project improves the computational efficiency of MCMC algorithms operating in a Riemannian geometry, making them a feasible alternative for Bayesian inference in probablistic programming and deep learning.

2022

2024