Abstract

This talk has two (for now) loosely connected parts: In the first part we aim to provide intuition for the key mechanisms underlying the sequential Monte Carlo (SMC) method (including the popular particle filters and smoothers). SMC provide approximate solutions to integration problems where there is a sequential structure present. The classical example of such a structure is offered by nonlinear dynamical systems, but we stress that SMC is significantly more general than most of us first thought. We will hint at a few ways in which SMC fits into the machine learning toolbox and mention a few interesting avenues for research. In the second part we develop a new approach to deep regression. While deep learning-based classification is generally addressed using standardized approaches, a wide variety of techniques are employed when it comes to regression. We have developed a new and general deep regression method with a clear probabilistic interpretation. We obtain good performance on several computer vision regression tasks (including a new state-of-the-art result on visual tracking). The loose connection lies in the use of the Monte Carlo idea in both topics. We do believe that the connection between the two seemingly disparate topics will be strengthened over the coming years.