Gurprit Singh from MPI will visit the Stream team next Monday and will give a talk entitled: How to Train Your Samples?

Abstract: Samples are the basic building block in many domains including computer graphics/vision and financial mathematics. In this talk, we will see how sample correlations play a crucial role in many areas including denoising, geometry reconstruction from point clouds, object placement and variance reduction during Monte Carlo integration for rendering. We briefly also touch upon the effectiveness of different sample shapes like points, segments and lines and look at different spatial and spectral tools that can be used to analyze sample correlations.

In the later part, I will present modern deep learning pipeline to generate these sample correlations in multiple dimensions. We show how–without any training data–a neural network can optimize correlations that were not possible to conceive earlier in a straight forward manner. The simplicity and expressivity of our architecture makes the tedious and mathematically involved pattern generation tasks a thing of the past.

Bio: Gurprit Singh is a senior researcher, leading the sampling and rendering group at the Max Planck Institute for Informatics (MPII), Saarburcken, Germany. He did his PhD from Université de Lyon 1, France, in 2015. This was followed by a postdoc at Dartmouth College, USA, with another two-year postdoc at MPII Saarbrucken, before he accepted the senior researcher position at MPII. His group focuses on understanding and synthesizing sample correlations for variance reduction in Monte Carlo and Quasi-Monte Carlo integration problems with direct application for image synthesis. He uses both theoretical and deep learning based tools to gain further understanding of how nature encodes correlations in objects that revolves around us.