Point correlations for graphics, vision and machine learning
Monday, February 20, 2023
Sampling is at the core of computer graphics, vision and machine learning. Sample correlations can significantly impact the accuracy of the algorithm. For example, global illumination for image synthesis involves Monte Carlo (MC) integration where negatively-correlated samples give better convergence and perceptually pleasing images compared to random sampling. In computer vision, the quality of surface reconstruction from point clouds largely depends on the distribution of points. In machine learning, negative correlations (e.g., antithetic sampling) has shown significant improvements in the training and at inference time. In this talk, we will see how randomness can be tailored for certain applications to improve the overall efficiency of existing models. We start with different correlations studied in computer graphics. Next, we discuss the tools used to characterize these correlations and how these tools are used to develop different correlations via neural networks. We conclude by looking at different applications where correlations can make a significant impact.
Post Talk Link: Click Here
Passcode: FgCX=981
Speaker/s
Gurprit Singh is leading the sampling and rendering group in the computer graphics department at the Max Planck Institute for Informatics in Saarbrücken, Germany. Before that, Gurprit spent two wonderful years at Dartmouth College working with Wojciech Jarosz followed by another two-year postdoc working with Karol Myszkowski. He obtained his PhD from Université Lyon 1 in France, under the supervision of Victor Ostromoukhov. His research revolves around sampling which is the basic building block in many domains including computer graphics, computer vision and machine learning.
Related
Polygenic Score Modeling to Investigate Genotype-Phenotype Associations
Read More
Trustworthy Machine Learning: Transparency, Collaboration, and Evaluation
Read More