Speaker: Prof. Tobias Ritschel (University College London)
I will discuss two tracks of methods which are "advanced" in so far as they explore designs beyond classic supervised learning of a given tunable rendering pipeline. The first track makes learning itself subject to learning (meta-learning). While normal learning optimizes parameters of a tunable pipeline, meta-learning optimized the parameters of the learning. I will discuss two instances that meta-learn learning rate and initialization, and another one that meta-learns sampling.
The second track investigates the differentiability of the rendering pipeline itself. We will recall why it is challenging to differentiate through rasterization or ray-tracing. Based on this framework, we will derive methods to optimize over the space of differentiable rasterizers as well as propose a simple and effective way to differentiate the light transport equation --which has a lot of dimensions to (MC) integrate over-- by adding even more dimensions.
Biography:
Professor Tobias Ritschel has received his PhD from Saarland University (Max Planck Institute) in 2009. He was a post-doctoral researcher at Telecom ParisTech / CNRS 2009-10 and a Senior Researcher at MPI 2010-15. Tobias was appointed Senior Lecturer at University College London in 2015 where he was named Full Professor of Computer Graphics in 2019. His work has received the Eurographics Dissertation (2010) and Young Researcher Award (2014). His interests include Image Synthesis and Human Visual Perception, now frequently including applied AI.
