Gradients, whatever it takes

Speaker: Tobias Ritschel, University College London

Abstract: Differentiation of tunable functions is at the core of AI, but some desired functions might not provide gradients in practice: Maybe it's the math that is in the way, they could be black boxes of which we lack the source code to do autodiff, sometimes the gradients are too expensive to evaluate or their symbolic expression is simply intractable. Even worse, some gradients might exist, but show to be useless because they are (numerically) zero.
In this talk we will discuss three example problems from that area and their solutions. The first is 'differentiating' functions with zero gradients that occur in visibility problems in Computer Graphics. The second differentiates black box systems using neural proxies. The last one numerically finds functions that have a desired other function as their (higher order) gradient that are intractable symbolically.

Biography: Tobias received his PhD from Saarland University / MPI Informatik 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 EG Dissertation (2010) and Young Researcher Award (2014). He served as general co-chair of Eurorgaphics 2025 in London. His interests include Image Synthesis and Human Visual Perception, now frequently including applied AI.

Chair: Prof. Piotr Didyk