Seminar

CVL Talk: Rethinking classical optimization in Graphics - Michael Fischer, UCL

Abstract:

Gradient-based optimization has become the de-facto standard for optimization in the Graphics community. However, it is rarely questioned whether we can do “better” than simple gradient descent (GD) - in fact, it is easy to construct examples where GD is easily outperformed by other methods. This talk will explore two such scenarios: Firstly, we will use meta-learning to optimize over GD itself, leading to improved convergence and insights for data sampling in appearance reproduction scenarios. Secondly, we will replace the traditional (and sometimes uninformative and brittle) gradients produced by Autodiff with a smoother formulation which we achieve via a convolution in (optimization-) parameter space. This variational formulation enables successful differentiable path tracing on problems where previous approaches did not converge due to plateaus in the cost landscape. 

Zoom Link

You can see the other CVPR 2023 UBC events here.