Zur NavigationLogin


IRTG Lecture

Friday, July 15, 2016 10:00 AM;

JGU Mainz, Physics, Newton-Raum


Michael Wand: Generative models for structuring geometric and visual data (and maybe more?)

My talk will consist of three parts, chronologically ordered: First, I will give a very brief overview of research topics and results in my research group from recent years. Our goal has been (and still is) to devise methods for a machine understanding of visual and geometric data. The key is to define structure models that capture redundancy in the data, such as correspondences (recurring pieces), shape grammars (assembly rules), or dynamical constraints (using physical priors for inferring compact descriptions).
The second part addresses a recent development in our field, the magnitude of the impact of which has caught many of us by surprise: The advent of powerful meta-programming algorithms, currently in particular in the form of "deep neural networks", which create such structure models in a much more automated fashion: Rather than hand-crafting structure models, these methods learn key structural properties along with processing algorithms from large quantities of example data. Very little human guided prior assumptions are required; this is mostly limited to elementary symmetry properties and model complexity constrains for regularization. I will show some recent experiments performed in our group that illustrate how generative models of complex classes of data models with complex correlation structure can be learned fully automatically. In this context, "generative model" means that we cannot only classify data, but also reproduce the described phenomenon by sampling and/or fitting to given boundary conditions.
This aspect motivates the third and last part of my talk, which will be a bit more speculative: Given the ability to capture complex generative data models, what are opportunities for research at the intersection of computer science and physics? Ideas might include using such representations to aid predictions and simulations of physical systems, as well as attempts to understand better why the "deep" models are so successful in characterizing real-world phenomena.


Fabrice Delbary: Determining a pair potential from a radial distribution function


Download iCal-file


This page uses cookies. By continuing to visit this site you agree to our use of cookies.

More info