Departement Informatik, Universität Basel
Departement Informatik, Universität Basel
Departement Informatik, Universität Basel
Departement Informatik, Universität Basel

Colloquium - 19.05.2016

TitleProbabilistic Numerics — Uncertainty for (AI) Computation
SpeakerPhilipp Hennig (Max-Planck-Institute for Intelligent Systems, Tübingen)
TimeThursday 19 May 2016, from 17:15 - 18:15 o'clock
LocationDepartment of Mathematics and Computer Science, Seminar Room 00.003, University of Basel, Spiegelgasse 1, Basel
Abstract

Numerical problems — integration, optimization, linear algebra, and the solution of differential equations — are the computational bottleneck of artificial intelligence. At the moment, these tasks are typically solved by standard algorithms. This “black box” approach saves development time, but restricts efficiency and reliability at runtime.

Numerical methods estimate unknown quantities given tractable computations. They can thus themselves be phrased as probabilistic inference, as elementary “intelligent” systems that act to collect information about a mathematical quantity. This insight gives rise to a new class of *probabilistic numerical methods*, which includes the classic methods as special cases, but allows novel functionality, and a more fine-grained description of individual numerical tasks. I will show that probabilistic numerics can provide mathematically well-founded, computationally efficient answers for central algorithmic challenges of intelligent machines.

Bio:

Philipp Hennig leads the Emmy Noether group on Probabilistic Numerics at the Max Planck Institute for Intelligent Systems. He received his PhD from the University of Cambridge in 2011, for a thesis on approximate inference in graphical models, under the supervision of David MacKay. Since then he has been interested in formulations of computation as the collection of information in a probabilistic sense. His group has contributed numerical methods for both the inner loop (linear algebra, convex and discrete optimization, differential equations) and outer loops (Bayesian optimization, active design) of learning machines. He is a co-founder of http://probabilistic-numerics.org , and has co-organized a number of workshops, meetings and tutorials on the topic of probabilistic numerical methods.

 

After the talk an informal apéro will be offered. I hope you’ll be joining us.