[Statlist] Seminar D. Janzing, 2012/10/24, 11:00 CAB H 52

Buhmann Joachim M. jbuhm@nn @end|ng |rom |n|@ethz@ch
Mon Oct 15 10:42:41 CEST 2012


Dear all

It is my pleasure to invite you to the following machine learning seminar:

**********************************************************************************
Dominik Janzing,
Max Planck Institute for Intelligent Systems,
Tübingen

DATE: Wednesday, 2012/10/24
TIME: 11:00
ROOM: CAB H 52

Introduction to information geometric causal inference
**********************************************************************************

Conventional methods of causal inference are based on conditional
independences and therefore require at least 3 observed variables.
New approaches also account for properties of the joint distribution
other than independences. They are in principle also able to infer
whether X causes Y or Y causes X for just two observed variables X,Y.
The idea is that if X causes Y then P(X) and P(Y|X) correspond to
independent mechanisms of the world, they therefore provide no
information about each other.

Information geometric causal inference focuses on the simple case
where Y is a deterministic function of X and vice versa. The idea is
that Y=f(X) is only plausible as causal model if P(X) contains no
information on the structure of f, while P(Y) does contain information
on f because it tends to be peaked in regions where the Jacobian of
f is small. We have developed a simple inference method that is based
on this observation. We obtained quite positive results on data sets
with known ground truth.

The method has a simple geometrical interpretation in the space of
probability distributions which suggests that it can be generalized to
non-deterministic relations.



References:
[1] Daniusis et al: Inferring deterministic causal relations, UAI 2010
[2] Janzing et al: Information-geometric approach to inferring
causal directions, AI 2012

(Host: J. Buhmann)


=====================================================================
Joachim M. Buhmann
Department of Computer Science                                     Tel.(office) : +41-44-63 23124
ETH Zentrum, CAB G 69.2                                               Tel.(secret.): +41-44-63 26496
Universitätstrasse 6                                                                       Fax: +41-44-63 21562
CH-8092 Zurich, Switzerland                                              email: jbuhmann using inf.ethz.ch<mailto:jbuhmann using inf.ethz.ch>
=====================================================================




More information about the Statlist mailing list