What's the right complexity measure for inferring causal relations?

Add to your list(s) Download to your calendar using vCal

• Dominik Janzing (Max Planck Institute Tuebingen)
• Friday 26 February 2010, 15:30-16:30
• MR12, CMS, Wilberforce Road, Cambridge, CB3 0WB.

If you have a question about this talk, please contact rbg24.

If X causes Y for two random variables X and Y, we expect that the factorization of P(X,Y) into P(X)P(Y|X) is simpler than the non-causal factorization P(Y)P(X|Y). This is because P(Y|X) describes the causal mechanism while P(X|Y) is “only a mathematical expression”.

Discussions have shown that a lot of researchers agree on this intuition. Since we would like to use this principle for inferring causal directions, we are left with two problems:

(1) what does “simple” mean?

(2) is there any deeper justification for this principle?

Our answer to question (2) is a clear “yes” if complexity vs simplicity is measured in terms of Kolmogorov complexity: I will present a theory of causal inference that generalizes the framework of Bayesian networks to algorithmic instead of statistical conditional dependences. I will show that our theory implies the above inference principle. However, since Kolmogorov complexity is uncomputable we still need complexity measures that are appropriate for practical implementations.

I will present some first small steps towards this challenging goal.

http://www.kyb.mpg.de/~janzing

This talk is part of the Statistics series.

This talk is included in these lists:

• All CMS events
• All Talks (aka the CURE list)
• CMS Events
• Cambridge Forum of Science and Humanities
• Cambridge Language Sciences
• Cambridge talks
• Chris Davis' list
• DPMMS Lists
• DPMMS info aggregator
• DPMMS lists
• Guy Emerson's list
• Hanchen DaDaDash
• Interested Talks
• MR12, CMS, Wilberforce Road, Cambridge, CB3 0WB
• Machine Learning
• School of Physical Sciences
• Statistical Laboratory info aggregator
• Statistics
• Statistics Group
• bld31
• custom
• rp587

Note that ex-directory lists are not shown.