Challenges in Causality
25Google Tech Talks
February, 11 2008
ABSTRACT
What affects your health, the economy, climate changes? And what actions will
have beneficial effects? These are some of the central questions of causal
discovery. A "causal model" is a model capable of making predictions under
changing circumstances, corresponding to actions of "external agents" on a
system of interest. For example, a doctor administering a drug to a patient, a
government enforcing a new tax law or a new environmental policy. It is often
necessary to assess the benefits and risks of potential actions using available
past data and excluding the possibility of experimenting. Experiments, which
are the ultimate way of verifying causal relationships, are in many cases too
costly, infeasible, or unethical. For instance, enforcing a law prohibiting to
smoke in public places is costly, preventing people from smoking may be
infeasible, and forcing them to smoke would be unethical. In contrast,
"observational data" are available in abundance in many applications. Recently,
methods to devise causal models from observational data have been proposed. Can
causal models thus obtained be relied upon to make important decisions? In this
presentation, we will challenge the hopes an promises of causal discovery and
present new means of assessing the validity of causal modeling techniques.
Want to play? Check the "causation and prediction" competition presently going
on: http://www.causality.inf.ethz.ch/challenge.php
Deadline April 30, 2008
Speaker: Isabelle Guyon
Isabelle Guyon is a researcher in machine learning and an independent
consultant. Prior to starting her consulting practice in 1996, she
worked at AT&T Bell Laboratories, where she pioneered applications of neural
networks to pen computer interfaces and invented Support Vector Machines (in
collaboration with B. Boser and V. Vapnik). Isabelle Guyon holds a Ph.D. degree
in Physical Sciences of the University Pierre and Marie Curie of Paris, France.
She is vice-president of the Unipen foundation, action editor of the Journal of
Machine Learning Research, and competition chair of the IJCNN conference.