January, Tuesday 23rd

14:30 (usual room 2014, building 660) (see location )

Olivier Goudet & Diviyan Kalainathan

(TAU team)

Title: End-to-end Causal Generative Neural Networks


We introduce ECGNN (End-to-end Causal Generative Neural Networks), a new
framework to discover causal relationships as a directed graph out of
observational data, using generative networks. This framework consists
in training multiple generative networks in parallel, using either the
maximum mean discrepancy metric or an adversarial discriminator to match
the distributions of the generated variables and the observed variables.
Considering the graph recovery problem as an optimization problem, ECGNN
successfully combines robust performance with fair computational cost
and scalability.

Contact: guillaume.charpiat at inria.fr
All TAU seminars: here

Contributors to this page: guillaume .
Page last modified on Monday 15 of January, 2018 12:27:23 CET by guillaume.