[Theory] [TTIC Talks] 1/9 TTIC Colloquium: Leon Bottou, Facebook

[Brandie Jones]

*When: *Monday, *January 9th at ** 11:30AM CT*

*Where: *Talk will be given *live, in-person* at

                    TTIC, 6045 S. Kenwood Avenue

                    5th Floor, Room 530


*Virtually: *via Panopto (Livestream
<https://uchicago.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=a8661e99-062f-42ff-9fd5-af780145c4d5>
)

*Who: *Leon Bottou, Facebook

*Title: *Out-of-distribution generalization, causation, and features

*Abstract:* Out-of-distribution generalization is only possible when we
assume that there is a connection between the training and testing
distribution. For instance, we can make the assumption that these
distributions share a same causal model but differ because of changes in
the distribution of uncontrolled variables. But how then to recover enough
information about this causal model and these uncontrolled variables.
Learning causal invariances (Arjovsky et al. 2019) offers an interesting
workaround but is plagued by optimization problems that severely limit its
impact.  We propose to first initialize the networks with a rich
representation containing a palette of potentially useful features, ready
to be used by even simple models. Such a representation is constructed with
a succession of specially crafted training episodes. This approach
consistently helps six OoD methods achieve top performance on ColoredMNIST
benchmark. The same technique substantially outperforms comparable results
on the Wilds Camelyon17 task, eliminates the high result variance that
plagues other methods, and makes hyperparameter tuning and model selection
more reliable.

*Bio: *Léon Bottou received the Diplôme d'Ingénieur de l'École
Polytechnique (X84) in 1987, the Magistère de Mathématiques Fondamentales
et Appliquées et d'Informatique from École Normale Supérieure in 1988, and a
Ph.D. in Computer Science from Université de Paris-Sud in 1991. His research
career took him to AT&T Bell Laboratories, AT&T Labs Research, NEC Labs
America, and Microsoft Research. He joined Facebook AI Research in 2015.
The long-term goal of Léon Bottou's research is to understand and replicate
human-level intelligence. Because this goal requires conceptual advances
that cannot be anticipated, Leon's research has followed many practical and
theoretical turns: neural networks applications in the late 1980s,
stochastic gradient learning algorithms and statistical properties of
learning systems in the early 1990s, computer vision applications with
structured outputs in the late 1990s, theory of large scale learning in the
2000s. During the last few years, Léon Bottou's research aims to clarify
the relation between learning and reasoning, with more and more focus on
the many aspects of causation (inference, invariance, reasoning,
affordance, and intuition.


*Host: Nati Srebro <nati at ttic.edu>*



-- 
*Brandie Jones *
*Executive **Administrative Assistant*
Toyota Technological Institute
6045 S. Kenwood Avenue
Chicago, IL  60637
www.ttic.edu
Working Remotely on Tuesdays
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/theory/attachments/20221229/22f408fa/attachment.html>