[Colloquium] [TTIC Talks] REMINDER: 11/11 TTIC Colloquium: Misha Belkin, Ohio State University

[Alicia McClarin]

*When:*      Monday, November 11th at 11:00am



*Where:*     TTIC, 6045 S. Kenwood Avenue, 5th Floor, Room 526



*Who: *       Misha Belkin, Ohio State University


*Title:        *Toward Unified Theory for Supervised Inference: The Lessons
of Deep Learning

*Abstract:  *For years Empirical Risk Minimization (ERM) served as the main
theoretical foundation for machine learning. Yet, even early on, certain
phenomena, such as lack of over-fitting in boosting, posed a challenge to
the ERM  view.  In recent years these foundational cracks have widened
precipitously, as best practices of deep learning have been  in direct
contradiction to the methodologies suggested by analyses of ERM or
bias-variance trade-off.   Indeed, as I will discuss, modern methods
operate in regimes, mostly overlooked in the past, that are best described
by interpolation, where the model is selected from the space of predictors
fitting the training data exactly, according  to a certain inductive bias.
Furthermore, I will show how  classical (capacity based) models and modern
models based on inductive biases  can  be unified within a single "double
descent" risk curve,  which subsumes the classical U-shaped bias-variance
trade-off. While  still speculative, I argue that a general unified theory
of supervised learning may be possible.

Finally, I will discuss some important implications for optimization, and
make a few comments and observations on the nature of the inductive biases
and the (considerable) gaps in our understanding.

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


For more information on the colloquium series or to subscribe to the
mailing list, please see http://www.ttic.edu/colloquium.php.

-- 
*Alicia McClarin*
*Toyota Technological Institute at Chicago*
*6045 S. Kenwood Ave., **Office 518*
*Chicago, IL 60637*
*773-834-3321*
*www.ttic.edu* <http://www.ttic.edu/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20191110/7e9ec580/attachment.html>