Search Mailing List Archives


Limit search to: Subject & Body Subject Author
Sort by: Reverse Sort
Limit to: All This Week Last Week This Month Last Month
Select Date Range     through    

[theory-seminar] Omer Reingold - Guilt-Free Interactive Data Analysis: The Reusable Holdout

Huacheng Yu yuhch123 at cs.stanford.edu
Fri Nov 14 16:15:20 PST 2014


Hi,
   We are going to have a theory seminar next Thursday, Nov 20th, at 4:15pm
in Gates 463A. Our speaker will be Omer Reingold. See info below!

- Huacheng Yu

--------------------------------------------------------

Guilt-Free Interactive Data Analysis: The Reusable Holdout

A great deal of effort has been made to reduce the risk of spurious
scientific discoveries, from the use of holdout sets and sophisticated
cross-validation techniques, to procedures for controlling the false
discovery rate in multiple hypothesis testing.  However, there is a
fundamental disconnect between the theoretical results and the practice of
science: the theory assumes a fixed collection of hypotheses to be tested,
or learning algorithms to be applied, selected non-adaptively before the
data are gathered, whereas science is by definition an adaptive process, in
which data are shared and re-used, and hypotheses and new studies are
generated on the basis of data exploration and previous outcomes.

Surprisingly, the challenges of adaptivity can be addressed using insights
from differential privacy, a field of study supporting a definition of
privacy tailored to private data analysis.  As a corollary we show how to
safely reuse a holdout set a great many times without undermining its power
of ``correctness protection,'' even when hypotheses and computations are
chosen adaptively.  Armed with this technique, the analyst is free to
explore the data ad libitum, generating and evaluating hypotheses,
verifying results on the holdout, and backtracking as needed.

Joint work with Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toni Pitassi
and Aaron Roth
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.stanford.edu/pipermail/theory-seminar/attachments/20141114/3b6b8a5a/attachment.html>


More information about the theory-seminar mailing list