[Ml-stat-talks] Fwd: [ORFE-Seminars] Today ORFE Department Colloquium: Tuesday, September 26, at 4:30 pm, 101 Sherrerd Hall

Barbara Engelhardt bee at princeton.edu
Tue Sep 26 10:21:15 EDT 2017

Talk of interest today.

* ===== Today ORFE Department Colloquium Announcement===== *

DATE:             Tuesday, September 26, 2017

TIME:             4:30pm

LOCATION:   Sherrerd Hall, Room 101

SPEAKER:      Sanjoy Dasgupta, UC, San Diego

Interactive Learning Structures

We introduce “interactive structure learning”, an abstract problem in which
a structure (classifier, clustering, topic model, knowledge graph, etc) is
to be learned, using unlabeled data as well as rounds of human interaction.
This formalism covers many situations of practical interest, such as
query-based classifier learning and interactive clustering. Interactive
structure learning presents a host of statistical and algorithmic
challenges. We discuss two such results. 1. Learning from partial
correction. Earlier models of interaction have typically adopted a
question-answer paradigm: the learner asks a question and a human expert
answers it. Our abstract model allows a richer and more flexible interface,
where the learner provides a small snapshot of its current model (for
instance, the restriction of a clustering to a few points), and the expert
can selectively fix any part of it. This kind of feedback is not iid.
Nonetheless, we show how statistical generalization bounds can be given for
structures learned in this way. The proof technique may also be of interest
in other situations with non-iid sampling. 2. Interactive hierarchical
clustering. We present a general-purpose algorithm for interactive
structure learning and illustrate how it works in the case of hierarchical
clustering. Along the way, we present a novel cost function for
hierarchical clustering, as well as an efficient algorithm for
approximately minimizing this cost.

Bio: Sanjoy Dasgupta is a Professor in the Department of Computer Science
and Engineering at UC San Diego. He works on algorithms for machine
learning, with a focus on unsupervised and interactive learning.


Barbara E Engelhardt
Assistant Professor
Department of Computer Science
Center for Statistics and Machine Learning
Princeton University
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/ml-stat-talks/attachments/20170926/53053271/attachment.html>

More information about the Ml-stat-talks mailing list