BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Ensembles for Discovery of Compact Structures and Learning Back-pr
 opagation Forests. - Madalina Fiterau\, Carnegie Mellon University
DTSTART:20150310T090000Z
DTEND:20150310T100000Z
UID:TALK58372@talks.cam.ac.uk
CONTACT:Microsoft Research Cambridge Talks Admins
DESCRIPTION:In many practical scenarios\, complex high-dimensional data co
 ntains low-dimensional structures that could be informative of the analyti
 c problems at hand. I will present a method that detects such structures i
 f they exist\, and uses them to construct compact interpretable models for
  different machine learning tasks that can benefit practical applications.
  \nTo start with\, I will formalize Informative Projection Recovery\, the 
 problem of extracting a small set of low-dimensional projections of data t
 hat jointly support an accurate model for a given learning task. Our solut
 ion to this problem is a regression-based algorithm that identifies inform
 ative projections by optimizing over a matrix of point-wise loss estimator
 s. It generalizes to multiple types of machine learning problems\, offerin
 g solutions to classification\, clustering\, regression\, and active learn
 ing tasks. Experiments show that our method can discover and leverage low-
 dimensional structures in data\, yielding accurate and compact models. Our
  method is particularly useful in applications in which expert assessment 
 of the results is of the essence\, such as classification tasks in the hea
 lthcare domain.\n\nIn the second part of the talk\, I will describe back-p
 ropagation forests\, a new type of ensemble that achieves improved accurac
 y over existing ensemble classifiers such as random forests classifiers or
  alternating decision forests. This research was performed under the mento
 rship of Dr. Peter Kontschieder and in collaboration with Dr. Samuel Rota-
 Bulò (FBK Trento\, IT). Back-propagation (BP) trees use soft splits\, suc
 h that a sample is probabilistically assigned to all the leaves. Also\, th
 e leaves assign a distribution across the labels. The splitting parameters
  are obtained through SGD by optimizing the log loss over the entire tree\
 , which is a non-convex objective. The probability distribution over the l
 eaves is computed exactly by maximizing a log concave procedure. In additi
 on\, I will present several proposed approaches for the use of BP forests 
 within the context of compact informative structure discovery.
LOCATION:Auditorium\, Microsoft Research Ltd\, 21 Station Road\, Cambridge
 \, CB1 2FB
END:VEVENT
END:VCALENDAR