BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Multiscale methods and recursion in data science - Piotr Fryzlewic
 z (London School of Economics)
DTSTART:20180323T113000Z
DTEND:20180323T123000Z
UID:TALK102847@talks.cam.ac.uk
CONTACT:INI IT
DESCRIPTION:The talk starts on a general note: we first attempt to define 
 a "multiscale" method / algorithm as a recursive program acting on a datas
 et in a suitable way. Wavelet transformations\, unbalanced wavelet transfo
 rmations and binary segmentation are all examples of multiscale methods in
  this sense. Using the example of binary segmentation\, we illustrate the 
 benefits of the recursive formulation of multiscale algorithms from the so
 ftware implementation and theoretical tractability viewpoints. <br><br>We 
 then turn more specific and study the canonical problem of a-posteriori de
 tection of multiple change-points in the mean of a piecewise-constant sign
 al observed with noise. Even in this simple set-up\, many publicly availab
 le state-of-the-art methods struggle for certain classes of signals. In pa
 rticular\, this misperformance is observed in methods that work by minimis
 ing a "fit to the data plus a penalty" criterion\, the reason being that i
 t is challenging to think of a penalty that works well over a wide range o
 f signal classes. To overcome this issue\, we propose a new approach where
 by methods learn from the data as they proceed\, and\, as a result\, opera
 te differently for different signal classes. As an example of this approac
 h\, we revisit our earlier change-point detection algorithm\, Wild Binary 
 Segmentation\, and make it data-adaptive by equipping it with a recursive 
 mechanism for deciding "on the fly" how many sub-samples of the input data
  to draw\, and w here to draw them. This is in contrast to the original Wi
 ld Binary Segmentation\, which is not recursive. We show that this signifi
 cantly improves the algorithm particularly for signals with frequent chang
 e-points. <br><br>Related Links<ul><li><a target="_blank" rel="nofollow" h
 ref="http://www-old.newton.ac.uk/cgi/https%3A%2F%2FCRAN.R-project.org%2Fpa
 ckage%3Dbreakfast">https://CRAN.R-project.org/package=breakfast</a> - R so
 ftware package "breakfast" (provides an implementation of Adaptive Wild Bi
 nary Segmentation)</li></ul>
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR