BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.2.3.1//EN
TZID:Europe/Paris
X-WR-TIMEZONE:Europe/Paris
BEGIN:VEVENT
UID:5513@i2m.univ-amu.fr
DTSTART;TZID=Europe/Paris:20130704T140000
DTEND;TZID=Europe/Paris:20130704T150000
DTSTAMP:20241029T173609Z
URL:https://www.i2m.univ-amu.fr/evenements/r-bailly-upc-spectral-learning-
 of-hidden-structure-or-how-nlp-natural-language-processing-is-just-a-matte
 r-of-compressive-sensing/
SUMMARY: (...): R. Bailly (UPC) : Spectral Learning of Hidden Structure or 
 "How NLP (Natural Language Processing) is just a matter of compressive sen
 sing"
DESCRIPTION:: Spectral Learning of Hidden Structure or "How NLP (Natural La
 nguage Processing) is just a matter of compressive sensing."\nBy Raphaël 
 Bailly\\\, Universitat Politècnica de Catalunya.\n\nThe spectral algorith
 m\\\, as it has been introduced\\\, deals with sequences of observables sy
 mbols with a hidden sequence of states. In this talk\\\, we consider a sup
 plementary hidden layer - e.g. a hidden tree structure over a sequence\\\,
  or a link structure between an input and an output sequence. We propose a
  framework to apply the spectral method to this kind of problem\\\, even i
 f the Hankel matrix is unknown. Each observed statistic (marginalization o
 ver all possible hidden structures for a given sequence) can be turned int
 o a linear constraint over the statistics concerning the joint probabiliti
 es\\\, which are the entries of the Hankel matrix. This leads to an optimi
 zation problem of rank minimization over linear constraints\\\, which can 
 be efficiently solved thanks to the nuclear norm convex relaxation of the 
 rank. This framework applies to all kinds of hidden structures\\\, provide
 d that there exists efficient ways of marginalizing over the hidden layer\
 \\, such as forward-backward or inside-outside algorithms. We will study i
 n details the case of the general transducers\\\, i.e. with possible empty
  observations: input and ouput can be generated jointly or independently\\
 \, thus inducing a hidden structure over pairs of sequences.
CATEGORIES:Séminaire,Signal et Apprentissage
END:VEVENT
BEGIN:VTIMEZONE
TZID:Europe/Paris
X-LIC-LOCATION:Europe/Paris
BEGIN:DAYLIGHT
DTSTART:20130331T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
END:DAYLIGHT
END:VTIMEZONE
END:VCALENDAR