BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Continuous feature structures: Can we learn structured representat
 ions with neural networks? - Guy Emerson\, NLIP\, University of Cambridge
DTSTART:20190607T110000Z
DTEND:20190607T120000Z
UID:TALK120529@talks.cam.ac.uk
CONTACT:Andrew Caines
DESCRIPTION:The basic data structure for neural network models is the vect
 or. While this is computationally efficient\, vector representations requi
 re a fixed number of dimensions\, which makes it impossible to encode even
  basic data structures that would be familiar to a first-year undergraduat
 e\, such as lists\, trees\, and graphs. In this talk\, I will focus on fea
 ture structures\, a general-purpose data structure which is notably used i
 n HPSG grammars. One challenge with learning such structured representatio
 ns is that they are discrete\, which rules out training with gradient desc
 ent. In this talk\, I will present a continuous relaxation of feature stru
 ctures\, which allows them to be used in neural networks trained by gradie
 nt descent. In particular\, I will show how these continuous feature struc
 tures can replace vectors in an LSTM\, which would make it possible to lea
 rn feature structure representations of sentences.
LOCATION:FW26\, Computer Laboratory
END:VEVENT
END:VCALENDAR