BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Lent Talklets: Costanza and Cătălina - Costanza Conforti (Depart
 ment of Theoretical and Applied Linguistics\, University of Cambridge)\; C
 ătălina Cangea (Department of Computer Science and Technology\, Universi
 ty of Cambridge)
DTSTART:20210226T130000Z
DTEND:20210226T140000Z
UID:TALK157663@talks.cam.ac.uk
CONTACT:Agnieszka Slowik
DESCRIPTION:Speaker 1: Costanza Conforti\n\nTitle: NLP-enhanced Sustainabl
 e Development: the case of Community Profiling in Rural Uganda\n\nAbstract
 : In recent years\, there has been an increasing interest in the applicati
 on of AI (and especially Machine Learning) to the field of Sustainable Dev
 elopment (SD). However\, until now\, NLP has not been systematically appli
 ed in this context. In this talk\, we discuss the high potential of NLP to
  enhance community profiling in developing countries. We introduce the new
  task of Automatic User-Perceived Value classification\, and we release an
  expert-annotated dataset of interviews carried out in rural Uganda. Exper
 imental results show that the problem is challenging\, and leaves consider
 able room for future research at the intersection of NLP and SD.\n\nSpeake
 r 2: Cătălina Cangea\n\nTitle: Exploiting multimodality and structure in
  world representations\n\nAbstract: In this talk\, I will give an overview
  of the major research works I have been involved in during my PhD\, which
  study and develop likely aspects of future intelligent agents. The first 
 contribution centers on vision-and-language learning\, introducing a chall
 enging embodied task that shifts the focus of Embodied Question Answering 
 to the visual reasoning problem\, along with several models that were eval
 uated on the novel dataset. The second work presents two ways of obtaining
  hierarchical representations of graph-structured data. These methods eith
 er scaled to much larger graphs than the ones processed by contemporary be
 st-performing models\, or incorporated theoretical properties via the use 
 of topological data analysis algorithms\; both competed with state-of-the-
 art graph classification methods. Finally\, the third contribution delves 
 further into relational learning\, presenting a probabilistic treatment of
  graph representations in complex settings such as few-shot & multi-task l
 earning and scarce labelled-data regimes. By adding relational inductive b
 iases to neural processes\, the resulting framework can model an entire di
 stribution of functions which generate datasets with structure. This yield
 ed significant performance gains in the aforementioned complex scenarios\,
  with semantically-accurate uncertainty estimates that drastically improve
 d over the neural process baseline.
LOCATION:Remote
END:VEVENT
END:VCALENDAR