Mark van Rossum and tbc |
Tue 14 Nov 2017, 11:00 - 12:00 |
IF 4.31/4.33 |
If you have a question about this talk, please contact: Gareth Beedham (gbeedham)
Mark van Rossum
Details to follow.
Michael Gutmann
Title: Conditional Noise-Contrastive Estimation
Noise-contrastive estimation (NCE) is a method for learning the parameters of models that are specified up their normalising partition function. Such models occur, for instance, in image or natural language processing. NCE works by training a classifier to distinguish between the observed and some reference data (the "noise" in noise-contrastive estimation). I present a version of NCE where the noise distribution is specified conditional on the observed data, which allows one to partly automate its specification.
This work was done as part of the MSc thesis by Ciwan Ceylan during his visit to Informatics this spring. The full thesis is available here:
https://kth.diva-portal.org/smash/get/diva2:1139156/FULLTEXT01.pdf