Dependent Component Analysis

Dependent Component Analysis

Seminar author:Janis Nötzel

Event date and time:04/28/2016 02:30:pm

Event location:GIQ Seminar room

Event contact:

We present an information-theoretic analysis of a question right at the heart of unsupervised learning approaches: 
Assume we are collecting a number K of observations about some event E from K different agents. Can we infer E from them without exactly knowing the behaviour of each of the agents? We model this task by letting the events be distributed according to a distribution p and the task is to estimate p under unknown and independent noise. It turns out that this task is feasible if
multiple copies of the true data are available. If the true distribution and the observations are modeled on the same finite alphabet, then the number of such copies is exactly three.
Results are taken from arXiv:1412.5831 .