Seminar: Conformal prediction and testing exchangeability on-line

SpeakerVladimir Vovk
AffiliationRoyal Holloway, University of London
DateThursday, 08 Mar 2012
Time13:00 - 14:00
LocationRoom 1.02, Malet Place Engineering Building
Event seriesCSML Joint Seminar Series

In the first part of this talk I will give a survey of the method of conformal prediction. Whereas standard machine learning algorithms produce point predictions, conformal predictors produce prediction sets that are valid in the sense of failing to contain the true label with a prespecified error probability. The validity of conformal predictors depends only on the assumption that the observed examples are generated independently from the same probability measure, which will be called the exchangeability assumption. Validity being achieved automatically, the quality of conformal predictors is determined by their other characteristics, such as efficiency (small size of prediction sets). Efficient conformal predictors can be designed on top of many standard machine-learning algorithms, such as nearest neighbours, support vector machines, neural networks, etc. In the second part of the talk I will discuss using the method of conformal prediction for testing exchangeability in the on-line mode. This is achieved by constructing "exchangeability martingales": gambling systems that are fair under the exchangeability assumption.

iCalendar csml_id_15.ics