Efficient Production of Training Data for Classification Supervised Learning

Speaker :	Elie de Panafieu
	Nokia Bell Labs France
Date:	01/04/2019
Time:	10:30 am - 12:00 pm
Location:	LINCS Meeting Room 26

Abstract

Supervised learning relies on the existence of training data, typically labeled by human experts. The size and quality of this data has a critical impact on the quality of the learned inferred function. But human experts are expensive. We consider the particular case of a classification problem. At each question, an expert is given two elements and tells whether they belong to the same class, with a small probability of error. We propose a scheme ensuring the quality of the final classification while minimizing the number of questions asked to the experts.