Title: Consequences and assessment of label noise
Authors: Benoit Frenay - University of Namur (Belgium) [presenting]
Abstract: When label noise pollutes a dataset, most machine learning algorithms will be affected. The consequences of label noise are diverse and well documented in the literature. They are reviewed in details, including changes in classification performances, learning requirements, complexity of learned models, observed frequencies and feature relevance. Then, we will show how a simple, generic probabilistic framework can be used to mitigate the impact of label noise in tasks such as classification, segmentation and feature selection. In practice, experimental validation of label noise tolerant algorithms is not trivial. There exist only a few datasets with clearly identified label noise and most works in the literature use simple random, uniform label noise that may not be realistic. Finally, methods that deal with label noise will be assessed.