Noisy Labels in Supervised Machine Learning: A Case Study

This thesis focuses on the aspect of label noise for real-life datasets. Due to the upcoming growing demand for annotated data in the field of supervised learning, concepts such as crowdsourcing and automatic-tagging are popular to obtain labeled data with limited costs. The straight-forward approach of tagging algorithms and missing expertise of inexpensive human annotators imply a deficit regarding label quality. The main part of this work is dedicated to the investigation of methods especially designed to take potentially corrupted samples into account. More precisely, I cover three countermeasures against label noise (noise-cleansing, loss-correction and curriculum learning) where each is based on another concept. For an exemplary case study I choose a noisy dataset provided by the organizers of DCASE, a community for “Research on Detection and Classification of Acoustic Scenes and Events”. The aim of this thesis is to evaluate label noise aware approaches in theory before their practical relevance is demonstrated in a case study on Audio Scene Classification. A detailed analysis regarding origin, characteristics and the labeling process of the dataset helps to justify the performance and applicability of all algorithms and select an appropriate method for similar problems. Final experiments show that the evaluated countermeasures outperform a standard baseline system. ; submitted by Alexander Moser, BSc ; Universität Linz, Masterarbeit, 2019 ; (VLID)4382120