Keynote: Training with imperfect and weak labels
Raúl Santos‐Rodríguez · 2021
The talk will focus on the task of training machine learning models with data with imperfect and weak labelling in scenarios where annotations consist of subsets of categories that may contain the true class and possibly several noisy classes. First, he will describe a general approach to such problems to then dive into specific applications, including learning from crowds or learning from label proportions. Finally, he will discuss strategies to measure label quality and how to learn in situations that involve combining annotations of different quality levels.