Accurate and Robust Weakly Supervised Learning with Candidate Labels

Machine learning algorithms underpin many modern technologies but typically rely on supervised learning, which assumes access to fully annotated, noise-free training data. In practice, this assumption rarely holds: real-world datasets are often noisy, incomplete, or ambiguous. In crowdsourcing, for example, experts can assign conflicting labels to the same instances. While such datasets can be manually cleaned, sanitizing data is costly. Partial-label learning (PLL) offers a principled framework to address such ambiguous data. Thereby, each training instance is associated with a set of candidate labels, of which only one is correct but unknown. The PLL framework aims at training classifiers in this setting that perform well on unseen samples. However, the absence of exact ground truth makes it particularly difficult to construct reliable and accurate classifiers. Ensuring that such models make reliable predictions thus requires robust algorithms that can deal with uncertainty, abstain from unreliable decisions, or adapt to imperfect data. This thesis addresses these challenges by developing accurate and robust partial-label learning methods. First, we introduce a method that allows models to abstain from uncertain predictions through a reject option, enabling safer deployment in human-in-the-loop scenarios such as medical diagnosis or crowdsourcing. Second, we present a probabilistic approach that models predictive uncertainty via a Dirichlet distribution over class probabilities, enhancing robustness against label noise, out-of-distribution samples, and adversarial perturbations. Third, we propose a refinement technique based on conformal prediction that iteratively cleans candidate label sets, providing confidence guarantees while improving overall accuracy. Finally, we formulate label disambiguation as a latent-variable inference problem and employ amortized variational inference to approximate the posterior distribution over true labels, achieving state-of-the-art accuracy in both synthetic and real-world experiments. Together, these contributions establish a comprehensive framework for accurate and robust weakly supervised learning with PLL candidate labels.