Amortized Variational Inference for Partial-Label Learning: A Probabilistic Approach to Label Disambiguation

Real-world data is frequently noisy and ambiguous. In crowdsourcing, for example, human annotators may assign conflicting class labels to the same instances. Partial-label learning (PLL) addresses this challenge by training classifiers when each instance is associated with a set of candidate labels, only one of which is correct. While early PLL methods approximate the true label posterior, they are often computationally intensive. Recent deep learning approaches improve scalability but rely on surrogate losses and heuristic label refinement. We introduce a novel probabilistic framework that directly approximates the posterior distribution over true labels using amortized variational inference. Our method employs neural networks to predict variational parameters from input data, enabling efficient inference. This approach combines the expressiveness of deep learning with the rigor of probabilistic modeling, while remaining architecture-agnostic. Theoretical analysis and extensive experiments on synthetic and real-world datasets demonstrate that our method achieves state-of-the-art performance in both accuracy and efficiency.

Many machine-learning systems learn from examples that have labels, but in the real world those labels are often uncertain or disputed. For instance, when several people label the same image or document, they may not agree, leaving the system with a small set of possible answers rather than one trusted answer. We study how to train models in this setting, where each example comes with several candidate labels and only one is correct. Our method treats the unknown true label as something the model should learn, and uses a neural network to estimate which candidate labels are most likely. This combines the flexibility of modern deep learning with a principled way to reason about uncertainty. The approach can work with different types of neural networks rather than being tied to one specific design. In experiments on both controlled and real-world data, it produced accurate predictions while still being easy to compute. This can make machine learning more reliable when labels are noisy, ambiguous, or expensive to verify.