Uncertainty-aware multi-instance partial-label learning via evidential deep model

doi:10.1016/j.neucom.2026.134578

Uncertainty-aware multi-instance partial-label learning via evidential deep model

2026-07-24

0

PRE

AI

DOI：10.1016/j.neucom.2026.134578

Original

Original request for help

Save

Abstract

En 中文

Multi-Instance Partial-Label Learning (MIPL) is a challenging weakly supervised learning paradigm in which instance-level ambiguity and label-level ambiguity coexist. Existing attention-based MIPL methods typically identify key instances using deterministic weights, but their inability to quantify uncertainty makes them vulnerable to noisy or ambiguous instances. To address this limitation, we propose the Uncertainty-Aware Multi-Instance Partial-Label Learning (UAMIPL) framework, which explicitly incorporates uncertainty into the instance aggregation process. Specifically, UAMIPL jointly captures (1) global uncertainty through Bayesian modeling of the instance-scoring parameters and (2) local uncertainty through an evidential uncertainty estimation mechanism at the instance level. The resulting reliability signal is used to calibrate base attention scores, yielding uncertainty-aware aggregation coefficients for more robust bag representation learning. Extensive experiments on benchmark and real-world datasets validate the effectiveness of UAMIPL, particularly under substantial noise and ambiguity. Further analyses show that the proposed uncertainty-aware aggregation mechanism improves robustness, enhances interpretability at the instance level, and benefits from the complementary effects of global and local uncertainty modeling.

Uncertainty-aware multi-instance partial-label learning via evidential deep model

Abstract

Journal

Organization