Learning with Imperfect Labels and Incomplete Views: A Review of Representation Methods for Weakly Supervised Perception
Alicia Green · 2025
Weakly supervised perception addresses the challenges of training representation models when supervision signals are incomplete, inexact, or inaccurate, and when multi-view or multi-modal data suffer from missing observations. This review consolidates academic advances in learning from noisy labels, coarse labels, and limited labeled data, as well as methods for incompleteview and incomplete-modality learning. Techniques include noise-robust loss functions, sample selection, multiple instance learning, constraint-based optimization, semi-supervised and self-supervised representation learning, shared latent subspace models, cross-view alignment, and federated feature integration. Representative domains include computer vision, 3D mapping, network traffic analysis, and evaluation of large language model social intelligence. The survey emphasizes observed empirical outcomes and established algorithms, aiming to provide a unified view of strategies that enable robust feature learning under weak supervision in perception tasks.