Soft-Label Training Preserves Epistemic Uncertainty

2025-11-17

AI Summary

In subjective tasks, aggregating crowd annotations into hard labels obscures inherent human judgment diversity, causing a misalignment between model confidence and true cognitive uncertainty. To address this, we propose modeling the annotation distribution itself as a direct representation of cognitive uncertainty—not as noise to be suppressed. Our method employs soft-label supervision, using the empirical probability distribution over crowd annotations as the learning target for end-to-end training. Crucially, we are the first to formally reinterpret the annotation distribution as a learnable signal of cognitive uncertainty, thereby aligning model confidence with observed human perceptual variability. Experiments across vision and NLP benchmarks demonstrate that, while preserving hard-label accuracy, our approach reduces KL divergence by 32% and improves the correlation between model output entropy and annotation entropy by 61%.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Humans and AI: Crowd Sourcing and Human Computation

Application Category

• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Many machine learning tasks involve inherent subjectivity, where annotators naturally provide varied labels. Standard practice collapses these label distributions into single labels, aggregating diverse human judgments into point estimates. We argue that this approach is epistemically misaligned for ambiguous data--the annotation distribution itself should be regarded as the ground truth. Training on collapsed single labels forces models to express false confidence on fundamentally ambiguous cases, creating a misalignment between model certainty and the diversity of human perception. We demonstrate empirically that soft-label training, which treats annotation distributions as ground truth, preserves epistemic uncertainty. Across both vision and NLP tasks, soft-label training achieves 32% lower KL divergence from human annotations and 61% stronger correlation between model and annotation entropy, while matching the accuracy of hard-label training. Our work repositions annotation distributions from noisy signals to be aggregated away, to faithful representations of epistemic uncertainty that models should learn to reproduce.

Problem

• Addresses subjectivity in machine learning annotation distributions
• Challenges collapsing diverse labels into false confidence models
• Preserves epistemic uncertainty through soft-label training approach

Innovation

• Soft-label training preserves epistemic uncertainty
• Treats annotation distributions as ground truth
• Matches accuracy while better aligning with human perception