Uncertainty Estimation by Human Perception versus Neural Models

2025-06-18

AI Summary

Neural networks often exhibit overconfidence and poor calibration, limiting their deployment in safety-critical applications requiring reliable uncertainty estimation. This work presents the first systematic evaluation of how well mainstream uncertainty estimation methods align with human perceptual uncertainty, assessed across three visual benchmarks featuring human disagreement and confidence annotations. We find consistently weak correlation (significantly low Spearman/Kendall coefficients) between model-derived and human-elicited uncertainty. To bridge this gap, we propose a human-aligned uncertainty training paradigm: multi-task learning supervised by crowdsourced confidence scores as soft labels. Our approach improves model calibration without compromising Top-1 accuracy. Experiments demonstrate an average 32% reduction in Expected Calibration Error (ECE), with robust performance across diverse tasks and evaluation metrics. This work establishes a novel pathway toward building trustworthy AI systems whose uncertainty estimates better reflect human intuition.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Modern neural networks (NNs) often achieve high predictive accuracy but remain poorly calibrated, producing overconfident predictions even when wrong. This miscalibration poses serious challenges in applications where reliable uncertainty estimates are critical. In this work, we investigate how human perceptual uncertainty compares to uncertainty estimated by NNs. Using three vision benchmarks annotated with both human disagreement and crowdsourced confidence, we assess the correlation between model-predicted uncertainty and human-perceived uncertainty. Our results show that current methods only weakly align with human intuition, with correlations varying significantly across tasks and uncertainty metrics. Notably, we find that incorporating human-derived soft labels into the training process can improve calibration without compromising accuracy. These findings reveal a persistent gap between model and human uncertainty and highlight the potential of leveraging human insights to guide the development of more trustworthy AI systems.

Problem

• Neural networks produce overconfident, poorly calibrated predictions
• Human perceptual uncertainty weakly aligns with model uncertainty
• Incorporating human insights improves model calibration without losing accuracy

Innovation

• Compare human and neural network uncertainty estimates
• Use human disagreement for model calibration
• Improve AI trust with human-derived soft labels