Human Uncertainty Inference via Deterministic Ensemble Neural Networks

Yujin Cha, Sang Wan Lee

Introduction

Presented at AAAI-21, this research explores the intersection of human psychology and machine learning by introducing a method to "read" human uncertainty using deterministic neural networks. The proposed Proxy Ensemble Network (PEN) aims to bridge the gap between the interpretable but "poorly accessible" nature of human doubt and the mathematically accessible but often uninterpretable nature of machine learning (ML) uncertainty.

The Challenge of Measuring Human Doubt

In human-in-the-loop systems, such as AI-assisted medical diagnosis, understanding when a human expert is unsure is critical. While humans can explain why they are doubtful, measuring this uncertainty accurately typically requires exhaustive behavioral sampling—asking a person to perform the same task numerous times—which is impractical in clinical or educational settings. This paper addresses a fundamental question: Can we leverage the accessible nature of ML models to infer the hard-to-reach internal uncertainty of a human expert?

Methodology: The Proxy Ensemble Network (PEN)

The researchers model human decision-making as a Bayesian process. They assume that when a human evaluates an image, their internal cognitive "weights" fluctuate, leading to varying levels of confidence or even different answers upon repeated exposure to the same stimulus.

To capture this without the computational overhead of a full Bayesian Neural Network (BNN), the authors developed the Proxy Ensemble Network (PEN). The PEN consists of an ensemble of simple, deterministic neural networks. Rather than being trained to find a "ground truth" label, these proxy networks are trained to predict the distribution of answers a human (modeled as a BNN) would provide.

Human behavioral experiments and the Subject module used to gather baseline uncertainty data.

The key innovation lies in using the disagreement between these deterministic ensemble members to map the "entropy range"—the bounds of doubt—of the human expert.

Uncertainty Decomposition

A core feature of the PEN is its ability to decompose uncertainty into two distinct categories:

  1. Aleatoric Uncertainty: Arising from noise inherent in the data (e.g., a blurry or low-quality X-ray).
  2. Epistemic Uncertainty: Arising from a lack of knowledge or experience (e.g., a rare medical condition the expert has seldom encountered).

The researchers discovered that human epistemic uncertainty is strongly associated with "response mismatch"—the tendency to give different answers for the same image over time. In contrast, aleatoric uncertainty lowers overall confidence but is less likely to cause inconsistent choices.

Relationship between uncertainty types and the ratio of same classification (consistency) across multiple trials.

Experimental Validation

The PEN model was validated using both simulated agents and real-world human data:

  • Benchmarks: SVHN, Fashion-MNIST, and HAM10000 (skin lesions).
  • Medical Imaging: Two chest X-ray datasets (CXR-A and CXR-B) focusing on abnormality detection and specific diagnoses.
  • Human Subjects: 64 licensed physicians (M.D.s) participated in behavioral experiments, classifying 50 images four times each to establish a "human ground truth."

Key Results

The PEN demonstrated remarkable efficiency and accuracy:

  • Predictive Accuracy: The model achieved a Predictive Uncertainty Inference Accuracy (PUIA) of 95.7% on the SVHN dataset and between 75-80% on complex medical tasks.
  • Diagnosis Prediction: The system predicted actual physician diagnoses with 92.8% accuracy for the CXR-A dataset.
  • Sample Efficiency: The model learned human uncertainty patterns from as few as 220 training images in the medical task, suggesting high utility in data-scarce environments.

Conclusion and Future Work

The study proves that an ensemble of "rigid" deterministic networks can effectively approximate the "fluid" uncertainty of the human brain. This has significant implications for active learning, where AI could identify specific cases that a human needs to review to maximize learning efficiency. However, the authors emphasize ethical considerations, noting that such systems must be implemented carefully to avoid reinforcing human biases or introducing security risks in clinical environments.

Summary updated

Abstract

The estimation and inference of human predictive uncertainty have great potential to improve the sampling efficiency and prediction reliability of human-in-the-loop systems for smart healthcare, smart education, and human-computer interactions. Predictive uncertainty in humans is highly interpretable, but its measurement is poorly accessible. Contrarily, the predictive uncertainty of machine learning models, albeit with poor interpretability, is relatively easily accessible. Here, we demonstrate that the poor accessibility of human uncertainty can be resolved by exploiting simple and universally accessible deterministic neural networks. We propose a new model for human uncertainty inference, called proxy ensemble network (PEN). Simulations with a few benchmark datasets demonstrated that the model can efficiently learn human uncertainty from a small amount of data. To show its applicability in real-world problems, we performed behavioral experiments, in which 64 physicians classified medical images and reported their level of confidence. We showed that the PEN could predict both the uncertainty range and diagnoses given by subjects with high accuracy. Our results demonstrate the ability of machine learning in guiding human decision making; it can also help humans in learning more efficiently and accurately. To the best of our knowledge, this is the first study that explored the possibility of accessing human uncertainty via the lens of deterministic neural networks.
BibTeX
@inproceedings{cha2021human,
  author = {Yujin Cha and Sang Wan Lee},
  title = {{Human Uncertainty Inference via Deterministic Ensemble Neural Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i7.16735},
}
Figure