Bayesian modeling of human–AI complementarity

Mark Steyvers, Heliodoro Tejeda, Gavin Kerrigan, Padhraic SmythProceedings of the National Academy of Sciences 2022 · DOI

TL;DR

This paper introduces a Bayesian framework to model how humans and AI can work together more effectively than either can alone. By analyzing how their predictions and confidence scores correlate, they found that human-AI complementarity is maximized when their errors are different and their confidence scores are incorporated, even if their individual accuracies differ. This work provides a theoretical and empirical foundation for building better human-AI hybrid systems.

What Problem Does This Paper Address?

As AI systems become more prevalent in real-world applications, there's a growing need to develop hybrid systems where humans and AI collaborate. The core challenge is understanding when and how such collaborations lead to better outcomes than either humans or AI working independently. While previous research has explored combining multiple AI models or multiple human experts, there's less understanding of the factors that drive effective human-AI partnerships. Specifically, this paper addresses how to combine predictions and confidence scores from humans and AI, considering their potentially different error patterns and ways of expressing confidence, to achieve "complementarity" – where the hybrid system outperforms its individual components.

Methods

The authors propose a novel Bayesian modeling framework designed to jointly model human and machine classifier predictions and confidence scores. Key aspects of their approach include:

  • Latent Correlation Modeling: The framework estimates the latent correlation between human and machine classifiers. This is crucial because high correlation can limit complementarity, even if individual accuracies differ.
  • Differentiated Confidence Modeling: It accounts for the different ways humans and machines express confidence. Machine classifiers typically provide probability distributions over all labels, while humans often provide a single confidence rating for their chosen label. The model uses an ordered probit model for human confidence.
  • Generative Process: A bivariate normal model is used to generate latent logit scores for human and machine classifiers, conditioned on the true label. This model captures dependencies between classifiers and across labels.
  • Theoretical Analysis: The paper derives a theoretical condition for complementarity based on classifier accuracies and latent correlations, providing a framework to predict when hybrid systems will outperform individual components.
  • Empirical Validation: The model is applied to a large-scale image classification dataset, allowing for empirical investigation of the theoretical findings.

Datasets

The study utilized a large-scale dataset consisting of 4,800 images across 16 classes. These images were derived from the ImageNet Large Scale Visual Recognition Challenge (ILSRVR) 2012 database and were distorted with varying levels of phase noise (ω = 80, 95, 110, 125) to create challenging classification tasks. The dataset includes:

  • Human classifications and confidence ratings: Collected from 145 participants, with at least six human classifications per image.
  • Machine classifier predictions: Generated by five well-known Convolutional Neural Network (CNN) architectures (AlexNet, DenseNet161, GoogleNet, ResNet152, VGG-19), with variants fine-tuned to different degrees of image noise.

Key Insights and Ideas

  • Complementarity is about differences, not just accuracy: Even if one component (human or AI) is less accurate, complementarity can still be achieved if their errors are distinct and their confidence scores are leveraged.
  • Correlation is the enemy of complementarity: High correlation between human and AI predictions significantly limits the potential for hybrid systems to outperform individual components. Independence is key.
  • Human confidence matters: Eliciting and incorporating human confidence ratings significantly improves hybrid performance, acting as a valuable signal for resolving disagreements.
  • Machine confidence is powerful: Machine confidence scores, especially when available across all labels, provide a substantial boost to hybrid performance, likely due to their richer information content compared to single human ratings.
  • Error differentiation is beneficial: Accounting for class-specific errors made by humans and machines can further enhance hybrid performance.

Key Results

  • Hybrid systems outperform: Hybrid human-machine (HM) pairs consistently achieved higher accuracy than non-hybrid human-human (HH) or machine-machine (MM) pairs, especially under high image noise conditions.
  • Lower HM correlations: Latent correlations between human and machine classifiers were found to be significantly lower (mean ~0.4) than between two humans (~0.7) or two machines (~0.65-0.75), supporting the idea of independent error sources.
  • Accuracy difference bounds: Theoretical analysis and empirical results showed that complementarity is achieved within a relatively narrow band of accuracy differences between human and machine classifiers. This band widens significantly if the human-AI correlation is reduced.
  • Factors improving performance: Including an error model, human confidence, and machine confidence scores all independently improved hybrid HM performance. Machine confidence had the largest impact.
  • Example: For high-noise conditions (110 noise level), a hybrid system with all factors included achieved an accuracy of 0.850, compared to 0.830 with only machine confidence, 0.844 with only human confidence, and 0.823 with only an error model.

Limitations and Future Work

  • Generalizability of Confidence: The paper notes that human confidence is often not elicited, and the current model assumes a simplified ordinal rating. Future work could explore more nuanced human confidence elicitation.
  • Individual Differences: The current model assumes a single set of parameters for all human classifiers. Extending the framework to account for individual differences in human parameters is a potential future direction.
  • Computational Cost: While MCMC was used for inference, exploring more efficient inference methods could be beneficial for larger-scale applications.
  • Broader Domains: The study focused on image classification. Applying the framework to other domains (e.g., medical diagnosis, natural language processing) would be valuable.
  • Dynamic Teaming: Future research could explore how to dynamically adjust the combination strategy based on the evolving performance and confidence of human and AI partners.

Summary updated

Abstract

SignificanceWith the increase in artificial intelligence in real-world applications, there is interest in building hybrid systems that take both human and machine predictions into account. Previous work has shown the benefits of separately combining the predictions of diverse machine classifiers or groups of people. Using a Bayesian modeling framework, we extend these results by systematically investigating the factors that influence the performance of hybrid combinations of human and machine classifiers while taking into account the unique ways human and algorithmic confidence is expressed.
BibTeX
@inproceedings{steyvers2022bayesian,
  author = {Mark Steyvers and Heliodoro Tejeda and Gavin Kerrigan and Padhraic Smyth},
  title = {{Bayesian modeling of human–AI complementarity}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2022},
  doi = {10.1073/pnas.2111547119},
}
Figure