Learning to Complement Humans

Bryan Wilder, Eric Horvitz, Ece Kamar

Abstract

A rising vision for AI in the open world centers on the development of systems that can complement humans for perceptual, diagnostic, and reasoning tasks. To date, systems aimed at complementing the skills of people have employed models trained to be as accurate as possible in isolation. We demonstrate how an end-to-end learning strategy can be harnessed to optimize the combined performance of human-machine teams by considering the distinct abilities of people and machines. The goal is to focus machine learning on problem instances that are difficult for humans, while recognizing instances that are difficult for the machine and seeking human input on them. We demonstrate in two real-world domains (scientific discovery and medical diagnosis) that human-machine teams built via these methods outperform the individual performance of machines and people. We then analyze conditions under which this complementarity is strongest, and which training methods amplify it. Taken together, our work provides the first systematic investigation of how machine learning systems can be trained to complement human reasoning.

Introduction

Most machine learning models are developed in a vacuum, with engineers aiming for peak accuracy on static datasets under the assumption that the model will be the sole decision-maker. However, in high-stakes fields like medicine or law, AI typically functions as a partner to human experts.

The authors argue that "optimizing for isolation" misses a critical opportunity for synergy. Humans and machines possess distinct strengths and systematic "blind spots." If a model is trained without awareness of its human partner, it may waste representational capacity on cases a human finds trivial, while failing on complex cases where it could provide unique value. The objective is complementarity: a system where the machine handles its areas of expertise and intelligently defers to human experts for the rest.

Methodology

The paper introduces two primary frameworks for building these complementary systems, contrasting "fixed" approaches with "joint" optimization.

Conceptual framework of discriminative and decision-theoretic approaches for human-AI collaboration.

Discriminative Approaches

These models learn a direct mapping from input features to a decision: either provide a prediction or query a human.

  • Joint Training: Instead of training a model and then building a query policy post-hoc, the authors use a differentiable surrogate loss function. This allows the model and the query policy to be trained together via backpropagation. The model "learns" which instances will be handled by the human and can focus its learning efforts on the remaining cases.

Decision-Theoretic Approaches

These are based on the Value of Information (VOI), where the system calculates the expected utility of making a decision immediately versus paying a "cost" to obtain a human's opinion.

  • Joint VOI: Traditional VOI uses fixed, pre-trained models. The authors introduce Joint VOI, which employs a differentiable surrogate for the VOI calculation. This allows the underlying probabilistic models—predicting the label, the human's response, and the final fused outcome—to be fine-tuned end-to-end to maximize total team utility.

Experimental Evaluation

The methods were evaluated on two real-world domains:

  • Galaxy Zoo (Scientific Discovery): Classifying galaxy images into five categories. Human data came from volunteer labels, with ground truth established by consensus.
  • CAMELYON16 (Medical Diagnosis): Detecting breast cancer metastasis in lymph node images. Human responses were sampled from pathologists under time pressure, while ground truth was established by an expert with unlimited time.

Key Insights and Results

The core finding is that joint training allows for a "specialization" effect. When trained jointly, the ML model essentially "gives up" on learning regions where the human is highly competent, allowing it to use its parameters to become significantly more accurate on the "hard" cases humans typically miss.

Error rates and fraction of queries by class, illustrating how joint training targets human errors and specializes the model.

Major Findings:

  • Capacity and Complementarity: The benefits of joint training are most pronounced when the ML model has limited capacity. A smaller model that knows its limits and defers to a human is often more effective than a larger model attempting to master every case.
  • Human-Error Targeting: In the medical task, the joint model successfully identified regions where humans were responsible for 68% of their total errors. By prioritizing machine-led decisions in these regions, the team reached near-perfect accuracy.
  • Performance Gains: In Galaxy Zoo, joint training led to a 21–73% reduction in loss compared to fixed baselines for simpler models. In CAMELYON16, joint training improved discriminative models by up to 20% and VOI models by up to 10%.

Conclusion and Future Work

This work demonstrates that training AI to be a "team player" significantly outperforms models trained in isolation. However, some limitations remain. The study uses static models of human behavior; in real-world deployments, human performance might "co-evolve" as experts learn to trust or distrust the AI. Additionally, the current framework focuses on machine-initiated queries. Future research could explore more symmetric, interleaved interactions where humans also have the initiative to query the machine or provide intermediate reasoning.

Summary updated

Abstract

A rising vision for AI in the open world centers on the development of systems that can complement humans for perceptual, diagnostic, and reasoning tasks. To date, systems aimed at complementing the skills of people have employed models trained to be as accurate as possible in isolation. We demonstrate how an end-to-end learning strategy can be harnessed to optimize the combined performance of human-machine teams by considering the distinct abilities of people and machines. The goal is to focus machine learning on problem instances that are difficult for humans, while recognizing instances that are difficult for the machine and seeking human input on them. We demonstrate in two real-world domains (scientific discovery and medical diagnosis) that human-machine teams built via these methods outperform the individual performance of machines and people. We then analyze conditions under which this complementarity is strongest, and which training methods amplify it. Taken together, our work provides the first systematic investigation of how machine learning systems can be trained to complement human reasoning.
BibTeX
@inproceedings{wilder2020learning,
  author = {Bryan Wilder and Eric Horvitz and Ece Kamar},
  title = {{Learning to Complement Humans}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/212},
}
Figure