Learning to Defer to a Population: A Meta-Learning Approach

Dharmesh Tailor, Aditya Patra, Rajeev Verma, Putra Manggala, Eric T. Nalisnick

Abstract

The Learning to Defer (L2D) framework enhances the safety and robustness of autonomous systems by selectively allocating difficult decisions to human experts. However, existing L2D systems are typically optimized for specific, well-identified experts and require retraining if the expert changes. This work introduces L2D-Pop, a meta-learning framework designed to handle never-before-seen experts at test-time. By utilizing a small context set of an expert's previous decisions, our system quickly adapts its deferral policy. We evaluate both optimization-based and model-based meta-learning variants, including an Attentive Neural Process (ANP) that leverages cross-attention to assess expert abilities locally. Experiments on image recognition, traffic sign detection, and skin lesion diagnosis demonstrate that L2D-Pop significantly outperforms traditional L2D baselines in dynamic, multi-expert environments.

Introduction

In high-stakes domains such as medical diagnostics and autonomous driving, AI systems often operate alongside humans. The "Learning to Defer" (L2D) paradigm allows a model to choose between making a prediction itself or handing the task to a human expert. While effective, current L2D systems are brittle: they are trained for a specific expert or a fixed group. If the expert changes—for instance, a night-shift radiologist replacing a morning-shift colleague—the system's deferral logic may no longer align with the new expert's specific strengths and weaknesses.

This paper addresses the challenge of generalizing to a population of experts. We propose L2D-Pop, a framework that treats expert identity as a latent variable and uses meta-learning to adapt to new individuals without requiring full retraining.

Methodology

L2D-Pop characterizes a new expert using a context set—a small collection of the expert’s past predictions. We explore two primary meta-learning strategies to utilize this context:

1. Meta-Optimization (Fine-tuning)

This approach treats the deferral policy as a model that can be quickly specialized. Starting from a set of meta-learned weights representing an "average" expert, the system performs a few steps of gradient descent using the new expert's context set. This allows the model to shift its decision boundaries to match the specific expertise of the current human collaborator.

2. Model-Based (Neural Processes)

We introduce a more efficient model-based approach using Neural Processes. This architecture uses a Deep Sets encoder to map the context set into a fixed-length latent representation in a single forward pass.

To improve precision, we employ an Attentive Neural Process (ANP). The ANP uses a cross-attention mechanism between the current test point and the context set. For example, if the AI is evaluating a specific medical image, it can "attend" to how the expert performed on similar images in the context set. This allows the system to recognize "local" expertise—identifying that an expert might be highly reliable for specific sub-classes (e.g., melanoma) even if their overall accuracy is average.

Experimental Results

We validated L2D-Pop across four benchmarks: synthetic 2D data, CIFAR-10/100, the German Traffic Sign Recognition Benchmark (GTSRB), and the HAM10000 skin lesion dataset.

Performance and Adaptability

L2D-Pop consistently outperformed single-expert baselines. In synthetic experiments, the model successfully learned to ignore "unskilled" experts (1% accuracy) while deferring difficult cases to "skilled" experts (95% accuracy). As the diversity of the expert population increased, the performance gap between L2D-Pop and traditional methods widened, highlighting the necessity of population-aware modeling.

The Importance of Context

Our findings indicate that context is vital. A context set of only 50–100 examples is sufficient to significantly boost system accuracy. Furthermore, the ANP variant proved particularly adept at identifying specialized knowledge. It could recognize when an expert was an "oracle" for specific categories, even when those categories were rare in the overall dataset.

Safety and Failure Modes

A critical feature of the Neural Process model is its safety profile. When the context set is missing or entirely unfamiliar, the model defaults to a high "coverage" (nearly 99%), meaning it relies on its own predictions rather than deferring to an unknown entity. This prevents the system from making dangerous assumptions about an uncharacterized expert.

Discussion and Conclusion

Traditional L2D systems suffer from the "marginal expert trap," where training for an average human leads to sub-optimal performance for every individual expert. L2D-Pop avoids this by treating adaptation as a first-class requirement. While the optimization-based approach is competitive, the model-based ANP offers superior speed and the ability to capture local expertise through attention.

Future work will address temporal shifts—modeling how an expert’s performance might change over time due to fatigue or learning—and testing the framework against a wider variety of real-world human decision-making data. By enabling AI to understand the population it serves, L2D-Pop moves us closer to truly collaborative and robust human-AI systems.

Summary updated

Abstract

The learning to defer (L2D) framework allows autonomous systems to be safe and robust by allocating difficult decisions to a human expert. All existing work on L2D assumes that each expert is well-identified, and if any expert were to change, the system should be re-trained. In this work, we alleviate this constraint, formulating an L2D system that can cope with never-before-seen experts at test-time. We accomplish this by using meta-learning, considering both optimization- and model-based variants. Given a small context set to characterize the currently available expert, our framework can quickly adapt its deferral policy. For the model-based approach, we employ an attention mechanism that is able to look for points in the context set that are similar to a given test point, leading to an even more precise assessment of the expert's abilities. In the experiments, we validate our methods on image recognition, traffic sign detection, and skin lesion diagnosis benchmarks.
BibTeX
@inproceedings{tailor2024learning,
  author = {Dharmesh Tailor and Aditya Patra and Rajeev Verma and Putra Manggala and Eric T. Nalisnick},
  title = {{Learning to Defer to a Population: A Meta-Learning Approach}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2024},
  url = {https://openreview.net/forum?id=lMV7LMtmO4},
}
Figure