Learning to Defer with Limited Expert Predictions

Patrick Hemmer, Lukas Thede, Michael Vössing, Johannes Jakubik, Niklas Kühl

Introduction

Learning to Defer (L2D) is a collaborative framework where an AI system and a human expert work together to make decisions. In high-stakes domains such as medical diagnostics or autonomous navigation, the system must decide whether to provide its own prediction or "defer" the task to a human expert. A significant challenge in training effective L2D models is the requirement for large datasets of expert predictions to model the expert's strengths and weaknesses. Collecting thousands of such labels is often prohibitively expensive or time-consuming. This paper addresses this "data hunger" by proposing a method to learn an expert's capabilities from a minimal sample of their work.

Methodology: A Three-Step Framework

The authors introduce a pipeline designed to "hallucinate" expert behavior across an entire dataset using only a handful of actual expert labels. This process is divided into three distinct phases:

  1. Feature Extraction (Embedding Model): A standard classification model is trained on ground truth labels to learn a rich feature representation (embeddings) of the data.
  2. Expertise Prediction (The "Expert Mimic"): A secondary model is trained to predict the binary outcome of whether a human expert will be correct or incorrect for a given input. To maximize efficiency, this step utilizes Semi-Supervised Learning (SSL) techniques like FixMatch or CoMatch, leveraging patterns in unlabeled data to generalize from a few expert examples.
  3. Artificial Label Generation: The "Expert Mimic" generates predicted human labels for the full training set. If the mimic predicts the expert would be correct, it assigns the ground truth label; otherwise, it assigns a random incorrect label.
The three-step framework for generating artificial expert predictions: 1) Training the embedding model, 2) Training the expertise predictor using SSL, and 3) Generating artificial labels.

This generated dataset allows any standard L2D algorithm to be trained as if a complete set of expert predictions were available.

Experimental Evaluation

The framework was evaluated on two datasets:

  • CIFAR-100: Used to simulate synthetic experts with specific categorical biases (e.g., high accuracy on animals, low accuracy on vehicles).
  • NIH Chest X-ray Dataset: A real-world medical dataset featuring predictions from individual radiologists compared against "gold standard" panel labels.

Key Results and Insights

The proposed "Embedding-SSL" approach demonstrates remarkable data efficiency. On the NIH dataset, the system achieved 98.83% of the performance of a model trained on the full expert dataset while using only six expert predictions per class.

System performance across various datasets showing how accuracy scales with the number of expert predictions. The proposed method consistently maintains high performance even with very few samples.

Key insights from the study include:

  • Binary Simplification: Modeling expertise as a binary "correct/incorrect" task significantly simplifies the learning problem compared to predicting specific class errors.
  • Transferability of Features: Features learned for object recognition are highly predictive of human error, particularly near decision boundaries where both AI and humans tend to struggle.
  • SSL as a Catalyst: Semi-supervised learning allows the model to map expert behavior across the entire latent space, rather than just memorizing specific points.

Conclusion and Future Directions

This research provides a viable path for deploying AI-human collaborative systems in data-scarce environments. By reducing the expert labeling requirement by orders of magnitude, L2D becomes practical for specialized medical tasks or rapidly changing expert pools. Future work may explore Active Learning to select the most informative samples for expert labeling and extending the framework to handle multiple experts simultaneously.

Summary updated

Abstract

Recent research suggests that combining AI models with a human expert can exceed the performance of either alone. The combination of their capabilities is often realized by learning to defer algorithms that enable the AI to learn to decide whether to make a prediction for a particular instance or defer it to the human expert. However, to accurately learn which instances should be deferred to the human expert, a large number of expert predictions that accurately reflect the expert's capabilities are required -- in addition to the ground truth labels needed to train the AI. This requirement shared by many learning to defer algorithms hinders their adoption in scenarios where the responsible expert regularly changes or where acquiring a sufficient number of expert predictions is costly. In this paper, we propose a three-step approach to reduce the number of expert predictions required to train learning to defer algorithms. It encompasses (1) the training of an embedding model with ground truth labels to generate feature representations that serve as a basis for (2) the training of an expertise predictor model to approximate the expert's capabilities. (3) The expertise predictor generates artificial expert predictions for instances not yet labeled by the expert, which are required by the learning to defer algorithms. We evaluate our approach on two public datasets. One with "synthetically" generated human experts and another from the medical domain containing real-world radiologists' predictions. Our experiments show that the approach allows the training of various learning to defer algorithms with a minimal number of human expert predictions. Furthermore, we demonstrate that even a small number of expert predictions per class is sufficient for these algorithms to exceed the performance the AI and the human expert can achieve individually.
BibTeX
@inproceedings{hemmer2023learning,
  author = {Patrick Hemmer and Lukas Thede and Michael Vössing and Johannes Jakubik and Niklas Kühl},
  title = {{Learning to Defer with Limited Expert Predictions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i5.25742},
}
Figure