Learning to Defer with Limited Expert Predictions

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

Abstract

Recent advancements in human-AI collaboration demonstrate that hybrid systems can outperform both AI models and human experts in isolation. A key mechanism for this collaboration is "learning to defer" (L2D), where an AI model learns to delegate tasks to a human expert when its own confidence is low or the expert's proficiency is high. However, traditional L2D algorithms require large volumes of expert predictions to model human behavior accurately—a requirement that is often impractical in high-stakes domains like healthcare. This paper proposes a three-step framework that utilizes semi-supervised learning to train L2D systems with a fraction of the usual expert data. Our results on synthetic and real-world medical datasets show that this approach maintains high performance even with minimal expert input, facilitating the deployment of human-AI teams in data-scarce environments.

Introduction: The Expert Data Bottleneck

The "Learning to Defer" (L2D) paradigm is designed to optimize the synergy between human intuition and algorithmic precision. In fields such as radiology, an AI might excel at identifying common patterns but struggle with rare pathologies that a human specialist would recognize immediately. An L2D system learns to recognize these boundaries, handing the "steering wheel" to the human when necessary.

Despite its potential, L2D faces a significant hurdle: the "expert data bottleneck." Training these systems typically requires thousands of labels from a specific expert to map their unique strengths and weaknesses. In practice, experts are often unavailable for such extensive labeling tasks, or the experts themselves may change frequently due to shifts or rotations. This paper addresses this challenge by reducing the reliance on large-scale expert datasets.

Methodology: A Three-Step Framework

The authors propose a pipeline to approximate expert behavior using limited data, centered on the generation of "artificial expert predictions."

1. Embedding Model

The first step involves training a standard feature extractor (e.g., ResNet or EfficientNet) using available ground truth labels. This model creates a high-dimensional embedding space where data points with similar features are clustered together. This "map" serves as the foundation for understanding where an expert is likely to succeed or fail.

2. Expertise Predictor

Instead of attempting to predict the exact class an expert might choose, the framework simplifies the task into a binary classification problem: Will the expert be correct or incorrect on this specific instance?

To train this predictor with minimal data, the authors employ Semi-Supervised Learning (SSL) techniques, specifically FixMatch and CoMatch. By using a small set of expert-labeled instances alongside a large volume of unlabeled data, the expertise predictor learns to identify the "competence regions" of the expert within the feature space.

3. Artificial Prediction Generation

The expertise predictor is then used to generate "artificial" predictions for the remainder of the training set. If the predictor indicates the expert would be correct for a given instance, the system assigns the ground truth label as the "expert's" prediction. If it predicts an error, it assigns a random incorrect label. These synthesized predictions are then used to train standard L2D algorithms as if a full expert dataset were available.

Experimental Evaluation

The framework was evaluated on two distinct datasets to test its versatility:

  • CIFAR-100: To simulate controlled environments, the researchers created "synthetic experts" with programmed biases (e.g., high proficiency in identifying "aquatic mammals" but poor performance on "vehicles").
  • NIH Chest X-ray Dataset: This real-world medical dataset provided a more rigorous test, featuring 4,374 images with predictions from individual radiologists and "gold standard" labels from a consensus panel. This allowed the model to be tested against the complex, non-linear error patterns of human specialists.

Key Insights and Results

The proposed "Embedding-SSL" approach demonstrated remarkable data efficiency and performance:

  • High Efficiency: On the CIFAR-100 dataset, the system achieved 99% of the performance of a model trained on a full expert dataset using only six expert predictions per class.
  • Superior Teamwork: In both synthetic and real-world scenarios, the Human-AI team consistently outperformed both the AI and the human expert working individually.
  • Medical Performance: On the NIH dataset, the framework outperformed standard baselines by over 38% in low-data regimes (fewer than 100 expert labels).
  • Fairness and Bias: A critical concern in medical AI is the amplification of demographic biases. The authors found no evidence that generating artificial predictions increased biases related to gender or age compared to models trained on full human datasets.

Discussion and Future Work

The success of this framework hinges on the "binary shortcut"—the realization that modeling an expert's competence is more data-efficient than modeling their specific classification choices. Furthermore, the reliance on the feature space suggests that if two images appear similar to an AI, a human's likelihood of success on those images is often correlated.

While the results are promising, several avenues for future research remain:

  • Active Learning: Currently, expert labels are sampled randomly. Integrating active learning could allow the AI to strategically request labels for instances where it is most uncertain about the expert's performance, further increasing efficiency.
  • Multi-Expert Dynamics: Future iterations could explore how to aggregate predictions from multiple experts with overlapping specialties.
  • Error Granularity: The binary "correct/incorrect" assumption may need refinement for tasks where specific types of errors (e.g., false negatives in cancer screening) carry significantly higher costs.

Conclusion

By bridging the gap between limited expert availability and the data requirements of L2D algorithms, this three-step framework makes human-AI collaboration more accessible. It provides a path forward for deploying sophisticated deferral systems in specialized fields where expert time is the most valuable resource.

Reference: Hemmer, P., Thede, L., Vossing, M., Jakubik, J., & Kühl, N. (2023). Learning to Defer with Limited Expert Predictions. Proceedings of the AAAI Conference on Artificial Intelligence.

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