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.
Introduction
In high-stakes domains such as medical diagnosis or autonomous driving, "Learning to Defer" (L2D) systems provide a safety net by allowing an AI to hand over difficult cases to a human expert. However, traditional L2D systems are brittle; they are typically optimized for a specific expert or a fixed group seen during training. If the human personnel changes—such as a new radiologist with a different specialty joining a hospital—the AI's deferral policy becomes obsolete. Retraining these systems for every new expert is computationally prohibitive.
To address this, we introduce L2D-Pop, a meta-learning framework that treats experts as random variables drawn from a population. By observing a small "context set" of an expert's past decisions, L2D-Pop adapts its deferral strategy on-the-fly to accommodate never-before-seen individuals.
Methodology
L2D-Pop employs two primary meta-learning strategies to handle expert variability:
- Optimization-based (Fine-tuning): The system maintains a "marginal" model representing the average expert. Upon encountering a new expert, the system performs a few gradient descent steps using the context set to specialize the policy.
- Model-based (Neural Processes): This approach uses a Deep Sets encoder to ingest the context set and generate a representation of the expert's specific strengths and weaknesses.
Attentive Neural Processes
A key innovation in our model-based approach is the use of an Attentive Neural Process. This architecture utilizes a cross-attention mechanism to compare a new test point against the expert's historical performance in the context set. If the expert has demonstrated high accuracy on similar cases previously, the system is more likely to defer.
Experimental Evaluation
We evaluated L2D-Pop across several benchmarks:
- Image Recognition: CIFAR-10 and CIFAR-100.
- Traffic Sign Detection: German Traffic Sign Recognition Benchmark (GTSRB).
- Medical Diagnosis: HAM10000 skin lesion dataset.
To simulate a diverse population, we constructed synthetic experts with varying "oracle" classes (classes they identify perfectly) and varying noise levels for others.
Results
L2D-Pop consistently outperformed the "Single-L2D" baseline, which only models the average expert. The performance gains were most pronounced in highly diverse expert populations.
Key findings include:
- Efficiency: The model-based approach (Neural Process) is significantly faster at test-time as it requires only a single forward pass for adaptation.
- Local Expertise: The attention mechanism successfully identifies sub-niches of expertise, allowing for more granular deferral decisions.
- Safe Defaults: In the absence of a context set, the model defaults to its own predictions (~99% coverage) rather than making erratic deferral choices.
Conclusion
L2D-Pop represents a significant step toward flexible human-AI collaboration. By shifting the focus from expert identity to expert behavior, we enable AI systems to adapt to new personnel without retraining. Future work will explore experts whose behavior shifts over time due to fatigue or learning, as well as more complex metric-based meta-learning techniques.
Summary updated