TL;DR
This paper introduces a novel approach for building more effective human-AI teams by jointly training a classifier and an allocation system. This system learns to delegate tasks to either the AI classifier or the most suitable human expert, effectively complementing individual human weaknesses and improving overall team performance.
What Problem Does This Paper Address?
As AI models become increasingly capable, they are often deployed alongside human experts. While AI can match or even surpass human performance in many areas, there are still specific instances where human intuition or nuanced understanding is superior. Prior work has explored scenarios where an AI defers difficult cases to a single human expert. However, real-world applications often involve multiple human experts with varying specializations and capabilities. This paper addresses the challenge of effectively leveraging a team of multiple human experts and an AI classifier, going beyond simply assigning tasks to the "best" expert and instead focusing on how the AI can specifically complement the collective weaknesses of the human team.
Methods
The core of the proposed approach is a joint training framework for two components:
- A Classifier (F): This component is trained to specifically learn to predict instances where human experts are less accurate.
- An Allocation System (A): This system acts as a "gating network" that learns to assign each instance to the most competent team member. This team member can be either the AI classifier or one of the available human experts.
The system is trained end-to-end by minimizing a cross-entropy loss function that considers the weighted predictions of all team members. The allocation system outputs probabilities for each team member, and these probabilities are used to combine their predictions into a final team decision. This approach is inspired by the mixture-of-experts (MoE) framework but adapted to combine a single AI classifier with multiple human experts.
Datasets
The paper evaluates its approach on three datasets:
- Twitter Hate Speech Dataset: A dataset of tweets used to create synthetic human experts based on dialect labels (African-American English vs. non-African-American English). This allowed for controlled experiments on team size and expert diversity.
- CIFAR-100 Dataset: A standard image classification dataset used to create synthetic human experts with varying levels of expertise across different subclasses. This dataset was used to investigate the impact of expert diversity on team performance.
- NIH Chest X-ray Dataset: A real-world dataset of chest X-ray images with annotations from multiple radiologists. This dataset was used to demonstrate the practical applicability of the approach in a high-stakes domain.
Key Insights and Ideas
- Complementarity is Key: The most significant insight is that the AI classifier doesn't just aim to be generally accurate; it specifically learns to cover the blind spots of the human experts. This targeted complementarity is crucial for maximizing team performance.
- Smart Delegation: The allocation system is not just a simple routing mechanism. It learns to dynamically assign tasks based on the predicted competence of each team member for a given instance, ensuring that the most appropriate decision-maker is chosen.
- Diversity Benefits: The approach shows that increasing the diversity of human experts' capabilities leads to better team performance. This is because a more diverse team covers a wider range of expertise, allowing the AI classifier to focus on an even smaller, more specialized set of difficult instances.
- Scalability: The method demonstrates effectiveness with varying team sizes, suggesting it can adapt to different resource availabilities.
Key Results
- Superior Performance: The proposed "Classifier & Expert Team" approach consistently outperformed various baselines, including the Joint Sparse Framework (JSF), a single classifier, random expert assignment, and the best single expert.
- Hate Speech Detection: On the Twitter dataset, the approach achieved significant performance gains over baselines, with an average improvement of 3.66% over JSF and 7.44% over a Classifier Team. The classifier's contribution was evident in its ability to accurately predict instances where human experts were less accurate.
- CIFAR-100: Similar improvements were observed on CIFAR-100, with gains of up to 25.86% over a One Classifier baseline. The experiments highlighted the positive impact of expert diversity, with the approach showing a 2.40% performance gain compared to an Expert Team alone.
- Medical Domain: On the NIH Chest X-ray dataset, the approach also outperformed baselines, demonstrating its real-world applicability. Table 1 shows team accuracies reaching up to 95.45% (±0.06) for the proposed method, outperforming other baselines.
Limitations and Future Work
- Data Requirements: The approach assumes access to human experts' predictions for all training data, which can be a significant manual labeling effort. Future work could explore methods that leverage data not annotated by all team members.
- Expert Availability: The current framework assumes experts are always available. Future research could investigate load balancing strategies or alternative assignment strategies when the most accurate team member is unavailable.
- Ethical Considerations: While the goal is to improve performance, the ability to estimate individual expert contributions raises ethical questions. The authors suggest this could be used for targeted training to improve individual experts rather than simply excluding them.
Summary updated