Source-linked AI summary
Learning to Complement Humans
Bryan Wilder, Eric Horvitz, Ece Kamar
TL;DR
The paper addresses the gap between optimizing machine accuracy in isolation and optimizing human-machine teamwork. It introduces end-to-end training that jointly learns predictions and human-query policies, then finds that complementarity is especially beneficial under limited model capacity and asymmetric error costs. The resulting framework systematically trains systems to use human and machine strengths together.
Problem
Existing systems for complementing people usually optimize machine performance in isolation rather than the combined performance of human-machine teams.
Method
The paper jointly trains a predictive model and a policy that allocates instances to humans or machines, fusing judgments while accounting for human-query costs.
Results
Complementarity-focused training provides the greatest benefits when model capacity is limited or error costs are asymmetric, and joint training produces structured differences between human and machine errors.
Takeaways & Limitations
Human-machine systems can be trained to exploit differences in human and machine capabilities rather than treating machine accuracy as the sole optimization target.
Takeaways & Limitations
The analysis notes that limited data requires restricting model capacity to prevent overfitting, while sufficiently large datasets could support arbitrarily complex models.
Abstract
from arXiv · showhide
A rising vision for AI in the open world centers on the development of systems that can complement humans for perceptual, diagnostic, and reasoning tasks. To date, systems aimed at complementing the skills of people have employed models trained to be as accurate as possible in isolation. We demonstrate how an end-to-end learning strategy can be harnessed to optimize the combined performance of human-machine teams by considering the distinct abilities of people and machines. The goal is to focus machine learning on problem instances that are difficult for humans, while recognizing instances that are difficult for the machine and seeking human input on them. We demonstrate in two real-world domains (scientific discovery and medical diagnosis) that human-machine teams built via these methods outperform the individual performance of machines and people. We then analyze conditions under which this complementarity is strongest, and which training methods amplify it. Taken together, our work provides the first systematic investigation of how machine learning systems can be trained to complement human reasoning.
1 Introduction
Existing human-AI systems typically optimize machine performance in isolation, despite complementary human capabilities. This paper proposes jointly training prediction and human-query decisions for team performance across scientific discovery and medical diagnosis.
- Existing machine-learning systems generally optimize model performance alone rather than human-machine teamwork.
- The proposed systems allocate instances between humans and machines and fuse their judgments while accounting for the cost of human queries.
- The authors evaluate complementarity-focused training on galaxy classification and breast-cancer-metastasis detection.
- Complementarity training is most important when model capacity is limited, because the machine must choose which parts of the task to prioritize.
- Complementarity training provides larger benefits when errors have asymmetric costs, increasing the need to prioritize among possible mistakes.
- Joint training changes machine errors and allocations so that human and machine errors become structured and qualitatively different.
2 Problem Formulation
The paper formalizes prediction with optional human consultation as a utility-maximization problem. A machine observes each instance, decides whether to query a human at cost c, and combines the resulting information into a prediction.
- The framework extends supervised multiclass prediction by associating each instance with a human label and a machine prediction.
- For each instance, the machine decides whether to predict independently or first consult a human expert.
- The query policy qθ(x) outputs 1 when the human is queried and 0 otherwise, after the machine observes x.
- Team utility is u(y, ŷ) without consultation and u(y, ŷ) − c when consultation occurs.
- Unlike conventional supervised learning, the objective includes both the query decision and the human information’s effect on overall team accuracy.
3 Approach
The approach replaces isolated model training with joint optimization of machine predictions and human-query policies, using discriminative and decision-theoretic methods to improve team utility. These methods allocate instances according to relative human-machine strengths and train models to focus on instances handled by the machine.
- Discriminative Approaches: Joint training optimizes the ML model and query policy together for human-machine complementarity, unlike fixed approaches that train the model in isolation.The joint discriminative method trains mθ end-to-end with qθ, while the fixed method constructs q after fixing m.
- Discriminative Approaches: The joint discriminative loss uses a differentiable relaxation of team utility, combining human and machine outputs while penalizing query probability.A smoother fractional-prediction loss improves training stability and lets the machine focus on instances where the query policy is lower.
- Discriminative Approaches: When the human is queried, the simplified runtime system outputs the human response, reflecting the observation that this performs better empirically than a separate post-query model.The query decision converts fractional q values into binary runtime actions using query probability and model uncertainty.
- Decision-Theoretic Approaches: Joint value-of-information training optimizes the combined system end-to-end while retaining probabilistic models for labels, human responses, and post-query outcomes.The method uses neural networks with calibration layers, updates them jointly, and estimates query value by comparing expected utilities with and without human input.
- Decision-Theoretic Approaches: The resulting system fits carefully to regions best handled by the machine and obtains human expertise for other instances.This allocation explicitly uses complementarity rather than maximizing each component's standalone performance.
4 Experiments
Experiments in Galaxy Zoo and CAMELYON16 compare fixed and jointly trained discriminative and value-of-information models, varying model capacity and error costs. Joint training generally matches or improves fixed models, with larger benefits under limited capacity and asymmetric costs, while analyses show machines reallocating effort to complement human strengths.
- Evaluation setup: Total loss combines classification error with the cost of human queries, and joint models are selected across different weightings of classification loss and query cost.The experiments vary network layers and hidden units to study model capacity.
- Overall results: Joint models uniformly outperform or tie fixed counterparts across the evaluated settings.For Galaxy Zoo, loss reductions reach 21–73% for one-layer VOI models and 29% for two-layer discriminative models; CAMELYON16 gains reach 20% for one-layer discriminative models.
- Model capacity: Up to 15% lower CAMELYON16 discriminative-model loss results from joint training when hidden units are reduced from 50 to 20.Overall, complementarity-focused training provides larger benefits with limited model capacity.
- Asymmetric costs: Up to 10% discriminative and 4.8% VOI improvements arise when false negatives are weighted twice as heavily as false positives in CAMELYON16.The gaps between fixed and joint models grow under asymmetric costs.
- Complementary allocation: Joint training queries humans more for machine-difficult Galaxy Zoo classes and prioritizes a CAMELYON16 region containing over 68% of human errors in 10% of instances.In CAMELYON16, the joint VOI model improves the fixed model’s 0.29 error rate to perfect accuracy in that region, yielding a 2% overall loss reduction despite increased errors elsewhere.
5 Conclusion and Future Work
The paper studies discriminative and decision-theoretic methods for optimizing human-machine complementarity in two real-world tasks. It analyzes which task characteristics increase the benefits of complementarity-focused training and frames the methods around human-machine teamwork.
- The paper optimizes human-machine complementarity using discriminative and decision-theoretic modeling methodologies across two real-world tasks.