Cooperation Instead of Competition: Harvard & Microsoft Research Optimizes AI-Human Teamwork

Machine learning (ML) systems are increasingly adept at tasks that have traditionally required human expertise. In challenges such as real-time gameplay, models like AlphaGo and AlphaStar have outperformed even the world’s top human professionals.

Machines beating up on humans is old news — maybe it’s time we all learn to work nicely together? A new paper explores the potentially richer optimizations that could result from a spirit of human-machine teamwork built on complementarity.

The first author of Learning to Complement Humans is a Harvard PhD student focused on optimization and ML for social impact. Bryan Wilder and his supervisors at Microsoft Research propose that optimizing AI performance in isolation overlooks common situations where human expertise can contribute complementary perspectives. The paper introduces methods for optimizing team performance wherein machines perform some parts of a given task and humans others.

Ideally a machine would be able to handle everything, but the researchers say this is rarely practical in complex domains: “When perfect accuracy is unattainable, the machine should focus its limited capacity on regions of the space where it offers the most benefit.”

The researchers set out to apply ML techniques on problems that are difficult for humans, and to identify instances that were difficult for the machine and could benefit from human input. They developed an end-to-end learning strategy that optimizes human-machine team performance by considering the distinct abilities of humans and machines; and introduced both discriminative and decision-theoretic approaches for model optimization.

The team began by proposing a family of related approaches for training ML systems for human-machine complementarity. The baseline method ML model predicted an answer to a given task, and built a policy for deciding when it should reach out to a human for help.

The team then conducted experiments in two real-world domains — scientific discovery and medical diagnosis — to explore opportunities and methods to best leverage human-machine complementarity. They chose a galaxy classification task and a breast cancer metastasis detection task respectively for the two domains. In comparative studies, the proposed methods improved on the performance of either machines or humans working alone.

The experiments show that joint training provides greater benefits when the ML model has limited capacity, as this forces the model to pick the parts of the task it is best at to focus on. The researchers also found that training for complementarity works best when there is an asymmetric cost to errors — in other words, optimizing for team utility reaps a higher payoff when prioritizing among potential errors is required.

The researchers demonstrated that their discriminative and decision-theoretic approaches can optimize the expected value of human-machine teamwork by training a model to identify and respond to the shortcomings of ML systems as well as the capabilities and blind spots of humans.

The paper identifies a possible area for future research on human-machine complementarity as optimization of team performance when interactions extend beyond the current level of querying humans for answers. This could include settings with more complex, interleaved interactions and those with different levels of human initiative and machine autonomy.

The paper Learning to Complement Humans is on arXiv.

Journalist: Yuan Yuan | Editor: Michael Sarazen

Thinking of contributing to Synced Review? Synced’s new column Share My Research welcomes scholars to share their own research breakthroughs with global AI enthusiasts.

We know you don’t want to miss any story. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.

Need a comprehensive review of the past, present and future of modern AI research development? Trends of AI Technology Development Report is out!

2018 Fortune Global 500 Public Company AI Adaptivity Report is out!
Purchase a Kindle-formatted report on Amazon.
Apply for Insight Partner Program to get a complimentary full PDF report.