Explore Artificial Intelligence research areas at Duke Computer Science.
AI for social good
Researchers at Duke use the tools of artificial intelligence to assist with various important societal problems, including (but not limited to) healthcare, antibiotic and cancer resistance, criminal justice, detecting fake news, allocation of public resources to those who need them, environmental sustainability, energy reliability, and political districting. For many of these applications, it is essential that the system satisfy certain interpretability, transparency, morality and/or fairness conditions.
Computational social choice
The theory of social choice concerns how to make decisions based on the conflicting preferences of multiple parties (agents). In computational social choice, these problems are studied from a computational perspective: what are well-motivated and efficient algorithms for solving these problems? Research at Duke has made foundational contributions to this research area, including on voting, fair resource allocation, budget allocation, setting societal priorities, and many other topics.
Computer vision
Computer vision designs algorithms that infer properties of the world from the outputs of a variety of imaging sensors. Application examples include the analysis of medical images for diagnostic purposes, the recognition of people from their faces, retinas, or fingerprints for authentication, the reconstruction of the three-dimensional shape of objects and scenes from multiple images for autonomous driving or architectural surveys, the recognition of objects for robotics applications or for assistance to the visually impaired, the analysis of biological microscopy images or video for measuring cell growth or for categorizing plant species from the shapes of their leaves, and much more. Computer vision research at Duke enables applications like these by developing foundational concepts and algorithms for video analysis, object and activity recognition, and shape reconstruction, and by collaborating with interdisciplinary teams to develop applications. More information is available: Duke Robotics Intelligence and Vision.
Machine learning
Machine learning algorithms allow computers to automatically learn from data to perform complicated tasks in vision, natural language processing and many other fields. Research at Duke addresses both theoretical and practical aspects of machine learning. In particular, researchers at Duke have made significant contributions in learning interpretable models, non-convex optimization and theoretical understanding of neural networks.
Moral AI
Now that AI algorithms are widely deployed in the world, it is becoming clear that the decisions that they make often have a significant moral component. Many of these algorithms require an objective to be specified that they then pursue, but how should we determine the right objective? Research on this topic at Duke focuses on combining insights from computer science, philosophy, and the social sciences to establish well-founded methodologies for determining the objective. Learn more on our Moral AI website.
Reinforcement learning
A reinforcement learning agent is tasked with interacting with an unknown environment and learning, through trial and error, a policy that minimizes long-term cost or maximizes long term reward. Problems as diverse as game playing, robotic control, disease management or user experience management fit this model. Research at Duke addresses fundamental questions in reinforcement learning including algorithm design, sample complexity, feature selection and state space representation.
Search and optimization
Many problems require searching through an exponentially large space of possibilities to find a feasible or optimal solution. Sometimes these problems allow efficient algorithms, whereas in other cases sophisticated techniques from areas such as search, integer programming, and branch decomposition can still be quite effective in practice. Researchers at Duke have used these techniques for solving large games, allocating resources efficiently, ranking alternatives, dead-end elimination, constraint satisfaction in protein structure determination, predicting antimicrobial resistance, and design of therapeutic proteins that are currently in clinical trials.