Learning & Modeling Initiative

We develop mathematical concepts and computational methods in areas such as perception, reasoning, learning, and planning. We investigate the data structures and algorithms for representing, reasoning with, and learning from visual, auditory, and other modes of input. The recent, rapid proliferation of quantitative models in the natural and social sciences and in medicine has underscored the importance of techniques for constructing models of complex systems and of algorithms for the efficient manipulation of these models.

Stochastic methods and machine learning techniques have proven to be powerful tools for tackling the increasing complexity of these models, and for harnessing the uncertainty inherent in both models and measurements. In scientific computing, new hierarchical and iterative techniques have multiplied the opportunities offered by the rapid increase in the power of today’s computers.

Our faculty are leaders in these and related areas. Their expertise ranges from scientific computing and machine learning to artificial intelligence, computer vision, and computational biology. For example, functional genomics is benefiting tremendously from techniques that extract reliable information from the huge quantity of imperfect data acquired through sequencing studies, gene-expression experiments, and from other sources. These techniques combine stochastic methods with machine learning and numerical optimization techniques.

In collaboration with the School of Medicine, the scientific computing group has developed techniques for detailed and computationally efficient models of the human heart and brain, of the chemical behavior of different types of stimulants, and of the structure of molecules and proteins. Together with researchers in the School of Engineering, the scientific computing group is also addressing the computational challenges raised by large sensor networks, and is participating in projects on compressive sensing and on tracking with optical devices, bringing both classical and probabilistic techniques to bear on these problems.

The artificial intelligence faculty have developed algorithms for solving stochastic control problems (MDPs) and robot mapping, as well as addressing issues in man-machine interfaces. In the robotic mapping domain, an algorithm developed by the AI group has been deployed on mobile robots donated by industrial partner SAIC to produce maps of physical environments with unprecedented accuracy.

In human-computer interaction, advanced models are at the basis of the “missing axiom theory” of dialogue: This theory proposes that participants in a dialogue exchange the information needed to achieve their respective goals, and interact to supply the missing axioms in the proof that the goal has been satisfied. This approach has been applied to man-machine dialogue systems for equipment repair problems, mystery solving, and automated tutoring. In recent work, the AI group has built a series of intelligent automatic telephone answering systems that make appointments or provide directory service. Computer vision research at Duke focuses on machine learning methods for the analysis of shape and motion in both images and video. For example, machine learning and linguistic analysis methods underlie solutions for the recognition of American Sign Language, finger spelling, and gesture languages for interacting with computers and devices. Methods for the automatic recognition of shape and for the detection of their changes over time are being employed for the dermatological study of skin lesions in the diagnosis of melanoma.

Another project focuses on the development of a theory of stochastic estimation for tracking the boundaries of objects whose topology can change over time. Examples of these are living cells in the field of view of a microscope, the consecutive cross-sections of a three-dimensional tomography scan of a human body, or a cloud of pollutants spreading in the ocean or in the atmosphere.