 |
Center for Nonlinear Studies |
This talk provides an overview of research in MIT’s Stochastic Systems Group (SSG). The (non-disjoint) intellectual themes of research in SSG broadly speaking all involve the representation and extraction of information involving complex data and phenomena, where our focus is on developing classes of models that capture rich classes of problems of practical importance and that also lead to scalable algorithms for problems of learning, inference, networked fusion, data mining, and large-scale data assimilation. Applications of our work include geophysical data assimilation and inversion (with specific applications ranging from oceanography to oil exploration), computer vision, situational awareness/multisensor fusion, sensor networks, and medical imaging. A critical component of much of our work is in developing methods that not only produce estimates but also measures of uncertainty in those estimates that are critical to their effective use. More specific areas of focus are on developing statistical methods for distributed phenomena, a research area that involves graphical models, Markov random fields, and network-constrained information transfer. A second focus of our work is on the representation and extraction of geometric information in complex problems. Curve evolution and very recent Markov Chain Monte Carlo methods are components of our work in this area, with applications in image segmentation/computer vision, in extracting and tracking dynamically evolving shapes (such as the heart), in estimating boundaries of sharp contrast including from remote sensing data, and in mapping subsurface gravitationally anomalous regions from gravity measurements. The third major component of our research is in machine learning, where our work includes applications in computer vision, multimodal data fusion, extraction of “links” among objects/signals in video and multimodal data, complexity-reduction methods in building models directly for fusion or classification, and blending physics and machine learning for data fusion and the extraction of statistically meaningful models for macroscopic behavior in the presence of finer-scale physics, uncertainties, and variability. A number of directions for future work will also be described.
Host: Frank Alexander