|
 |
Mathematical, Computational and Applied Data Science Lab is a research lab in the Khoury College of Computer Sciences at the Northeastern University, led by Dr. Ehsan Elhamifar. Our research includes 1) development of mathematical and computational techniques to model and analyze complex problems involving massive and heterogenous high-dimensional data, 2) design of robust, efficient and scalable algorithms to perform learning, inference and decision making on big data using developed mathematical models, and 3) application of our methodologies and techniques to solve challenging real-world problems in computer vision, robotics and other areas. Current research in the lab includes coherent, structured and interpretable data summarization, procedure learning from instructional data, large-scale multi-label and zero-shot recognition, clustering and completion of multi-manifold data and active learning in visual data. For more information about our research activities, please visit Research and Publications pages. |
Structured Data Summarization
One of the fundamental challenges in science and engineering is the massive amounts of unlabeled data that need to be processed and analyzed. We develop mathematical and computationally efficient, robust and scalable methods for summarization of massive datasets, that handle complex structured dependencies in high-dimensional data, adapt to tasks and require minimum supervision. We use convex and non-convex optimization and deep learning to develop algorithms, analyze their theoretical guarantees and apply them to real-world applications.
|
Zero-Shot and Multi-Label Recognition
Recognition of all labels in an image or video is a challenging problem with numerous applications, including self-driving cars, surveillance systems and assistive robots. We develop scalable and robust multilabel recognition systems that recognize tens of thousands of labels with limited or no annotations while incorporate and infer dependencies across labels and images. We use and extend deep Convolutional Networks, multilabel learning, semi-supervised learning along with attention modeling and apply our techniques to large-scale datasets.
|
Procedure Learning from Complex Activities
Procedure learning from complex activity data, which is to discover and learn visual models for the sequence of key-steps needed to perform a complex task, is an emerging problem that enables to teach autonomous agents to perform complex tasks or build assistive technologies. We develop automatic methods for procedure learning from unconstrained, noisy, unaligned, multimodal instructional video data. We develop methods based on structured subset selection and deep learning that that generate procedural key-steps and their ordering and learn models of procedures in videos.
|
Optimization for Machine Learning
The increasing complexity, size and variety of today's machine learning models along with the growing size of heterogenous incomplete data call for reassessment of existing optimization methods and development of new efficient tools. We address the development and analysis of continuous and discrete optimization algorithms that can efficiently handle challenges posed by machine learning models and massive high-dimensional data. We apply our tools to problems such as image and video understanding in the wild and multimodal machine learning.
|
|
