|
 |
Mathematical 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 focuses on the development of robust, scalable and interpretable mathematical and computational methods to solve challenging problems involving complex and/or massive visual data with minimum human intervention. Current research in the lab includes learning tasks from procedural videos of complex activities, scalable and interpretable low-shot and multi-label recognition, adversarial attacks on fine-grained and multi-label models, and coherent and structured data summarization. For more information about our research activities, please visit the Publications page. |
Procedural Task Learning from Complex Activities
We develop video recognition, segmentation and anticipation methods to learn models of complex tasks from procedural (e.g., instructional) videos. We use these methods to develop assistive technologies that guide users of all skill levels through familiar and unfamiliar tasks via AR/VR technology.
|
Zero-Shot Fine-Grained Recognition
Fine-grained recognition requires distinguishing visually very similar classes. Training deep networks, however, requires many training samples for each class, which are hard to gather. We study zero-shot attribute-based fine-grained recognition methods that can efficiently generalize to previously unseen classes without requiring strong supervision.
|
Multi-Label Human-Object Interaction Recognition
Images often contain multiple labels such as objects or human-object interactions (HOIs), which are actions performed on objects. We develop robust and scalable multi-label and HOI recognition, based on attention models, that learn from weak supervision, without expensive bounding-boxes.
|
Adversarial Attacks and Defenses
Deep neural networks are sensitive to impreceptible perturbations of input data, which has motivated research on designing attacks and defense mechanisms for DNNs. In this project, we study efficient and generalizable attacks and defenses for fine-grained and multi-label recognition problems.
|
Structured Data Summarization
We design robust and scalable summarization methods that handle structured dependencies in massive and complex data, adapt to tasks and require minimum/no supervision. We use convex and non-convex optimization and deep learning to develop algorithms, analyze their guarantees and apply them to real-world tasks.
|
Optimization for Machine Learning
We address the development and analysis of continuous and discrete optimization algorithms that can efficiently handle challenges posed by machine learning models and massive or high-dimensional data. We apply our tools to problems such as image and video understanding in the wild and multimodal machine learning.
|
|
