January 31, 2018: Invited Speaker Seminar: Various Flavors of Machine Learning

Ahmed Ashraf, Ph.D.
Carnegie Mellon University

Wednesday, January 31, 2018 at 11:45 am

Room EV003.309

Abstract

Fourier SVMs, Medical Imaging, and Deep Learning for Pain Detection

SVMs in the Fourier Domain: In classical machine learning, feature extraction and learning have been two distinct tasks. A key result of my Ph.D. work was that in certain contexts, the two tasks can be subsumed together into a single learning objective that helps us interpret the role played by different feature extraction techniques. By casting a support vector machine objective in the Fourier domain, I showed a mathematical equivalence between linear margin maximizing machine learning classifiers based on high dimensional filter-bank responses, and classifiers trained in the original low dimensional space which optimize a weighted Euclidean distance margin. This equivalence opened the door to the exploration of image resolutions and filter bank sizes previously unimaginable, as the computational and memory requirements of learning a linear classifier are now shown to be independent of the number of filter banks. Several other offshoots of the same idea such as Fourier image alignment, and filter learning would be discussed.

Modeling Cancer Heterogeneity for Therapy Selection: Cancer is a heterogeneous disease. Tumor heterogeneity is emerging as one of the biggest challenges in the way of disease treatment. In a heterogeneous tumor, different parts of the same tumor can have different subtypes of cancer, and as a result therapies targeted to treat only one subtype fail. Traditionally, tumor characteristics have been analyzed on the basis of selective biopsy tissue samples, which typically represent only a portion of generally heterogeneous tumor. I will present my work on building imaging descriptors for capturing tumor heterogeneity from the entire 3d volume of the tumor. These descriptors were then used for training machine learning algorithms for predicting cancer recurrence and choosing therapy options for breast cancer patients based on MRI images of the breast.

Recurrent Neural Networks for Pain Detection: Older adults living in long term care facilities with advanced dementia are unable to verbally communicate their pain. As a result, their pain condition remains under-diagnosed and under-treated because of sole dependence on nursing staff. Shortage of trained nursing staff is well recognized, and often the pain-condition (e.g. a fracture) of older adults remains unnoticed for prolonged periods of time. During my Ph.D., I worked on developing the first computer vision based pain recognition system using automated facial expression. I am currently part of a multi-institute project (Toronto Rehab Institute and University of Regina) for deploying the system in two long term care facilities in Regina, SK. A real life setting with videos from older adults brings along unique challenges requiring new pain detection methods. Fresh methods based on deep learning and recurrent neural networks will be discussed for pain detection.

Biography

Biography Ahmed Ashraf is a post-doctoral fellow at Toronto Rehabilitation Institute, University Health Network, Toronto. He was awarded NCE AGE-WELL fellowship in 2015. Ahmed is a Fulbright Scholar and was a recipient of a 5-year Fulbright Fellowship from 2004-2009. He received his Ph.D. in Electrical and Computer Engineering from Carnegie Mellon University in 2010. He worked as a post-doctoral researcher at the University of Pennsylvania from 2010-2013. In 2016, Ahmed completed the AGE-WELL’s Innovators of Tomorrow training (a yearlong program). His research interests include machine learning, deep Bayesian learning, reinforcement learning, and computer vision; and their applications to biomedical and healthcare settings.