LANS Seminar

Seminar Title:Error‐controlled feature selection for ultrahigh‐dimensional and highly correlated feature spaces using deep learning – An application in neuroimaging studies

Speaker:Arka Ganguli, Postdoctoral Researcher, Argonne National Laboratory, Specializing in Statistical Machine Learning

Date/Time: May 8, 2024/ 10:30 AM-11:30 AM
Location: See Meeting URL on the cels-seminars website which will require an Argonne login.

Description: Deep learning has been at the center of analytics in recent years due to its impressive empirical success in analyzing complex data objects. While its application in feature selection holds promise for uncovering insightful predictors, challenges persist, particularly in managing ultrahigh-dimensional, correlated features, and elevated noise levels To bridge this gap, we propose a novel screening and cleaning method that integrates deep learning to achieve data-adaptive discovery of highly correlated predictors while controlling error rates. Extensive empirical evaluations across simulated scenarios and real datasets demonstrate our method’s efficacy in achieving high power while minimizing false discoveries.Transitioning to neuroimaging, we explore cognitive reserve—a phenomenon observed in older adults maintaining cognitive abilities despite neuropathological diseases. Leveraging diffusion MRI tractography, we investigate subcortical white matter connections as potential cognitive reserve markers. However, traditional statistical analyses encounter challenges in handling tractography data’s high dimensionality and correlation. To overcome this, we introduce a flexible feature selection algorithm combining deep learning for cluster-level discovery with controlled error rates. Through simulations and application to clinical neuroimaging data, our approach reveals meaningful discoveries in brain regions linked to neurodegeneration risk and resilience. This integration of screening, cleaning, and deep learning offers a comprehensive solution for the neuroimaging study, facilitating deeper insights into cognitive reserve and related neurobiological substrates.

Bio: Arka Ganguli is a Postdoctoral Researcher at Argonne National Laboratory, specializing in statistical machine learning. He holds a Ph.D. in Statistics from Michigan State University, where his research focused on advanced statistical methodologies for feature selection in ultra-high dimensional datasets. Arka’s academic background includes a Bachelor’s and Master’s degrees in Statistics from the University of Calcutta, India. His research interests revolve around developing and applying statistical methods for analyzing high-dimensional datasets, encompassing feature selection, deep learning, and generative models.

Please note that the meeting URL for this event can be seen on the cels-seminars website which requires an Argonne login.