This webinar will showcase some of the game-changing research supported by our Catalyst Grants program.

This event was recorded and will be on the UM Youtube channel shortly.

Speakers

Stephen Smith
Associate Professor of Ecology and Evolutionary Biology
University of Michigan

The Emergence of Biological Complexity and Evolutionary Innovation in Plant Genomes

BLANK

Xun Huan
Assistant Professor of Mechanical Engineering
University of Michigan

Towards Bayesian Uncertainty Quantification in Deep Learning Models for Brain Tumor Segmentation

While the use of deep learning (DL) models in healthcare has grown rapidly in recent years, the uncertainty/confidence information in their predictions is often unavailable and unreported. A lack of such information can render decision-making dangerous, and prompt clinicians to hesitate in using and trusting these machine learning technologies. We propose to adopt principles and computational methods of uncertainty quantification for medical artificial intelligence applications, focusing on a problem of brain tumor segmentation from MRI scans. As a first step, we assess the robustness and sensitivity of two such DL models, U-Net and SqueezeU-Net, with respect to uncertainty in model weights, which may arise due to sparsity and noise in training data features as well as labels. We achieve this through Monte Carlo uncertainty propagation of noise injected on trained weight values. The resulting uncertainty of segmentation maps can then be presented and visualized through robustness maps and summarizing box-plots of the Dice coefficients, which can help indicate the regions where our models do not predict well and most susceptible to training noise. In our on-going work, we seek to compute the Bayesian posterior distributions for the weights directly from training data. However, performing a full-scale inference for the millions of weights in U-Net and SqueezeU-Net would be prohibitive. Instead, we develop a procedure to use sensitivity analysis to identify the most important subset of weights (or layers), and perform a targeted Bayesian inference on this lower-dimensional parameter space.

Monica Valluri
Research Professor of Astronomy
University of Michigan

Probing the nature of dark matter by modeling the Milky Way

Despite nearly four decades of research in astrophysics and particle physics, the nature of dark matter, the substance that comprises 85% of the matter in the universe, is unknown. The shape of the Milky Way’s dark matter distribution and the variation of this shape with radius are important probes of the nature of dark matter. Mapping the detailed formation history of the Milky Way, especially the number of satellites that were assimilated by our Galaxy and their masses and their time of infall will provide clues to the dark matter distribution in satellites as well as evidence for nearby streams and dark matter satellites. We are developing a multi-pronged approach to understanding the nature of dark matter with new dynamical tools, new simulations and analysis of large cosmological simuations. I will describe progress on our efforts to enhance the galactic dynamics package AGAMA (Vasiliev, 2019)by adding GPU acceleration for the potential and action solvers. I will provide an update on how we are using positions and velocities for old stars in the Milky Way’s halo to determine the three dimensional shape of the dark matter distribution and its variation with radius.I will describe new simulations of the evolution of satellites that merge with our Milky Way that can lead to insights into the fundamental nature of dark matter. Finally I will descibe the use of two cluster finding tools (a self organizing mapping and multi-dimensional density estimation), that when applied to action-space properties of stars in the Milky Way’s halo, can yield insights into the accretion history of our Galaxy. This concert of efforts will significantly advance our goal of understanding the fundamental nature of dark matter using the properties of stars in the Milky Way.