The Research Computing Center launched a new speaker series that will bring experts from various fields and areas of research to share their experiences and insights. We will host speakers from the University, industry, and other academic institutions. The speaker series will share the annual theme of Mind Bytes.
This year’s theme is Computing, Data, and Beyond-Impact on our world.
Upcoming Speaker Series Events:
August 23, 2022, 11 AM – 12 PM CDT, via Zoom
Salman Habib, Director of Computational Science Division and Distinguished Fellow at Argonne National Lab.
The Exascale Scientific Computing Landscape: Challenges and Opportunities.
Computing has become an essential enabling and empowering component across all of science. In particular, large-scale computing is playing an ever-increasing role in theory and modeling and in data collection and analysis, as well as accounting for the increasing influence of machine learning. Early exascale systems are harbingers of a future computing roadmap that promises major leaps in capability, but one that poses significant challenges that will need to be faced by the scientific community. In this talk I will survey some of the history, current trends, and indications of near-future directions in scientific computing, focusing on a subset of examples that lie at the interface of high-performance computing, large data sets, and machine learning.
Register here on Eventbrite.
February 21, 2022, 2 PM – 3 PM CST, via Zoom
Shashi Shekhar, McKnight Distinguished University Professor, University of Minnesota
What is special about Geo-AI and Spatial Data Science?
Rapid expansion in spatial big data (e.g., trajectories, remote-sensing) is fueling growth of Geo-AI for making previously unimaginable maps, answering trail-blazing geo-content based queries, discovering groundbreaking spatial patterns, etc. Applications span from apps for navigation, ride-sharing, and delivery to monitoring global crops, climate change, diseases, and smart cities to understanding cellular or urban patterns of life.
However, one-size-fit-all machine learning performs poorly (e.g., salt-n-pepper noise, inaccuracy) due to spatial autocorrelation and variability, which violate the common i.i.d. assumption (i.e. data samples are generated independently and from identical distribution). Furthermore, high cost of spurious patterns requires guardrails such as noise tolerance, and modeling of spatial concepts (e.g., polygons) and implicit relationships (e.g., distance, inside). In addition, methods discretizing continuous space face the modifiable areal unit problem (e.g., gerrrymandering).
Thus, the talk suggests spatial data science approaches and describes methods for spatial classification and prediction (e.g., spatial auto-regression, spatial decision trees, spatial variability aware neural networks) along with techniques for discovering patterns such as noise-robust hotspots (e.g., SaTScan, linear, arbitrary shapes), interactions (e.g., co-locations, tele-connections ), spatial outliers, and their spatio-temporal counterparts (e.g., cascade , mixed-drove co-occurrence ). It concludes by calling for inclusion of spatial perspectives in data science courses and curricula.