AI Infrastructure and Security Workshop Abstracts
Title: Creating Intelligent Cyberinfrastructure for Democratizing AI: Overview of the Activities at the NSF-AI Institute ICICLE
Title: NVIDIA and the Future of AI
Title: The need for trustworthy and ethical development and deployment of AI
OAK Supercomputing Conference Track I - Main Session Abstracts
Title: Future of AI
Title: The Future of the DataCenter With an Eye on Sustainability
Bio: Matthew received his Bachelor of Arts in Molecular, Cellular and Developmental Biology from the University of Colorado, Boulder and went on to work with the National Renewable Energy Laboratories in Golden, CO on biofuel projects. Matthew then switched his energies to learning how to design and architect x86-based supercomputers primarily for use in life science projects. He left life science research and joined IBM onto the North America Advanced Technical Support team where he broadened his scope of High Performance Computing (HPC) designs into other sectors such as Oil and Gas, Digital Media, Weather/Atmospheric Sciences and General Research. As the market for Intel-based clusters continued to expand, Matthew progressed to the role of Architect in the System x Product Marketing team at IBM which was then acquired by Lenovo in 2014. At Lenovo, Matthew continued to work on the worldwide HPC team as the Director of HPC Performance and Architecture and has recently moved into a new role focusing solely on liquid-cooling in server systems and sustainability in the data center.
Title: Using HPC to Support Applied Learning in an Industrial Math
Abstract: Spatial relations of objects are ubiquitous in constructed environments. Consequently, robots designed to aid in tasks within kitchens, offices, and homes would greatly benefit from the ability to identify the spatial relationships among objects, especially when robots collaborate with humans. To enable robots with the ability to learn spatial relationships among objects, robots need to understand the spatial relationships of unseen objects because we cannot enumerate all objects. While previous research has focused on recognizing pairwise semantic relationships and sequentially manipulating objects to alter these relationships, none have demonstrated the capacity to organize objects into intricate structures such as circles or table settings. To tackle this challenge, we propose a novel deep learning model based on the Bayesian neural network (BNN) called Spatial Bayesian. This network takes as input a partial-view point cloud representing the current arrangement of objects and a structured language command describing the desired configuration of objects. Through rigorous experimentation, we demonstrate that Spatial Bayesian empowers a physical robot to rearrange unfamiliar objects into semantically meaningful structures, leveraging multi-object relational constraints inferred from the language command.
Title: Work-In-Progress: “Transforming Edge Computing with Machine Learning to Optimize Heterogeneous Cloud”
Abstract: This work introduces a methodology to apply machine learning to optimize heterogeneous cloud by transforming edge computing. First, using VisualSim simulator, we model and simulate a heterogeneous system with cloud servers (CSs), core routers (CRs), edge servers (ESs), edge routers (ERs), distribution routers (DRs), and edge devices (EDs), where EDs are of different types. Then, we optimize the system by applying machine learning models for optimal utilization, latency, and power consumption through offloading computations from cloud to edges.
Krishna Marupaka
Deepika Nuthalapati
Title: Multi-Modal Spatial-Temporal Vision Transformer for Analyzing Climate Change Impact in South Dakota
Abstract: Understanding the effects of climate change is imperative, especially in regions like South Dakota, where unique environmental challenges require specialized analytical tools. This work proposes a Multi-Modal Spatial-Temporal Vision Transformer-based model designed to comprehensively analyze the impact of climate change in South Dakota. The proposed model integrates three key components: a Multi-Modal Transformer, a Spatial Transformer, and a Temporal Transformer. It utilizes visual remote sensing data and short-term meteorological information to simulate and understand the immediate effects of climate variability. The Spatial Transformer captures fine-grained spatial dependencies between various regions in South Dakota, providing enhanced insights into the localized impact of climate change. Concurrently, the Temporal Transformer learns long-range temporal dependencies, offering a nuanced understanding of how climate change unfolds over time. Through extensive analysis and experimentation, conducted in the context of South Dakota's unique climate challenges, our results highlight the effectiveness of the proposed model in providing a comprehensive assessment of climate change impacts. By seamlessly integrating satellite imagery and meteorological data, our work offers valuable insights into the combined effects of short-term climate fluctuations and long-term climate change in South Dakota. Our findings provide a robust foundation for informed decision-making, aiding policymakers and environmental scientists in addressing the dynamic challenges posed by climate change in South Dakota.
Title: Using High-throughput Computing for Viral Genomics
Abstract: Classification of viruses is important for monitoring and early detection of outbreaks. Currently, there are more than 15,000 different types of viruses in the NCBI RefSeq database, and nearly all of these have recently been giving hierarchical classification, from the International Committee on Taxonomy of Viruses. All of the RefSeq viral genomes can be placed into forty different classes, and for each class we used a k-mer based method (Mash) to cluster all of the genomes. We then used the 40 medioid genomes from each class to build a dendogram, representing all known viral classes. This will be useful as a reference in displaying all of the viruses detected in an environmental sample, representing for example wastewater collected at a particular time. Continual monitoring will allow for establishing the baseline ecological structures, and rapid detection of potential viral outbreaks.
Title: Bringing Big Data Analytics in Accounting into Classrooms: A Case Combing XBRL Data and High-Performance Computing (HPC) Resources
Abstract: Big data analytics is the process of collecting, processing, and analyzing large and complex data sets to extract valuable insights and support decision making. Despite the critical importance of big data analytics skills in this age when data volume is experiencing explosive growth, very few accounting programs teach big data analytics effectively. To help students gain knowledge and experiences of applying big data analytics in accounting, we developed a case for financial misreporting detection using multiple large real-world financial statement datasets based on eXtensible Business Reporting Language (XBRL). To deal with the challenges of high requirements on computer resources for efficiently sharing and processing large XBRL formatted data, we take advantage of High-Performance Computing (HPC) resources that can be accessed by educators and students easily for free at most universities and other public research institutions. Through this case, students will learn about XBRL concepts such as extension taxonomies and how to use free and popular tools such as Python to process and analyze XBRL data. They will learn how to read in large raw data sets, extract financial line items, transform the data, and perform basic data analysis such as descriptive statistics and data visualization for the distribution of earnings and other financial line items that can indicate potential financial misreporting cases. This approach will equip accounting students who have little or no statistical background and programming skills with basic big data analytics skills for detailed financial data analysis.
Title: Fueling Regional Prosperity: Wichita State University and Connected Nation Forge Path with $5M Grant for Carrier-Neutral Internet Exchange Point
Title: GPN Initiative for R2 and Emerging Research Institutions
Abstract: The Great Plains Network is exploring opportunities for and seeking partners to for an initiative to support colleges and universities emerging into supercomputing research and advanced applications. This brainstorming session will set the stage for an upcoming consortium-wide grant proposal.
Title: Overview of the GP-ENGINE Project
Title: Tornado Tracking with Waves in the Earth
Abstract: The use of modern radar methods is very successful if tracking tornados in real time. However, the radar aperature that sees the tornado loses focus in the very near earth surface. It is of great benefit to know if the tornado is actually on the ground. Hilly terrain blocks the more or less straight line radar signals. Using elastic wave equations that correctly model the four waves P, S, R and L it is possible to track these on the ground tornados is real time. The joint inversion of what the radar sees and says, and what the seismic signals see will impove the quality of the warning processes and help humans seek shelter and safety. In this talk I will demonstrate using movies of simple wave models of a tornado touching an elastic earth surface and generating seismic waves into the subsurface. The computational challenge for solving the track and prediction process will be reviewed.
Title: Using HPC To Train Spatial Bayesian Neural Networks for Human-Robot Collaboration
Abstract: Spatial relations of objects are ubiquitous in constructed environments. Consequently, robots designed to aid in tasks within kitchens, offices, and homes would greatly benefit from the ability to identify the spatial relationships among objects, especially when robots collaborate with humans. To enable robots with the ability to learn spatial relationships among objects, robots need to understand the spatial relationships of unseen objects because we cannot enumerate all objects. While previous research has focused on recognizing pairwise semantic relationships and sequentially manipulating objects to alter these relationships, none have demonstrated the capacity to organize objects into intricate structures such as circles or table settings. To tackle this challenge, we propose a novel deep learning model based on the Bayesian neural network (BNN) called Spatial Bayesian. This network takes as input a partial-view point cloud representing the current arrangement of objects and a structured language command describing the desired configuration of objects. Through rigorous experimentation, we demonstrate that Spatial Bayesian empowers a physical robot to rearrange unfamiliar objects into semantically meaningful structures, leveraging multi-object relational constraints inferred from the language command.