Category: News & Events

Please join the MRC on April 17 from 11AM-12PM in Room 928 of the College of Computing and Informatics for “Accelerating Artificial Intelligence for Data-Driven Discovery” a talk delivered by Shih-Chieh Hsu, PhD, University of Washington.

Abstract

As scientific datasets become progressively larger, algorithms to process this data quickly become more complex. In response, Artificial Intelligence (AI) has emerged as a solution to efficiently analyze these massive datasets. Emerging processor technologies such as graphics processing units (GPUs) and field-programmable gate arrays (FPGAs) allow AI algorithms to be greatly accelerated. The Accelerated AI Algorithms for Data-Driven Discovery (A3D3) Institute sponsored by the National Science Foundation under the Harnessing the Data Revolution program is established to enable real-time AI at scale for broad applications. In this talk, Hsu will give an overview about the challenges of high energy physics, multi-messenger astrophysics and neuroscience regarding AI across latency and throughput regimes. He will introduce various techniques for model compression using state-of-the-art techniques such as pruning and quantization for edge computing. He will demonstrate that acceleration of AI inference as a web service represents a heterogeneous computing solution. Finally Hsu will discuss how A3D3 can bring together disparate communities that are threaded by common data-intensive grand challenges to accelerate discovery in science and engineering.

Biography

Shih-Chieh Hsu, PhD is a professor in physics and adjunct professor in electrical and computer engineering at University of Washington (UW), and director of NSF HDR Institute: Accelerated Artificial Intelligence Algorithms for Data-Driven Discovery. He earned the BS/MS in physics from National Taiwan University and the PhD in Physics from University of California San Diego. He is working on experimental particle physics using proton-proton collision data from the Large Hadron Collider. His research interests range from dark matter searches with the ATLAS experiment neutrino cross-section measurements with the FASER experiment innovative artificial intelligence algorithms for data-intensive discovery and accelerated machine learning with heterogeneous computing.

ID4: Institute of Data Driven Dynamical Design

Hosted by the Metadata Research Center, College of Computing & Informatics, Drexel University

• DATE: April 15-16, 2024
• LOCATION: Quorum – University City Science Center, 3675 Market Street, Philadelphia PA, 19104

AI-ready data refers to the high-quality and well-prepared data that is optimized for use in artificial intelligence (AI) applications. AI-ready data increasingly encompasses the inclusion of metadata and ontologies to enhance the value and usability of data. Metadata provides essential context and information about the data, and ontologies offer structured semantic representation of a particular domain. These additional layers of information help data scientists,data scientists, researchers, and AI systems understand, interpret, and apply appropriate algorithms and models for analysis. Metadata and ontologies enable consistent data integration, interoperability, and knowledge sharing across systems, while facilitating more knowledgeable AI applications. Additionally, these systems are proving vital for supporting the FAIR (Findable, Accessible, Interoperable, and Reusable) principles and reproducible computational research (RCR).

Despite these capacities, approaches for developing, implementing, and sustaining metadata and ontologies within AI-ready data pipelines remain inconsistent, cumbersome, and lack sufficient support. Challenges underlie the full data lifecycle from data creation, collection, and research, to longer-term aims of data preservation, archiving, reuse and support for research reproducibility. Collective, community driven efforts are needed to address current obstacles and maximize the value and reliability of data. The AI-Ready Data: Navigating the Dynamic Frontier of Metadata and Ontologies workshop is a step toward addressing this challenge. This workshop will bring together a community of individuals with expertise across the data lifecycle to discuss issues, share solutions, and chart a path forward for addressing key challenges in preparing AI-ready data for scientific research. 

Specific workshop goals are to:

1. Collectively define the state of AI-ready data challenges in the metadata and ontology space
2. Share current successes and solutions leveraging metadata standards and ontologies.
3. Contribute to a road map to accelerate the preparation of data for artificial intelligence (AI) applications.

Current topics

• What is AI ready data
• Research Bottlenecks: Data Life Cycle Challenges and Solutions with Scientific Data
• Metadata and Ontologies: Human in the Loop in the Era of LLMs
• Annotation: Large-scale Data and Balancing Human and Machine Driven Approaches 
• Standards Development, Adoption, and Implementation: Realities and Fictions
• Knowledge Graphs
• Ontology Guided Knowledge Extraction: Leveraging Scholarly Big Data for Scientific Discovery
• Future Directions with Metadata and Knowledge Organization Systems

Science has traditionally harnessed data to inform decisions. Historically, data was sufficiently low-dimensional and manageable for human processing. However, the rapid expansion of sensing technologies across disciplines has overwhelmed traditional human-centric methods with vast, high-velocity data streams from diverse and often unreliable sources. Despite the remarkable advances in computers and large language models like ChatGPT, their capabilities remain limited. Current AI algorithms predominantly excel in interpolation, not extrapolation, leading to unrealistic and nonsensical outputs when stretched beyond their training data.

This talk explores the intersection of massive data influx and AI, focusing on their limitations and potential in enhancing decision-making, particularly in data-driven infrastructure. We propose a “humanistic carrot” – not the “stick” approach to address pressing challenges in scientific data management, spotlighting DataFed – a comprehensive data management system. This platform facilitates autonomous pipelines for the curation, sharing, searching, and fine-grain access control of data and metadata. We demonstrate how DataFed can streamline data management for experimentalists, enhancing data stewardship while reducing their workload.

We also delve into the intricacies of handling high-velocity data streams, where gigabits per second of data necessitate immediate processing for critical decision-making or autonomous control. This section covers deploying high-availability inference servers for on-demand data analysis and reduction. Additionally, we explore the concept of AI co-design, where algorithms are optimized to fit on programmable logic, enabling rapid, intelligent analysis, decision-making, and control on ultra-low cost, low-power devices at unprecedented speeds. Finally, we discuss the broad applicability of these methodologies across various fields, from particle physics to astronomy, highlighting their potential to revolutionize our approach to data and AI integration.

Dr. Joshua C. Agar is an Assistant Professor in the Department of Mechanical Engineering and Mechanics at Drexel University. With a foundational background in experimental materials science, Dr. Agar is predominantly renowned for his pioneering contributions to AI algorithms, computing infrastructure, and the development of cyber-physical systems in the fields of materials synthesis and microscopy. His expertise has been applied across a wide array of disciplines, including particle and plasma physics, materials science, and fluid dynamics. An active member of various AI communities, particularly the FastML community, which emphasizes ultra-low latency ML co-design, Dr. Agar has earned recognition as a leader in AI innovation. His work has garnered attention from prestigious institutions such as the National Academy of Engineering and the National Science Foundation

Woo hoo

Ajani Levere, a Drexel University STAR Scholar working with Drs. Jane Greenberg and David Breen presented their research imageomics at the STAR Scholars showcase on August 31, 2023. Their presentation was titled “Computational Fish Specimen Classification: Advancing Machine Learning Model Accuracy” and was part of the NSF-HDR: Biology-guided Neural Networks for Discovering Phenotypic Traits. Ajani’s research is continuing under the guidance of Dr. Greenberg. They describe their project as follows:

Digital specimen metadata is valuable for scientific research and discovery, yet sparse specimen metadata availability restricts its potential. In addition to computational efforts made to remedy this issue, Machine Learning (ML) classification was performed on a computed metadata component, the outline extracted from fish specimen images. An ML model (MLM) approach provided a computational genus classification for a given fish outline. This research improves the MLM’s ability to accurately classify fish from their 2D outlines and demonstrates the expressiveness of this computed metadata item.  

In our analysis, we inspected the outlines of the error cases, followed by a statistical review of their numerical data. We discovered our dataset limited higher MLM accuracy potential. Refactoring the dataset with a reduced feature length thus enhanced our dataset for MLM interpretability. Experimental results indicate a 96% accuracy, a 5% improvement over previous results. These results confirm the outline as a unique and highly distinguishable metadata component. Computing metadata components of this nature aids the development of a more robust metadata catalog for ML researchers.

The Metadata Reasearch Center congratulates Jane Greenberg, its Director and Founder, for receiving the Association for Information Science & Technology’s (ASIS&T) 2023 Research in Information Science Award. The award “recognizes an individual or team who has made an outstanding contribution to information science research. The award is for a systematic “program of research” in a single area at a level beyond the single study.” ASIS&T recognized Dr. Greenberg’s wide-ranging contributions, including her current positions as principal investigator on the Metadata Capital Initiative (MetaDataCAPT’L) and the NSF-funded Institute for Data Driven Dynamical Design (ID4), and the IMLS-funded project LEADING (LIS Education and Data Science Integrated Network Group). The award committee also singled out her work with the Biology-guided Neural Network (BGNN) project and the Helping Interdisciplinary Vocabulary Engineering (HIVE) tool. Please click here for the full press release from ASIS&T, “Jane Greenberg Receives Association for Information Science and Technology (ASIS&T) Research in Information Science Award.”

On Friday, May 19th, the 2023 LEADING Forum took place in the Science Center/Quorum at Drexel University. Highlights form the Forum included a panel on ChatGPT, a fellows panel, a fellow poster session, and keynote presentations from Florence Hudson (Executive director, Northeast Big Data Innovation Hub at Columbia University), and Laurie Allen (Chief, Digital Innovation Lab (LC Labs), Library of Congress). The 2023 OCLC/LEADING Data Challenge preceded the forum on Thursday, May 18th.

In moving forward, we welcome the incoming LEADING 2023 fellows. This year’s cohort includes 18 fellows from 14 iSchools and LIS institutions from across the country. June kicks off the 2023 LEADING boot camp, which precedes the 6-month fellowship period. Read more about the 2023 fellows here.

Scott McClellan, a second year doctoral student, presented research results to the “Data representation in materials and chemicals based on harmonised domain ontologies” birds of a feather group at the Research Data Alliance’s 20th Plenary meeting in Gothenburg, Sweden on March 21-23, 2023. His presentation, titled “Along the Border: Term Overlap Among 5 Matportal Ontologies,” focused on term overlap among a subset of ontologies maintained at the Matportal repository. It looked at how term matching algorithms for materials science semantic artifacts differed when locating terminological or URI results. His presentation stemmed from prior research done with Drs. Yuan An and Jane Greenberg and fellow graduate student Xintong Zhao. [Slides]

MRC’s Jane Greenberg and Richard Marciano, Advanced Information Collaboratory (AI Collaboratory) University of Maryland, presented at the 2022 DLF Forum on Wednesday, October 12th.

Their panel, titled “Innovating Data Science Education and Computational Thinking: Connecting iSchools and LAMs,” presented about two national Institute of Museum and Library Services (IMLS) initiatives connecting leading GLAMs (galleries, libraries, archives, and museums) and educators, and innovative data science education. Jane presented about the LEADING (The LIS Education and Data Science Integrated Network Group) fellowship project, and Richard presented about the TALENT (Training of Archival & Library Educators with iNnovative Technologies) Network.

The presentation slides are available here: [LINK].