At Georgia Tech, innovation flourishes where disciplines converge. By encouraging joint appointments, the Institute breaks down traditional academic silos and enables researchers to revolutionize the energy landscape.
The symposium is a chance for faculty to meet new robotics students on campus, as well as a chance to get a better idea of what IRIM colleagues are up to these days. The goal of the symposium is to spark new ideas, new collaborations, and even new friends!
New research from Georgia Tech is giving scientists more control options over generative artificial intelligence (AI) models in their studies. Greater customization from this research can lead to discovery of new drugs, materials, and other applications tailor-made for consumers.
The Tech group dubbed its method PRODIGY (PROjected DIffusion for controlled Graph Generation). PRODIGY enables diffusion models to generate 3D images of complex structures, such as molecules from chemical formulas.
Scientists in pharmacology, materials science, social network analysis, and other fields can use PRODIGY to simulate large-scale networks. By generating 3D molecules from multiple graph datasets, the group proved that PRODIGY could handle complex structures.
In keeping with its name, PRODIGY is the first plug-and-play machine learning (ML) approach to controllable graph generation in diffusion models. This method overcomes a known limitation inhibiting diffusion models from broad use in science and engineering.
“We hope PRODIGY enables drug designers and scientists to generate structures that meet their precise needs,” said Kartik Sharma, lead researcher on the project. “It should also inspire future innovations to precisely control modern generative models across domains.”
PRODIGY works on diffusion models, a generative AI model for computer vision tasks. While suitable for image creation and denoising, diffusion methods are limited because they cannot accurately generate graph representations of custom parameters a user provides.
PRODIGY empowers any pre-trained diffusion model for graph generation to produce graphs that meet specific, user-given constraints. This capability means, as an example, that a drug designer could use any diffusion model to design a molecule with a specific number of atoms and bonds.
The group tested PRODIGY on two molecular and five generic datasets to generate custom 2D and 3D structures. This approach ensured the method could create such complex structures, accounting for the atoms, bonds, structures, and other properties at play in molecules.
Molecular generation experiments with PRODIGY directly impact chemistry, biology, pharmacology, materials science, and other fields. The researchers say PRODIGY has potential in other fields using large networks and datasets, such as social sciences and telecommunications.
These features led to PRODIGY’s acceptance for presentation at the upcoming International Conference on Machine Learning (ICML 2024). ICML 2024 is the leading international academic conference on ML. The conference is taking place July 21-27 in Vienna.
Assistant Professor Srijan Kumar is Sharma’s advisor and paper co-author. They worked with Tech alumnus Rakshit Trivedi (Ph.D. CS 2020), a Massachusetts Institute of Technology postdoctoral associate.
Twenty-four Georgia Tech faculty from the Colleges of Computing and Engineering will present 40 papers at ICML 2024. Kumar is one of six faculty representing the School of Computational Science and Engineering (CSE) at the conference.
Sharma is a fourth-year Ph.D. student studying computer science. He researches ML models for structured data that are reliable and easily controlled by users. While preparing for ICML, Sharma has been interning this summer at Microsoft Research in the Research for Industry lab.
“ICML is the pioneering conference for machine learning,” said Kumar. “A strong presence at ICML from Georgia Tech illustrates the ground-breaking research conducted by our students and faculty, including those in my research group.”
Visit https://sites.gatech.edu/research/icml-2024 for news and coverage of Georgia Tech research presented at ICML 2024.
Bryant Wine, Communications Officer
bryant.wine@cc.gatech.edu
A new machine learning (ML) model created at Georgia Tech is helping neuroscientists better understand communications between brain regions. Insights from the model could lead to personalized medicine, better brain-computer interfaces, and advances in neurotechnology.
The Georgia Tech group combined two current ML methods into their hybrid model called MRM-GP (Multi-Region Markovian Gaussian Process).
Neuroscientists who use MRM-GP learn more about communications and interactions within the brain. This in turn improves understanding of brain functions and disorders.
“Clinically, MRM-GP could enhance diagnostic tools and treatment monitoring by identifying and analyzing neural activity patterns linked to various brain disorders,” said Weihan Li, the study’s lead researcher.
“Neuroscientists can leverage MRM-GP for its robust modeling capabilities and efficiency in handling large-scale brain data.”
MRM-GP reveals where and how communication travels across brain regions.
The group tested MRM-GP using spike trains and local field potential recordings, two kinds of measurements of brain activity. These tests produced representations that illustrated directional flow of communication among brain regions.
Experiments also disentangled brainwaves, called oscillatory interactions, into organized frequency bands. MRM-GP’s hybrid configuration allows it to model frequencies and phase delays within the latent space of neural recordings.
MRM-GP combines the strengths of two existing methods: the Gaussian process (GP) and linear dynamical systems (LDS). The researchers say that MRM-GP is essentially an LDS that mirrors a GP.
LDS is a computationally efficient and cost-effective method, but it lacks the power to produce representations of the brain. GP-based approaches boost LDS's power, facilitating the discovery of variables in frequency bands and communication directions in the brain.
Converting GP outputs into an LDS is a difficult task in ML. The group overcame this challenge by instilling separability in the model’s multi-region kernel. Separability establishes a connection between the kernel and LDS while modeling communication between brain regions.
Through this approach, MRM-GP overcomes two challenges facing both neuroscience and ML fields. The model helps solve the mystery of intraregional brain communication. It does so by bridging a gap between GP and LDS, a feat not previously accomplished in ML.
“The introduction of MRM-GP provides a useful tool to model and understand complex brain region communications,” said Li, a Ph.D. student in the School of Computational Science and Engineering (CSE).
“This marks a significant advancement in both neuroscience and machine learning.”
Fellow doctoral students Chengrui Li and Yule Wang co-authored the paper with Li. School of CSE Assistant Professor Anqi Wu advises the group.
Each MRM-GP student pursues a different Ph.D. degree offered by the School of CSE. W. Li studies computer science, C. Li studies computational science and engineering, and Wang studies machine learning. The school also offers Ph.D. degrees in bioinformatics and bioengineering.
Wu is a 2023 recipient of the Sloan Research Fellowship for neuroscience research. Her work straddles two of the School’s five research areas: machine learning and computational bioscience.
MRM-GP will be featured at the world’s top conference on ML and artificial intelligence. The group will share their work at the International Conference on Machine Learning (ICML 2024), which will be held July 21-27 in Vienna.
ICML 2024 also accepted for presentation a second paper from Wu’s group intersecting neuroscience and ML. The same authors will present A Differentiable Partially Observable Generalized Linear Model with Forward-Backward Message Passing.
Twenty-four Georgia Tech faculty from the Colleges of Computing and Engineering will present 40 papers at ICML 2024. Wu is one of six faculty representing the School of CSE who will present eight total papers.
The group’s ICML 2024 presentations exemplify Georgia Tech’s focus on neuroscience research as a strategic initiative.
Wu is an affiliated faculty member with the Neuro Next Initiative, a new interdisciplinary program at Georgia Tech that will lead research in neuroscience, neurotechnology, and society. The University System of Georgia Board of Regents recently approved a new neuroscience and neurotechnology Ph.D. program at Georgia Tech.
“Presenting papers at international conferences like ICML is crucial for our group to gain recognition and visibility, facilitates networking with other researchers and industry professionals, and offers valuable feedback for improving our work,” Wu said.
“It allows us to share our findings, stay updated on the latest developments in the field, and enhance our professional development and public speaking skills.”
Visit https://sites.gatech.edu/research/icml-2024 for news and coverage of Georgia Tech research presented at ICML 2024.
Bryant Wine, Communications Officer
bryant.wine@cc.gatech.edu
Some Georgia Tech researchers solve cosmic mysteries such as how supermassive black holes were born — and others now are getting a better, sharper look at those black holes.
The International Conference for High Performance Computing, Networking, Storage, and Analytics, or “Supercomputing” (SC) for short, was held in Denver from November 12-17 and hosted nearly 14,000 attendees. SC is the premier event for advances in algorithms, applications, architectures and networks, clouds and distributed computing, data analytics and visualization, machine learning and HPC, programming systems, system software, and state of the practice in large-scale data storage, deployment and integration. Each year, SC provides a unique opportunity to meet leaders in the field of high-performance computing, including researchers at universities and government labs, and hardware vendors like Intel, AMD, NVIDIA, and Penguin Computing.
The theme for this year’s event was “I am HPC; Impact and Future Direction,” focusing on the ways in which HPC impacts society as well as the community of researchers in HPC. An interdisciplinary cohort comprised of Georgia Tech researchers from the Partnership for an Advanced Computing Environment, the Center for High-Performance Computing, the School of Computational Science and Engineering, the Center for Research into Novel Computing Hierarchies, the Institute for Data Engineering and Science, and the School of Computer Science at various levels of their career were in attendance to present technical talks, participate in workshops and promote HPC research at Georgia Tech with a booth in the exhibit hall.
Georgia Tech teams were well represented across the research themes, including presentations on; large graph analytics, new methods for high-performance data-intensive workloads, challenges presented by the exascale computing, and Large Language Models (LLMs) applications in codesign. Jeffrey Young and Richard Vuduc were feature speakers in the tutorial “Leveraging SmartNICs for HPC Applications” that offered attendees an in-depth exploration of the state-of-the-art for SmartNICs and the emerging software ecosystems supporting them. Georgia Tech was also represented in the Birds of a Feather Co-Hort Panel Discussions on “Software Testing for Scientific Computing in HPC” and “Scientific Software and the People Who Make It Happen: Building Communities of Practice”. Of special mention is the Special Topic Workshop on “Machine Learning with Graphs in High Performance Computing Environments” organized by Richard Vuduc of Georgia Tech as well as Seung-Hwan Lim, Catherine Schuman, and Jose Moreira.
Georgia Tech researchers had numerous discussions with potential collaborators and new partners for initiatives in high performance computing, including conference attendees from universities, government labs, and industry. The team also had a great opportunity to reconnect with numerous alumni, who stopped by the GT booth to tell us about their careers since graduation. Georgia Tech graduates are doing some amazing things in computing hardware, algorithms, and software, with applications across a wide range of engineering and science problems.
Christa M. Ernst - Research Communications Program Manager
christa.ernst@research.gatech.edu
In keeping with a strong strategic focus on AI for the 2023-2024 Academic Year, the Institute for Data Engineering and Science (IDEaS) has announced the winners of its 2023 Seed Grants for Thematic Events in AI and Cyberinfrastructure Resource Grants to support research in AI requiring secure, high-performance computing capabilities. Thematic event awards recipients will receive $8K to support their proposed workshop or series and Cyberinfrastructure winners will receive research support consisting of 600,000 CPU hours on the AMD Genoa Server as well as 36,000 hours of NVIDIA DGX H-100 GPU server usage and 172 TB of secure storage.
Proposed Workshop: “Foundation of scientific AI (Artificial Intelligence) for Optimization of Complex Systems”
Primary PI: Raphael Pestourie, Assistant Professor, School of Computational Science and Engineering
Secondary PI: Peng Chen, Assistant Professor, School of Computational Science and Engineering
Proposed Series: “Guest Lecture Seminar Series on Generative Art and Music”
Primary PI: Gil Weinberg, Professor, School of Music
Title: Human-in-the-Loop Musical Audio Source Separation
Topics: Music Informatics, Machine Learning
Primary PI: Alexander Lerch, Associate Professor, School of Music
Co-PIs: Karn Watcharasupat, Music Informatics Group | Yiwei Ding, Music Informatics Group | Pavan Seshadri, Music Informatics Group
Title: Towards A Multi-Species, Multi-Region Foundation Model for Neuroscience
Topics: Data-Centric AI, Neuroscience
Primary PI: Eva Dyer, Assistant Professor, Biomedical Engineering
Title: Multi-point Optimization for Building Sustainable Deep Learning Infrastructure
Topics: Energy Efficient Computing, Deep Learning, AI Systems OPtimization
Primary PI: Divya Mahajan, Assistant Professor, School of Electrical and Computer Engineering, School of Computer Science
Title: Neutrons for Precision Tests of the Standard Model
Topics: Nuclear/Particle Physics, Computational Physics
Primary PI: Aaron Jezghani - OIT-PACE
Title: Continual Pretraining for Egocentric Video
Primary PI: : Zsolt Kira, Assistant Professor, School of Interactive Computing
Co-PI: Shaunak Halbe, Ph.D. Student, Machine Learning
Title: Training More Trustworthy LLMs for Scientific Discovery via Debating and Tool Use
Topics: Trustworthy AI, Large-Language Models, Multi-Agent Systems, AI Optimization
Primary PIs: Chao Zhang, School of Computational Science and Engineering & Bo Dai, College of Computing
Title: Scaling up Foundation AI-based Protein Function Prediction with IDEaS Cyberinfrastructure
Topics: AI, Biology
Primary PI: Yunan Luo, Assistant Professor, School of Computational Science and Engineering
Christa M. Ernst - Research Communications Program Manager
Robotics | Data Engineering | Neuroengineering
The Institute for Data Engineering and Science, in conjunction with several Interdisciplinary Research Institutes (IRIs) at Georgia Tech, have awarded seven teams of researchers from across the Institute a total of $105,000 in seed funding geared to better position Georgia Tech to perform world-class interdisciplinary research in data science and artificial intelligence development and deployment.
The goals of the funded proposals include identifying prominent emerging research directions on the topic of AI, shaping IDEaS future strategy in the initiative area, building an inclusive and active community of Georgia Tech researchers in the field that potentially include external collaborators, and identifying and preparing groundwork for competing in large-scale grant opportunities in AI and its use in other research fields.
Overview: The goal of this initiative is to bring together expertise in machine learning/AI, high-throughput computing, computational chemistry, and experimental materials synthesis and characterization to accelerate material discovery. Computational chemistry and materials simulations are critical for developing new materials and understanding their behavior and performance, as well as aiding in experimental synthesis and characterization. Machine learning and AI play a pivotal role in accelerating material discovery through data-driven surrogate models, as well as high-throughput and automated synthesis and characterization.
Overview: Zhigang Jiang is currently leading an initiative within IMAT entitled “Quantum responses of topological and magnetic matter” to nurture multi-PI projects. By crosscutting the IMAT initiative with this IDEAS call, we propose to support and feature the applications of AI on predictive and inverse problems in quantum materials. Understanding the limit and capabilities of AI methodologies is a huge barrier of entry for Physics students, because researchers in that field already need heavy training in quantum mechanics, low-temperature physics and chemical synthesis. Our most pressing need is for our AI inclined quantum materials students to find a broader community to engage with and learn. This is the primary problem we aim to solve with this initiative.
Overview: Our objective is to use large language models (LLMs) in conjunction with AI algorithms to identify effective driver proteins, develop screening algorithms that target appropriate binding sites while avoiding deleterious ones, and consider bioavailability and drug resistance factors. LLMs can rapidly analyze vast amounts of information from literature and bioinformatics tools, generating hypotheses and suggesting molecular modifications. By bridging multiple disciplines such as biology, chemistry, and pharmacology, LLMs can provide valuable insights from diverse sources, assisting researchers in making informed decisions. Our aim is to establish a first-in-class, LLM driven research initiative at Georgia Tech that focuses on designing highly effective small molecule libraries to treat challenging diseases. This initiative will go beyond existing AI approaches to molecule generation, which often only consider simple properties like hydrogen bonding or rely on a limited set of proteins to train the LLM and therefore lack generalizability. As a result, this initiative is expected to consistently produce safe and effective disease-specific molecules.
Overview: We are committed to building a team of interdisciplinary & transdisciplinary researchers and practitioners with a shared goal: developing a new framework which model future climatic variations and the interconnected and interdependent energy infrastructure network as complex systems. To achieve this, we will harness the power of cutting-edge climate model outputs, sourced from the Coupled Model Intercomparison Project (CMIP), and integrate approaches from Machine Learning and Deep Learning models. This strategic amalgamation of data and techniques will enable us to gain profound insights into the intricate web of future climate-change-induced extreme weather conditions and their immediate and long-term ramifications on energy infrastructure networks. The seed grant from IDEaS stands as the crucial catalyst for kick-starting this ambitious endeavor. It will empower us to form a collaborative and inclusive community of GT researchers hailing from various domains, including City and Regional Planning, Earth and Atmospheric Science, Computer Science and Electrical Engineering, Civil and Environmental Engineering etc. By drawing upon the wealth of expertise and perspectives from these diverse fields, we aim to foster an environment where innovative ideas and solutions can flourish. In addition to our internal team, we also have plans to collaborate with external partners, including the World Bank, the Stanford Doerr School of Sustainability, and the Berkeley AI Research Initiative, who share our vision of addressing the complex challenges at the intersection of climate and energy infrastructure.
Overview: Our research team envisions a significant trend in the exploration of AI applications for urban flooding hazard forecasting. Georgia Tech possesses a wealth of interdisciplinary expertise, positioning us to make a pioneering contribution to this burgeoning field. We aim to harness the combined strengths of Georgia Tech's experts in civil and environmental engineering, atmospheric and climate science, and data science to chart new territory in this emerging trend. Furthermore, we envision the potential extension of our research efforts towards the development of a real-time hazard forecasting application. This application would incorporate adaptation and mitigation strategies in collaboration with local government agencies, emergency management departments, and researchers in computer engineering and social science studies. Such a holistic approach would address the multifaceted challenges posed by urban flooding. To the best of our knowledge, Georgia Tech currently lacks a dedicated team focused on the fusion of AI and climate/flood research, making this initiative even more pioneering and impactful.
Overview: Most asset management and recycling-use technology has not changed for decades. The use of bar codes and RFID has provided some benefits, such as for retail returns management. Automated sorting of recyclables using magnets, eddy currents, and laser plastics identification has improved municipal recycling. Yet the overall field has been challenged by not-quite-easy-enough identification of products in use or at end of life. AI approaches, including computer vision, data fusion, and machine learning provide the additional capability to make asset management and product recycling easy enough to be nearly autonomous. Georgia Tech is well suited to lead in the development of this application. With its strength in machine learning, robotics, sustainable business, supply chains and logistics, and technology commercialization, Georgia Tech has the multi-disciplinary capability to make this concept a reality; in research and in commercial application.
Overview: Artistic human movement at large, stands at the precipice of a data-driven renaissance. By leveraging novel tools, we can usher in a transparent, data-driven, and accessible training environment. The potential ramifications extend beyond dance. As sports analytics have reshaped our understanding of athletic prowess, a similar approach to dance could redefine our comprehension of human movement, with implications spanning healthcare, construction, rehabilitation, and active aging. Georgia Tech, with its prowess in AI, HCI, and biomechanics is primed to lead this exploration. To actualize this vision, we propose the following research questions with ballet as a prime example of one of the most complex types of artistic movements: 1) What kinds of data - real-time kinematic, kinetic, biomechanical, etc. captured through accessible off-the-shelf technologies, are essential for effective AI assessment in ballet education for young adults?; 2) How can we design and develop an end-to-end ML architecture that assesses artistic and technical performance?; 3) What feedback elements (combination of timing, communication mode, feedback nature, polarity, visualization) are most effective for AI- based dance assessment?; and 4) How does AI-assisted feedback enhance physical wellness, artistic performance, and the learning process in young athletes compared to traditional methods?
Christa M. Ernst | Research Communications Program Manager
Robotics | Data Engineering | Neuroengineering
christa.ernst@research.gatech.edu