Pedro Domingos is a professor of computer science at the University of Washington and the author of "The Master Algorithm". He is a winner of the SIGKDD Innovation Award, the highest honor in data science. He is a Fellow of the Association for the Advancement of Artificial Intelligence, and has received a Fulbright Scholarship, a Sloan Fellowship, the National Science Foundation’s CAREER Award, and numerous best paper awards. His research spans a wide variety of topics in machine learning, artificial intelligence, and data science, including scaling learning algorithms to big data, maximizing word of mouth in social networks, unifying logic and probability, and deep learning.

If you are interested in attending please RSVP via the following link as space is limited.  Please contact the event organizer, Jeff Young, if you have any questions.

• https://www.eventbrite.com/e/nvidia-gpu-deep-learning-symposium-tickets-33441905640

• Time: Monday, April 17, 2017, from 11am – 7pm.
• Location: Technology Square Research Building, First Floor (Banquet Rooms TSRB 132/134) on Georgia Tech Campus (Map | Parking).
• Schedule Summary (Detailed schedule)
  • 11 am: Introduction (Featuring Center leadership and Dean of the College of Computing)
  • 11:30am: Lunch
  • 12:00n – 1:00pm: Panel with Georgia Tech’s Interdisciplinary Research Institutes Leadership
  • 1:00pm – 5:00pm: Quickfire talks (5 minutes each) from 30+ faculty from all Schools / Colleges of GA Tech (listed here)
  • 5:00pm – 7:00pm: Reception and Poster Session
    • Including statements from Steve Cross, (EVPR, GA Tech) and others.
    • Student Posters (32 and counting … listed here)
• RSVP required HERE.

I will present a framework for addressing this challenge based on the idea of embedding algorithm steps into nonlinear spaces, and learn these embedded algorithms from problem instances via either direct supervision or reinforcement learning. In contrast to traditional algorithm design where every step in an algorithm is prescribed by experts, the embedding design will delegate some difficult algorithm choices to nonlinear learning models so as to avoid either large memory requirement, restricted assumptions on the data, or limited design space exploration. I will illustrate the benefit of this new design framework using large scale real world data, including a materials discovery problem, a recommendation problem over dynamic information networks, and a problem of learning combinatorial algorithms over graphs. The learned algorithms can reduce memory usage and runtime by orders of magnitude, and sometimes result in drastic improvement in predictive performance.

Bio: Le Song is an Associate Professor in the Department of Computational Science and Engineering, College of Computing, and an Associate Director of the Center for Machine Learning, Georgia Institute of Technology. He received his Ph.D. in Machine Learning from University of Sydney and NICTA in 2008, and then conducted his post-doctoral research in the Department of Machine Learning, Carnegie Mellon University, between 2008 and 2011. Before he joined Georgia Institute of Technology in 2011, he was a research scientist at Google briefly. His principal research direction is machine learning, especially nonlinear models, such as kernel methods and deep learning, and probabilistic graphical models for large scale and complex problems, arising from artificial intelligence, network analysis, computational biology and other interdisciplinary domains. He is the recipient of the Recsys’16 Deep Learning Workshop Best Paper Award, AISTATS’16 Best Student Paper Award, IPDPS’15 Best Paper Award, NSF CAREER Award’14, NIPS’13 Outstanding Paper Award, and ICML’10 Best Paper Award. He has also served as the area chair for many leading machine learning and AI conferences such as ICML, NIPS, AISTATS, AAAI and IJCAI, and the action editor for JMLR.

Liao’s expertise in artificial intelligence — specifically machine learning and deep learning — earned her the invitation to give one of three addresses at the March AMS meeting. Liao will present some statistical learning theory of deep neural networks where data are concentrated on or near a low-dimensional manifold. (Low- and high-dimensions refer to the number of key features versus observations in a dataset.)

“When data are sampled on a low-dimensional manifold, the sample complexity crucially depends on the intrinsic dimension of the manifold, which demonstrates that deep neural networks are adaptive to low-dimensional geometric structures in data,” Liao says.

Liao is organizing a separate special session on mathematical advances in machine learning during the meeting. Two School of Mathematics visiting assistant professors are also organizing special sessions: Papri Dey’s session is on combinatorics of geometry in polynomials, and Austin Christian’s will focus on topological interactions of contact and symplectic manifolds.

“For Wenjing to give a presentation like this at an AMS Spring Southeastern Sectional meeting, at this early stage of her career, speaks volumes about the impact her research into machine learning and especially low dimensional loci within data is having on the field,” says Michael Wolf, professor and chair of the School of Mathematics.

“It is my honor and privilege to speak at the Sectional Meeting,” Liao says. “This is a great opportunity for me to share our research on deep learning theory with the community.”

Georgia Tech and AMS Connections

Last spring the School of Mathematics hosted the AMS 2023 Southeastern Sectional Meeting, which saw alumna Blair Dowling Sullivan (MATH 2003) deliver one of four invited addresses. 

Several former and current faculty and graduate students and postdoctoral scholars also organized special sessions and lectures during the AMS 2023 meeting, including Christine Heitsch, Kevin Shu, Mehrdad Ghadiri, Brandon Jerome Legried, Gong Chen, Ryan Dickmann, Abdoul Karim Sane, Dan Margalit, Benjamin Jaye, Naga Manasa Vempati, Galyna Livshyts, Orli Herscovici, Anton Bernshteyn, Matthew Powell, Burak Hatinoglu, Zhiyu Wang, Xingxing Yu, Miriam Kuzbary, Jon Simone, and Nur Saglam.

“We were excited and proud to host approximately 800 mathematicians on campus for the March AMS Southeastern Sectional Meeting,” Wolf shares. “It cemented for us just how important these regional conferences are in our discipline. We got to hear the latest on issues like combinatorics, the intersection of math and biology, and quantum systems. We look forward to continuing our regional relationships at the upcoming meeting at Florida State.”

Wolf says other current and past School of Mathematics faculty who have spoken at previous AMS Sectional meetings include Professors Matt Baker, Greg Blekherman, Rafael de la Llave, John Etnyre, Heitsch, Michael Lacey, Michael Loss, Margalit, Dana Randall, Prasad Tetali, Robin Thomas, Wolf, and XingXing Yu.

About Wenjing Liao 

Liao received her Ph.D. in Applied Mathematics at the University of California, Davis, and joined Georgia Tech in 2017. In addition to machine learning, her research interests include imaging, signal processing, and high-dimensional data analysis.

Liao was also part of the Georgia Tech contingent attending this summer’s International Conference for Machine Learning in Honolulu. She recently won a U.S. Department of Energy (DOE) Early Career Award for her work on how deep learning might be leveraged to make mathematical advances in achieving more efficient modeling techniques.

Liao’s machine learning research also won her a National Science Foundation (NSF) Faculty Early Career Development Program (CAREER) Award in 2022.

Machine learning versus deep learning

While machine learning relies on algorithms to search for predictability and patterns in sets of structured data, deep learning algorithms imitate how the brain’s neural networks work to find patterns in large sets of unstructured data. 

“It is a common belief that deep neural networks are capable of learning various geometric structures hidden in data sets,” Liao said. “One of the central interests in deep learning theory is to understand why deep neural networks are successful, and how they utilize low-dimensional data structures.” 

Liao’s upcoming AMS address, “Exploiting Low-Dimensional Data Structures in Deep Learning,” will discuss how deep learning leverages low dimensional data structures and achieves great success.

Ph.D. student Alec Helbling is the creator of ManimML, a tool that renders common ML concepts into animation. This development will enable new ML technologies by allowing designers to see and share their work in action. 

Helbling presented ManimML at IEEE VIS, the world’s highest-rated conference for visualization research and second-highest rated for computer graphics. It received so much praise at the conference that it won the venue’s prize for best poster. 

“I was quite surprised and honored to have received this award,” said Helbling, who is advised by School of Computational Science and Engineering Associate Professor Polo Chau.

“I didn't start ManimML with the intention of it becoming a research project, but because I felt like a tool for communicating ML architectures through animation needed to exist.”

[RELATED: Polo Chau is One of Three College of Computing Faculty to Receive 2023 Google Award for Inclusion Research]

ManimML uses animation to show ML developers how their algorithms work. Not only does the tool allow designers to watch their projects come to life, but they can also explain existing and new ML techniques to broad audiences, including non-experts.

ManimML is an extension of the Manim Community library, a Python tool for animating mathematical concepts. ManimML connects to the library to offer a new capability that animates ML algorithms and architectures.

Helbling chose familiar platforms like Python and Manim to make the tool accessible to large swaths of users varying in skill and experience. Enthusiasts and experts alike can find practical use in ManimML considering today’s widespread interest and application of ML.

“We know that animation is an effective means of instruction and learning,” Helbling said. “ManimML offers that ability for ML practitioners to easily communicate how their systems work, improving public trust and awareness of machine learning.”

ManimML overcomes what has been an elusive approach to visualizing ML algorithms. Current techniques require developers to create custom animations for every specific algorithm, often needing specialized software and experience.

ManimML streamlines this by producing animations of common ML architectures coded in Python, like neural networks.

A user only needs to specify a sequence of neural network layers and their respective hyperparameters. ManimML then constructs an animation of the entire network.

“To use ManimML, you simply need to specify an ML architecture in code, using a syntax familiar to most ML professionals,” Helbling said. “Then it will automatically generate an animation that communicates how the system works.”

ManimML ranked as the best poster from a field of 49 total presentations. IEEE VIS 2023 occurred Oct. 22-27 in Melbourne, Australia. This event marks the first time IEEE held the conference in the Southern Hemisphere.

ManimML has more than 23,000 downloads and a demonstration on social media has hundreds of thousands of views.

ManimML is open source and available at: https://github.com/helblazer811/ManimML

The workshop focused AI technology toward optimization of complex systems. Presentations of climatological and electromagnetic simulations showed these techniques resulted in more efficient and accurate computer modeling. The workshop also progressed AI research itself since AI models typically are not well-suited for optimization tasks.

The School of Computational Science and Engineering (CSE) and Institute for Data Engineering and Science jointly sponsored the workshop.

School of CSE Assistant Professors Peng Chen and Raphaël Pestourie led the workshop’s organizing committee and moderated the workshop’s two panel discussions. The duo also pitched their own research, highlighting potential of scientific AI.

Chen shared his work on derivative-informed neural operators (DINOs). DINOs are a class of neural networks that use derivative information to approximate solutions of partial differential equations. The derivative enhancement results in neural operators that are more accurate and efficient. 

During his talk, Chen showed how DINOs makes better predictions with reliable derivatives. These have potential to solve data assimilation problems in weather and flooding prediction. Other applications include allocating sensors for early tsunami warnings and designing new self-assembly materials.

All these models contain elements of uncertainty where data is unknown, noisy, or changes over time. Not only is DINOs a powerful tool to quantify uncertainty, but it also requires little training data to become functional.

“Recent advances in AI tools have become critical in enhancing societal resilience and quality, particularly through their scientific uses in environmental, climatic, material, and energy domains,” Chen said. 

“These tools are instrumental in driving innovation and efficiency in these and many other vital sectors.”

[Related: Machine Learning Key to Proposed App that Could Help Flood-prone Communities]

One challenge in studying complex systems is that it requires many simulations to generate enough data to learn from and make better predictions. But with limited data on hand, it is costly to run enough simulations to produce new data.

At the workshop, Pestourie presented his physics-enhanced deep surrogates (PEDS) as a solution to this optimization problem. 

PEDS employs scientific AI to make efficient use of available data while demanding less computational resources. PEDS demonstrated to be up to three times more accurate than models using neural networks while needing less training data by at least a factor of 100. 

PEDS yielded these results in tests on diffusion, reaction-diffusion, and electromagnetic scattering models. PEDS performed well in these experiments geared toward physics-based applications because it combines a physics simulator with a neural network generator.

“Scientific AI makes it possible to systematically leverage models and data simultaneously,” Pestourie said. “The more adoption of scientific AI there will be by domain scientists, the more knowledge will be created for society.”

[Related: Technique Could Efficiently Solve Partial Differential Equations for Numerous Applications]

Study and development of AI applications at these scales require use of the most powerful computers available. The workshop invited speakers from national laboratories who showcased supercomputing capabilities available at their facilities. These included Oak Ridge National Laboratory, Sandia National Laboratories, and Pacific Northwest National Laboratory.

The workshop hosted Georgia Tech faculty who represented the Colleges of Computing, Design, Engineering, and Sciences. Among these were workshop co-organizers Yan Wang and Ebeneser Fanijo. Wang is a professor in the George W. Woodruff School of Mechanical Engineering and Fanjio is an assistant professor in the School of Building Construction.

The workshop welcomed academics outside of Georgia Tech to share research occurring at their institutions. These speakers hailed from Emory University, Clemson University, and the University of California, Berkeley.

A new multidisciplinary research effort co-led by Anna (Anya) Ivanova, assistant professor in the School of Psychology at Georgia Tech, alongside Kyle Mahowald, an assistant professor in the Department of Linguistics at the University of Texas at Austin, is working to uncover just that.

Their results could lead to innovative AIs that are more similar to the human brain than ever before — and also help neuroscientists and psychologists who are unearthing the secrets of our own minds. 

The study, “Dissociating Language and Thought in Large Language Models,” is published this week in the scientific journal Trends in Cognitive Sciences. The work is already making waves in the scientific community: an earlier preprint of the paper, released in January 2023, has already been cited more than 150 times by fellow researchers. The research team has continued to refine the research for this final journal publication. 

“ChatGPT became available while we were finalizing the preprint,” Ivanova explains. “Over the past year, we've had an opportunity to update our arguments in light of this newer generation of models, now including ChatGPT.”

Form versus function

The study focuses on large language models (LLMs), which include AIs like ChatGPT. LLMs are text prediction models, and create writing by predicting which word comes next in a sentence — just like how a cell phone or email service like Gmail might suggest what next word you might want to write. However, while this type of language learning is extremely effective at creating coherent sentences, that doesn’t necessarily signify intelligence.

Ivanova’s team argues that formal competence — creating a well-structured, grammatically correct sentence — should be differentiated from functional competence — answering the right question, communicating the correct information, or appropriately communicating. They also found that while LLMs trained on text prediction are often very good at formal skills, they still struggle with functional skills.

“We humans have the tendency to conflate language and thought,” Ivanova says. “I think that’s an important thing to keep in mind as we're trying to figure out what these models are capable of, because using that ability to be good at language, to be good at formal competence, leads many people to assume that AIs are also good at thinking — even when that's not the case.

It's a heuristic that we developed when interacting with other humans over thousands of years of evolution, but now in some respects, that heuristic is broken,” Ivanova explains.

The distinction between formal and functional competence is also vital in rigorously testing an AI’s capabilities, Ivanova adds. Evaluations often don’t distinguish formal and functional competence, making it difficult to assess what factors are determining a model’s success or failure. The need to develop distinct tests is one of the team’s more widely accepted findings, and one that some researchers in the field have already begun to implement.

Creating a modular system

While the human tendency to conflate functional and formal competence may have hindered understanding of LLMs in the past, our human brains could also be the key to unlocking more powerful AIs. 

Leveraging the tools of cognitive neuroscience while a postdoctoral associate at Massachusetts Institute of Technology (MIT), Ivanova and her team studied brain activity in neurotypical individuals via fMRI, and used behavioral assessments of individuals with brain damage to test the causal role of brain regions in language and cognition — both conducting new research and drawing on previous studies. The team’s results showed that human brains use different regions for functional and formal competence, further supporting this distinction in AIs. 

“Our research shows that in the brain, there is a language processing module and separate modules for reasoning,” Ivanova says. This modularity could also serve as a blueprint for how to develop future AIs.

“Building on insights from human brains — where the language processing system is sharply distinct from the systems that support our ability to think — we argue that the language-thought distinction is conceptually important for thinking about, evaluating, and improving large language models, especially given recent efforts to imbue these models with human-like intelligence,” says Ivanova’s former advisor and study co-author Evelina Fedorenko, a professor of brain and cognitive sciences at MIT and a member of the McGovern Institute for Brain Research.

Developing AIs in the pattern of the human brain could help create more powerful systems — while also helping them dovetail more naturally with human users. “Generally, differences in a mechanism’s internal structure affect behavior,” Ivanova says. “Building a system that has a broad macroscopic organization similar to that of the human brain could help ensure that it might be more aligned with humans down the road.” 

In the rapidly developing world of AI, these systems are ripe for experimentation. After the team’s preprint was published, OpenAI announced their intention to add plug-ins to their GPT models. 

“That plug-in system is actually very similar to what we suggest,” Ivanova adds. “It takes a modularity approach where the language model can be an interface to another specialized module within a system.” 

While the OpenAI plug-in system will include features like booking flights and ordering food, rather than cognitively inspired features, it demonstrates that “the approach has a lot of potential,” Ivanova says.

The future of AI — and what it can tell us about ourselves

While our own brains might be the key to unlocking better, more powerful AIs, these AIs might also help us better understand ourselves. “When researchers try to study the brain and cognition, it's often useful to have some smaller system where you can actually go in and poke around and see what's going on before you get to the immense complexity,” Ivanova explains.

However, since human language is unique, model or animal systems are more difficult to relate. That's where LLMs come in. 

“There are lots of surprising similarities between how one would approach the study of the brain and the study of an artificial neural network” like a large language model, she adds. “They are both information processing systems that have biological or artificial neurons to perform computations.” 

In many ways, the human brain is still a black box, but openly available AIs offer a unique opportunity to see the synthetic system's inner workings and modify variables, and explore these corresponding systems like never before.

“It's a really wonderful model that we have a lot of control over,” Ivanova says. “Neural networks — they are amazing.”

Along with Anna (Anya) Ivanova, Kyle Mahowald, and Evelina Fedorenko, the research team also includes Idan Blank (University of California, Los Angeles), as well as Nancy Kanwisher and Joshua Tenenbaum (Massachusetts Institute of Technology).

DOI: https://doi.org/10.1016/j.tics.2024.01.011

Researcher Acknowledgements

For helpful conversations, we thank Jacob Andreas, Alex Warstadt, Dan Roberts, Kanishka Misra, students in the 2023 UT Austin Linguistics 393 seminar, the attendees of the Harvard LangCog journal club, the attendees of the UT Austin Department of Linguistics SynSem seminar, Gary Lupyan, John Krakauer, members of the Intel Deep Learning group, Yejin Choi and her group members, Allyson Ettinger, Nathan Schneider and his group members, the UT NLL Group, attendees of the KUIS AI Talk Series at Koç University in Istanbul, Tom McCoy, attendees of the NYU Philosophy of Deep Learning conference and his group members, Sydney Levine, organizers and attendees of the ILFC seminar, and others who have engaged with our ideas. We also thank Aalok Sathe for help with document formatting and references.

Funding sources

Anna (Anya) Ivanova was supported by funds from the Quest Initiative for Intelligence. Kyle Mahowald acknowledges funding from NSF Grant 2104995. Evelina Fedorenko was supported by NIH awards R01-DC016607, R01-DC016950, and U01-NS121471 and by research funds from the Brain and Cognitive Sciences Department, McGovern Institute for Brain Research, and the Simons Foundation through the Simons Center for the Social Brain.

Dr. Polo Chau, an associate professor in the School of Computational Science and Engineering at Georgia Tech and associate director of the Master of Science Analytics program, and Didier Contis, executive director of Academic Technology, Innovation, and Research Computing within OIT, have embarked on an ambitious project to create a short, asynchronous course designed to equip IT professionals and business citizen developers with the knowledge to apply LLMs in enterprise environments.

The course will initially provide a focused, online tutorial that covers LLMs, including advanced AI techniques such as fine-tuning processes. A future version of the tutorial will focus on Retrieval Augmented Generation (RAG) and fine-tuning processes. RAG enables language models to dynamically retrieve external data, improving response accuracy, while fine-tuning allows for the customization of these models to meet specific business needs, from unique vocabularies to specialized tasks.

Two graduate students from Chau’s research group are developing the curriculum to offer hands-on learning experiences. The course will demonstrate the practical application of LLMs, giving learners the tools to innovate and automate within their industries.

The development of this course is timely, given the growing demand for AI capabilities in business environments. It addresses the critical need for skills in managing and deploying LLMs, potentially leading to increased efficiency and the creation of sophisticated user experiences.

This educational initiative is expected to play a crucial role in democratizing AI, making advanced AI techniques accessible to a wider audience beyond AI researchers and data scientists. In doing so, Chau and Contis are not only contributing to the professional development of individuals but also fostering the growth of an ecosystem that embraces technological progress and innovation.

The result of their efforts is a future in which AI integration is not a complex challenge, but a standard practice.

The minor’s Ivan Allen College track will give students the skills necessary to ethically and effectively use such tools in digital humanities and computational social sciences contexts, said Kaye Husbands Fealing, dean and Ivan Allen Jr. Chair in the Ivan Allen College of Liberal Arts.

“Artificial intelligence and machine learning have important applications in nearly every social science and humanities discipline. So, it is essential for tomorrow’s leaders in these fields to have a strong foundation not just in how to use these tools, but also when and why, and to understand the difficult ethical and policy issues as we strive for outcomes that improve lives,” Husbands Fealing said.

In fact, the AI ethics and policy course is the only required course for all students taking the minor — whether they choose the Ivan Allen College track or the College of Engineering track, said Shatakshee Dhongde, associate dean for academic affairs in the Ivan Allen College who spearheaded the minor for the Ivan Allen College.

“The fact that every student seeking this minor will be required to take an Ivan Allen College policy course highlights the importance Georgia Tech places on graduating leaders with an understanding of the ethical implications of this ever-changing technology,” Dhongde said.

In addition to the ethics class, the Ivan Allen College track requires one course in probability and statistics, an introduction to AI and machine learning, and two electives. Students can choose from classes in linguistics, philosophy of computation, the role of technology such as AI in international affairs, and machine learning in economics.

“This is truly a collegewide minor with courses from several schools in the Ivan Allen College,” Dhongde said.

Ivan Allen College majors can also pursue the College of Engineering track, which provides additional exposure to the technical fundamentals of AI and machine learning and their applications in biomedical engineering, process engineering, mechanical engineering, and robotics.

“We wanted to ensure that this minor is not just about the technical aspects of creating or using an AI platform, but also about thinking through the ways in which AI can be used to meet global challenges,” Dhongde said.

The School of Public Policy has offered an AI Ethics and Policy course since 2023. Taught by Associate Professor Justin Biddle, the class prepares students to think critically about AI's impact on humanity and to contribute to AI governance and policy.

"AI systems are value-laden because they're human creations," Biddle said at the time. "Humans generate, design, develop, distribute, and monitor AI systems. Human decisions are made all along the way, and those human decisions, reflecting our values, impact AI systems in a very consequential way."

The minor is open to all students pursuing degrees from the Ivan Allen College or the College of Engineering. In addition to Public Policy, the School of Economics, the School of Literature, Media, and Communication, the School of Modern Languages, and the Sam Nunn School of International Affairs will contribute to the minor.

Recently, School of Electrical and Computer Engineering (ECE) Professor Chris Rozell traveled to Washington, D.C. to share the impact of his BRAIN Initiative research with U.S. Congressional offices — and offer insights on how critical this program is to society. The briefing took on a particular urgency because BRAIN Initiative funding was cut over 40% this year, and future funding appears to be in jeopardy in the current federal budget climate. 

“The millions of patients suffering with intractable neurologic disorders and mental illness deserve a moonshot to develop new solutions for their conditions,” said Rozell, who also holds the Julian T. Hightower Chair in ECE and serves on the executive committee for Georgia Tech’s Neuro Next Initiative. “You can't get to the moon with a paper plane, and you can’t get there without a map. The BRAIN Initiative is a vital program because it's one of the few places that brings together interdisciplinary teams that include the scientists who have been building maps of brain circuits and the engineers who have been building rockets to understand and intervene with those circuits. 

“I'm proud to have had the chance to represent not only our own research, but the incredible community here at Georgia Tech and around the country working to understand many different aspects of the brain, developing new neurotechnologies, and advancing therapies for neurologic disorders.” 

Interdisciplinary impacts 

“The main message we presented to Congress is that the interdisciplinary combination of rigorous science and technical innovation can have enormous societal impact over the next few decades,” said Rozell. 

A stark example of that impact was published in Nature this past fall. In this research, Rozell and his collaborators at the Icahn School of Medicine at Mount Sinai and Emory University School of Medicine identified the first known biomarker of disease recovery with deep brain stimulation in treatment-resistant depression. 

“The fact that an engineer can advance clinical therapies is a testament to the new era we're in,” says Rozell, “where disciplinary boundaries are fading, and technological innovation accelerates our scientific and translational breakthroughs.” 

This research served as a focal point of the congressional briefing, where Rozell presented with BRAIN Initiative Director John J. Ngai, clinical collaborators, and a family whose lives have been transformed by this work.  

“Events like last week are dream come true,” shared Jon Nelson, who was treated with deep brain stimulation as part of the study and presented with Rozell in D.C. After living through 10 years of debilitating, treatment-resistant depression, Nelson says “remission of depression still doesn't feel real. It's been a year and a half, and I still am in awe every single day. 

“The fact that I have come out of this study and found that the disease is purely an electrical deficiency in my brain has fueled me to completely pulverize the stigma of mental illness,” Nelson explained. “When you have an opportunity to go speak to Congress — that’s about as great of a platform as you can get for that. Being able to put a face to what the BRAIN Initiative funding can do for people was just amazing.” 

When meeting with local representatives, Rozell also relayed his work as co-executive leader of the Neuro Next Initiative, a budding Interdisciplinary Research Institute at Georgia Tech. 

“I was thrilled to highlight that Georgia Tech is leading the charge with the Neuro Next Initiative, which will evolve into a full Interdisciplinary Research Institute in 2025,” said Rozell. “Georgia Tech has the ingredients to become a leading center for modern technology-driven interdisciplinary brain research and workforce development. 

“This visit was a reminder to me that research funding is not guaranteed and it’s important to keep communicating the critical value that research plays in advancing our understanding, training our workforce, fueling our economy, and ultimately making a better tomorrow for society.” 

Farsight is a tool designed for developers who use large language models (LLMs) to create applications powered by artificial intelligence (AI). Farsight alerts prototypers when they write LLM prompts that could be harmful and misused.

Downstream users can expect to benefit from better quality and safer products made with Farsight’s assistance. The tool’s lasting impact, though, is that it fosters responsible AI awareness by coaching developers on the proper use of LLMs.

Machine Learning Ph.D. candidate Zijie (Jay) Wang is Farsight’s lead architect. He will present the paper at the upcoming Conference on Human Factors in Computing Systems (CHI 2024). Farsight ranked in the top 5% of papers accepted to CHI 2024, earning it an honorable mention for the conference’s best paper award.

“LLMs have empowered millions of people with diverse backgrounds, including writers, doctors, and educators, to build and prototype powerful AI apps through prompting. However, many of these AI prototypers don’t have training in computer science, let alone responsible AI practices,” said Wang.

“With a growing number of AI incidents related to LLMs, it is critical to make developers aware of the potential harms associated with their AI applications.”

Wang referenced an example when two lawyers used ChatGPT to write a legal brief. A U.S. judge sanctioned the lawyers because their submitted brief contained six fictitious case citations that the LLM fabricated.

With Farsight, the group aims to improve developers’ awareness of responsible AI use. It achieves this by highlighting potential use cases, affected stakeholders, and possible harm associated with an application in the early prototyping stage. 

A user study involving 42 prototypers showed that developers could better identify potential harms associated with their prompts after using Farsight. The users also found the tool more helpful and usable than existing resources. 

Feedback from the study showed Farsight encouraged developers to focus on end-users and think beyond immediate harmful outcomes.

“While resources, like workshops and online videos, exist to help AI prototypers, they are often seen as tedious, and most people lack the incentive and time to use them,” said Wang.

“Our approach was to consolidate and display responsible AI resources in the same space where AI prototypers write prompts. In addition, we leverage AI to highlight relevant real-life incidents and guide users to potential harms based on their prompts.”

Farsight employs an in-situ user interface to show developers the potential negative consequences of their applications during prototyping. 

Alert symbols for “neutral,” “caution,” and “warning” notify users when prompts require more attention. When a user clicks the alert symbol, an awareness sidebar expands from one side of the screen. 

The sidebar shows an incident panel with actual news headlines from incidents relevant to the harmful prompt. The sidebar also has a use-case panel that helps developers imagine how different groups of people can use their applications in varying contexts.

Another key feature is the harm envisioner. This functionality takes a user’s prompt as input and assists them in envisioning potential harmful outcomes. The prompt branches into an interactive node tree that lists use cases, stakeholders, and harms, like “societal harm,” “allocative harm,” “interpersonal harm,” and more.

The novel design and insightful findings from the user study resulted in Farsight’s acceptance for presentation at CHI 2024.

CHI is considered the most prestigious conference for human-computer interaction and one of the top-ranked conferences in computer science.

CHI is affiliated with the Association for Computing Machinery. The conference takes place May 11-16 in Honolulu.

Wang worked on Farsight in Summer 2023 while interning at Google + AI Research group (PAIR).

Farsight’s co-authors from Google PAIR include Chinmay Kulkarni, Lauren Wilcox, Michael Terry, and Michael Madaio. The group possesses closer ties to Georgia Tech than just through Wang.

Terry, the current co-leader of Google PAIR, earned his Ph.D. in human-computer interaction from Georgia Tech in 2005. Madaio graduated from Tech in 2015 with a M.S. in digital media. Wilcox was a full-time faculty member in the School of Interactive Computing from 2013 to 2021 and serves in an adjunct capacity today.

Though not an author, one of Wang’s influences is his advisor, Polo Chau. Chau is an associate professor in the School of Computational Science and Engineering. His group specializes in data science, human-centered AI, and visualization research for social good.  

“I think what makes Farsight interesting is its unique in-workflow and human-AI collaborative approach,” said Wang. 

“Furthermore, Farsight leverages LLMs to expand prototypers’ creativity and brainstorm a wide range of use cases, stakeholders, and potential harms.”

"This application problem is indeed significant and worthy of serious consideration. While you may assert that I lack the capability to resolve it, it remains undeniable that this issue holds considerable importance."  

With this wisdom imparted by his advisor, Professor Jianjun Shi, Jianjun Shi, Shancong Mou started his academic journey on developing Artificial Intelligence (AI) and Machine Learning (ML)-enabled data fusion methodologies aimed at addressing real and significant engineering challenges. 

As a fan of both the epic Star Wars saga and machine learning, Mou, expressed his enthusiasm for leveraging the force of ML for quality and productivity improvement in advanced manufacturing systems. 

"If a problem remains unsolved, it highlights both its complexity and the urgent need for creative solutions." 

Jedi-Level Precision 

From the iconic lightsaber to the X-wing aircraft, Mou shared his fascination with Star Wars’ depiction of advanced control systems and aircraft maneuverability.  

“The spacecraft in those films maneuver with remarkable precision, navigating the narrow corridors of the Death Star effortlessly. Crafting such aircraft would entail integrating millions of intricate parts manufactured to the highest standards. Achieving such precision in manufacturing, coupled with stringent quality control, would indeed be challenging but groundbreaking.” 

Much like a Jedi, governed by a problem-solving philosophy and mindset, one of Mou’s research project explores physics-informed machine learning for the control and design optimization of complex engineering systems.  

One major application is the reduction of variation in fuselage assembly processes, a critical step in the manufacturing process of modern airplanes, such as the Boeing 787.[1] 

In another research avenue, Mou innovates with generative models, specifically generative adversarial networks (GANs), to learn and interpret the underlying patterns of normal signals.  

This approach, termed 'robust GAN inversion', transcends traditional statistical methods to reconstruct signals from corruption, offering a distributional-assumption-free perspective, which provides a tool for unsupervised fine-grained anomaly detection.  

These elements are crucial in high-value and safety-critical industrial applications, such as personal electronic manufacturing process quality monitoring. 

Advanced Sensors & ML: A New Hope for Manufacturing  

The synergy between increasingly advanced sensor capabilities and the development of cutting-edge ML methodologies acts as a crucial factor in achieving unprecedented levels of product detail monitoring and defect detection.  

Mou noted, "Quality and productivity improvement is the goal of my research. The development of sensing technology offers new opportunities and challenges for further adopting/developing advanced ML algorithms to solve this problem." 

Mou’s vision for the integration of ML in IE, mirrors the innovative spirit seen in the Star Wars saga, potentially leading towards a future where technology and human expertise converge to create smarter, cleaner and more efficient manufacturing systems. 

Author: Atharva Anand Dave 

In this role, Professor Davenport will serve as the School’s primary representative on all matters related to graduate academics. He will oversee the graduate affairs team, lead the graduate admission process, develop programs and assessments, and collaborate closely with the chair and associate chairs on academic initiatives.

"We are thrilled to have Mark step into this vital role,” said Arijit Raychowdhury, Steve W. Chaddick School Chair and professor. “His extensive experience and forward-thinking approach will be invaluable as we continue to attract top Ph.D. talent. I am confident that his innovative ideas and collaborative approach will significantly enhance our graduate programs."

Georgia Tech's ECE graduate program is among the largest in the country, with over 1,300 students. It holds a top-five national ranking, including No. 2 among public universities, per U.S. News & World Report.

Since 2018, Davenport has served as the graduate recruitment coordinator for ECE, where he played a strategic role in working with the ECE Graduate Affairs Office and the School’s Graduate Student Recruitment Committee to recruit top Ph.D. candidates and improve the matriculation rate of exceptional applicants.

Joining the ECE faculty in 2012, Davenport’s previous roles include serving as an NSF Mathematical Sciences Postdoctoral Research Fellow at Stanford University and a visitor with the Laboratoire Jacques-Louis Lions at the Université Pierre et Marie Curie. He holds a B.S.E.E., M.S., and Ph.D. from Rice University.

His research focuses on the role of low-dimensional models and optimization in signal processing, statistical inference, and machine learning. He has received several prestigious awards, including the NSF CAREER Award, the Air Force Office of Scientific Research Young Investigator Award, the Sloan Research Fellowship, and the Presidential Early Career Award for Scientists and Engineers (PECASE). Additionally, he currently serves as a Senior Editor for IEEE Transactions on Signal Processing and is a two-time winner of the fastest faculty member award in the Pi Mile Road Race.

Davenport succeeds Mathieu Bloch, who was named Associate Dean of Academic Affairs in the College of Engineering in January.

What models will work best for military AI-human teams? That’s the question Nunn School of International Affairs Associate Professor Margaret E. Kosal will work to help answer as part of a Georgia Tech Research Institute-led project examining the use of human-AI teams. 

Jain is the creator of VCoder, a new tool that enhances the visual perception capabilities of multimodal large language models (MLLMs). Jain said MLLMs like GPT-4 with vision (GPT-4V) are prone to miss obscure objects that blend in with other objects in an image.

Jain paused his TV as he watched The One with the Halloween Party episode of the popular TV Series Friends. 

Chandler stood out the most in a pink bunny costume while holding hands with Ross in a potato costume. As the two prepared for an arm-wrestling match with Joey and Phoebe, multiple groups socialized behind them.

Jain wondered how accurate GPT-4V would be if he prompted itto describe everything happening in this image.

“I watch a lot of TV series, so I frequently think about ways to leverage or include some aspects of those into my work,” said Jain, a Ph.D. student in the School of Interactive Computing. “The scene was very cluttered, so I thought, what questions could I ask GPT-4 about this show.”

On the surface, the generative AI chatbot knew much about the image. It knew which show and episode it was from and recognized the man in the bunny costume as Chandler. It knew the main characters were hosting a Halloween party.

But when Jain prompted the chatbot to count the number of people in the image, he discovered that GPT-4V and its open-source counterparts fell short of performing the simplest task.

It answered 10 when the correct answer was 14. In the right corner of the image, there is a group of people standing in front of a dark curtain that GPT-4V had missed. 

AI Paradox

Jain had a theory — the MLLMs had not been trained for the object perception task and did not have the necessary information to perceive the objects in the foreground and background.

“We started testing it with different pictures, and GPT-4V kept underperforming,” Jain said. “The key takeaway is that it couldn’t do a simple task such as counting the people in the scene, but it knew complex information such as what was happening and who the characters were. This phenomenon is Moravec’s Paradox in Perception — the MLLMs visually reason about complex questions but fail at simple object perception tasks like counting.”

Jain said he has worked on image segmentation tools for the past two years. That includes when he was a research intern at Picsart AI under his now Ph.D. advisor Humphrey Shi, an associate professor in the School of Interactive Computing.

The core idea behind VCoder is to act as a perceptive eye for the MLLM, using segmentation and depth maps obtained through established computer vision frameworks with minimal training costs. The tool also proposes evaluation metrics for object perception tasks like counting and ordering.

Its training and evaluation set consists of images from Common Objects in Context (COCO), a widely used object perception dataset. Associate Professor James Hays from the School of Interactive Computing was one of the academic collaborators who worked with Microsoft to create COCO.

Improving MLLMs

Though VCoder didn’t know which show the image was from, it accurately described everything, including the number of people. It showed as much as 10% more accuracy than its nearest competitor.

It could also identify the order of objects in a scene.

Jain designed VCoder to integrate easily with existing MLLMs. He said augmenting MLLMs with VCoder leads to an MLLM with sound general reasoning and object perception capabilities.

However, he added he was unsure if integration would happen because companies like Open AI, which created GPT-4V, may overlook it.

“There’s no way to know if they will integrate since GPT-4V is a closed model, and their main motivation is to make a product useful to consumers in general,” he said. “They often ignore these small issues.”

Jain’s paper was accepted into the Institute of Electrical and Electronics Engineers’ 2024 Conference on Computer Vision and Pattern Recognition (CVPR), occurring June 17-21 in Seattle. CVPR is the highest-ranked conference in computer vision according to Google Scholar.

Neural networks do the opposite, making the same decisions each time. Now, Georgia Tech researchers in Associate Professor Dobromir Rahnev’s lab are training them to make decisions more like humans. This science of human decision-making is only just being applied to machine learning, but developing a neural network even closer to the actual human brain may make it more reliable, according to the researchers.

In a paper in Nature Human Behaviour, “The Neural Network RTNet Exhibits the Signatures of Human Perceptual Decision-Making,” a team from the School of Psychology reveals a new neural network trained to make decisions similar to humans.

Decoding Decision

“Neural networks make a decision without telling you whether or not they are confident about their decision,” said Farshad Rafiei, who earned his Ph.D. in psychology at Georgia Tech. “This is one of the essential differences from how people make decisions.” 

Large language models (LLM), for example, are prone to hallucinations. When an LLM is asked a question it doesn’t know the answer to, it will make up something without acknowledging the artifice. By contrast, most humans in the same situation will admit they don’t know the answer. Building a more human-like neural network can prevent this duplicity and lead to more accurate answers.

Making the Model

The team trained their neural network on handwritten digits from a famous computer science dataset called MNIST and asked it to decipher each number. To determine the model’s accuracy, they ran it with the original dataset and then added noise to the digits to make it harder for humans to discern. To compare the model performance against humans, they trained their model (as well as three other models: CNet, BLNet, and MSDNet) on the original MNIST dataset without noise, but tested them on the noisy version used in the experiments and compared results from the two datasets. 

The researchers’ model relied on two key components: a Bayesian neural network (BNN), which uses probability to make decisions, and an evidence accumulation process that keeps track of the evidence for each choice. The BNN produces responses that are slightly different each time. As it gathers more evidence, the accumulation process can sometimes favor one choice and sometimes another. Once there is enough evidence to decide, the RTNet stops the accumulation process and makes a decision. 

The researchers also timed the model’s decision-making speed to see whether it follows a psychological phenomenon called the “speed-accuracy trade-off” that dictates that humans are less accurate when they must make decisions quickly. 

Once they had the model’s results, they compared them to humans’ results. Sixty Georgia Tech students viewed the same dataset and shared their confidence in their decisions, and the researchers found the accuracy rate, response time, and confidence patterns were similar between the humans and the neural network.

“Generally speaking, we don't have enough human data in existing computer science literature, so we don't know how people will behave when they are exposed to these images. This limitation hinders the development of models that accurately replicate human decision-making,” Rafiei said. “This work provides one of the biggest datasets of humans responding to MNIST.” 

Not only did the team’s model outperform all rival deterministic models, but it also was more accurate in higher-speed scenarios due to another fundamental element of human psychology: RTNet behaves like humans. As an example, people feel more confident when they make correct decisions. Without even having to train the model specifically to favor confidence, the model automatically applied it, Rafiei noted. 

“If we try to make our models closer to the human brain, it will show in the behavior itself without fine-tuning,” he said.

The research team hopes to train the neural network on more varied datasets to test its potential. They also expect to apply this BNN model to other neural networks to enable them to rationalize more like humans. Eventually, algorithms won’t just be able to emulate our decision-making abilities, but could even help offload some of the cognitive burden of those 35,000 decisions we make daily.

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.

“Congratulations to all the exceptional nominees and to the individuals who were selected to receive this year's Institute Research Awards,” said Chaouki Abdallah, executive vice president for Research at Georgia Tech. “These outstanding researchers were nominated by their peers for their diligent research efforts, and we are proud to acknowledge them for their commitment to advance science and technology and to improve the human condition.”

Awardees were selected in nine areas, from achievements in innovation to engagement and outreach. Two of the awards were given to groups of researchers who are making an impact collectively. This year, more than 150 researchers were nominated for these prestigious awards.

• Outstanding Achievement in Advancing Diversity, Equity, and Inclusion:Iris Tien, CEE, SEI
• Outstanding Achievement in Early Career Research:Marta Hatzell, ME, IMat, SEI  
• Outstanding Achievement in Research Enterprise Enhancement:Robert Knotts, Federal Relations 
• Outstanding Achievement in Research Innovation:Younan Xia, SoCB, IEN, Imat, IBB
• Outstanding Doctoral Thesis Advisor:John Reynolds, SoCB, IMat, RBI
• Outstanding Faculty Research Author:Wilbur Lam, BME, IEN, IBB 
• Outstanding Achievement in Research Engagement and Outreach:Pascal Van Hentenryck, ISyE, IDEaS, SEI
• Outstanding Achievement in Research Program Development: The Spaceflight Project Group at GT: Glen Lightsey, AE, IRIM; Jud Ready, GTRI, IEN, IMat, SEI; Christopher Valenta, GTRI; Christopher Carr, AE; Brian Gunter, AE, BBISS, IRIM; Sterling Peet, AE; Ian Harrison, GTRI
• Outstanding Achievement in Research Program Impact:Partnership for Inclusive Innovation: Debra Lam, IDEaS, IPaT, SEI; Clarence Anthony Jr., Kayla Burns, Cody Cocchi, Jamal Lewis, Polly Sattler, all from EI2

Awardees will be recognized at the Faculty and Staff Honors Luncheon on Friday, April 21. 

Call for Proposals: A Gateway to Innovation

Launched in early 2024, the Call for Proposals invited researchers from across Georgia Tech to submit their innovative ideas on GenAI. The scope was broad, encouraging proposals that spanned foundational research, system advancements, and novel applications in various disciplines, including arts, sciences, business, and engineering. A special emphasis was placed on projects that addressed responsible and ethical AI use.

The response from the Georgia Tech research community was overwhelming, with 76 proposals submitted by teams eager to explore this transformative technology. After a rigorous selection process, eight projects were selected for support. Each awarded team will also benefit from access to Microsoft’s Azure cloud resources..

Recognizing Microsoft’s Generous Contribution

This successful initiative was made possible through the generous support of Microsoft, whose contribution of research resources has empowered Georgia Tech researchers to explore new frontiers in GenAI. By providing access to Azure’s advanced tools and services, Microsoft has played a pivotal role in accelerating GenAI research at Georgia Tech, enabling researchers to tackle some of the most pressing challenges and opportunities in this rapidly evolving field.

Looking Ahead: Pioneering the Future of GenAI

The awarded projects, set to commence in Fall 2024, represent a diverse array of research directions, from improving the capabilities of large language models to innovative applications in data management and interdisciplinary collaborations. These projects are expected to make significant contributions to the body of knowledge in GenAI and are poised to have a lasting impact on the industry and beyond.

IDEaS and the Cloud Hub are committed to supporting these teams as they embark on their research journeys. The outcomes of these projects will be shared through publications and highlighted on the Cloud Hub web portal, ensuring visibility for the groundbreaking work enabled by this initiative.

Congratulations to the Fall 2024 Winners

• Annalisa Bracco | EAS "Modeling the Dispersal and Connectivity of Marine Larvae with GenAI Agents" [proposal co-funded with support from the Brook Byers Institute for Sustainable Systems]
• Yunan Luo | CSE “Designing New and Diverse Proteins with Generative AI”
• Kartik Goyal | IC “Generative AI for Greco-Roman Architectural Reconstruction: From Partial Unstructured Archaeological Descriptions to Structured Architectural Plans”
• Victor Fung | CSE “Intelligent LLM Agents for Materials Design and Automated Experimentation”
• Noura Howell | LMC “Applying Generative AI for STEM Education: Supporting AI literacy and community engagement with marginalized youth”
• Neha Kumar | IC “Towards Responsible Integration of Generative AI in Creative Game Development”
• Maureen Linden | Design “Best Practices in Generative AI Used in the Creation of Accessible Alternative Formats for People with Disabilities”
• Surya Kalidindi | ME & MSE “Accelerating Materials Development Through Generative AI Based Dimensionality Expansion Techniques”
• Tuo Zhao | ISyE “Adaptive and Robust Alignment of LLMs with Complex Rewards”

The award-winning paper, “HLSFactory: A Framework Empowering High-Level Synthesis Datasets for Machine Learning and Beyond,” introduces an open-source infrastructure that simplifies the contribution and sharing of hardware designs.

This platform allows developers to submit their designs to a common pool, making it easy for machine learning practitioners to access the data they need to train their algorithms, ultimately driving advancements in hardware design.

The framework addresses a major challenge in applying machine learning to chip design:the scarcity of high-quality training data. By being open-source, HLSFactory encourages a wide range of contributions from the community, creating a richer dataset for researchers and developers.

“HLSFactory creates unified platform that the entire community can leverage to drive more scientific discoveries and improve design outcomes,” said Hao.

The software was developed through Hao's Software/Hardware Co-Design Lab.

ECE Ph.D. candidate Stefan Abi-Karam was the lead author on the paper. Abi-Karam previously won the Community Award at the 33rd International Conference on Field-Programmable Logic and Applications (FPL) 2023, for his work on “GNNBuilder: An Automated Framework for Generic Graph Neural Network Accelerator Generation, Simulation, and Optimization.” The award recognized his contributions to the open-source community, underscoringthe group’s dedication to open-source tools and hardware design, according to Hao.

ECE Ph.D. candidates Rishov Sarkar and Hanqiu Chen also contributed to the paper, alongside collaborators Allison Seigler, Sean Lowe, Zhigang Wei, Nanditha Rao, and Lizy Kurian John from the University of Texas at Austin, and Aman Arora from Arizona State University.

Hao, who joined the Georgia Tech ECE faculty in 2022, specializes in efficient hardware design and machine learning algorithms. Her research focuses on reconfigurable, high-efficiency computing and developing practical electronic design automation tools. She has previously been recognized with several prestigious awards, including the NSF CAREER Award and the Intel Rising Star Faculty Award, for her groundbreaking work in hardware design research.
 

Artificial Intelligence at Georgia Tech (AI@GT) is a student-led organization that offers a platform for seasoned and novice students to explore AI. 

Vamvoudakis received $400,000 from the National Science Foundation for his proposal, “Improving Safety by Synthesizing Interacting Model-based and Model-free Learning Approaches.” This is the first grant on Safe Learning-enabled Systems (SLES) awarded to Georgia Tech from NSF. He and his team will establish a framework that leads to the design and implementation of SLES in which safety is ensured with high confidence levels. The framework will leverage tools from control theory, multi-agent autonomy, and formal methods for rigorously probabilistic reasoning to create safe learning-enabled systems. Before anyone releases an autonomous machine, the public expects it to be safe for those around it. For example, sensors in drones and other machines are sensitive to infiltration, malfunction, and the environment. If the wind is strong, the drone would need to be able to adjust to the environment, stay on course, and perhaps change altitude. If the drone encounters a telephone pole or even a person in its path, it would be able to adjust accordingly without waiting for human control. 

His research approach will take elements from various theories and combine them to improve the safety of these LES within the machine.

“Our approach algorithmically combines model-based and model-free reinforcement learning for enhancing safety by using the learned model to predict how well a safe policy will behave and then update the resulting actions,” Vamvoudakis explained. “As a result, our approach does not rely on improving the model and does not require an infinite amount of time for convergence. Instead, our plan optimally enhances safety and combines the predefined time-convergent actions generated to achieve high performance in the specified task.”

The fundamental knowledge created in his research could inform how future-assured autonomous systems with embodied intelligence can be built. Their results could inform the design of key enablers of the global economy, including smart and connected cities, networked actions of smart and autonomous systems by enabling system flexibility, efficiency, and capacity, and automated financial trading, such as creating automated news digests around finance.

Gaming Strategies Inform Military LES Frameworks

Autonomous machines are changing the way that the military operates. Uncrewed battles between autonomous systems require the systems to learn and adapt to unknown environments and to distinguish allies from enemies. Learning-enabled systems are trained to take the circumstances at hand and give recommendations for the desired response.

When humans have control over these machines, this is considered humans in the loop. When humans move further into the background and give the machines decision-making autonomy, it is called humans on the loop. Humans would still have oversight, but the machine could ultimately decide without human approval.

Through his newly awarded $480,000 project “Embodied and Secure Physical Intelligence with Possible Humans-on-the Loop in Complex Adaptive Systems” with the Army Research Office (ARO), Vamvoudakis and his team are developing decision-making algorithms to assist during conflict in adversarial environments. This is needed because military maneuvers can be unpredictable, and autonomous machines need to be able to adapt accordingly. He will use game-theoretic strategies to inform his work.

Vamvoudakis’ team has created algorithms in the context of games, where a “defender” wants to regulate a cyber-physical system around a trimming point, but an “attacker” intends to disrupt this regulation as much as possible.

They also employed level-k thinking to capture the behavior of the attacker. Particularly, instead of assuming that the attacker can reason perfectly about the behavior of the defender, the employed level-k thinking model imposed that the attacker can only make finitely-many (though arbitrarily many) steps of reasoning about what the defender might do, how the attacker can best respond to that, how the defender can then best respond. 

The project is a continuation of his ARO YIP award that developed a way to understand different types of attackers in a unified framework. Attackers who think a little ahead are called low-level, while those who think more strategically, like those near a Nash equilibrium, are called high-level. This understanding helps create better defense strategies without assuming that attackers always act perfectly. 

To demonstrate how this model works in real military situations, he and his students looked at it through the lens of a pursuit-evasion game. They found that using level-k thinking to understand and respond to attackers is more effective than assuming attackers always optimize their strategies perfectly.

MathWorks Gift to Enhance Learning for Artificial Intelligence

Current methods for protecting closed-loop reinforcement learning systems (artificial intelligence where the system continuously learns and adapts based on feedback from the environment) don't work well against potential threats. These existing methods often rely on guesswork, need a deep understanding of the system, and require a lot of training time. They also fail to guarantee safety when facing adversaries. 

Vamvoudakis’ MathWorks gift, “Adversarial Reinforcement Learning” aims to create a new generation of smart, flexible, autonomous systems that can learn and adapt. This is the first-ever gift from MathWorks made to Georgia Tech. 

“We will develop the next generation of agile, highly adaptive autonomous systems that use mechanisms from cognition and learning to process information from distributed sensors. In particular, looking to autonomous systems appearing in nature for inspiration,” he said. Specifically, behavioral scientists have validated the need for intermittent data sharing in learning tasks. They have shown that the central nervous system in human beings minimizes effort and sorts through impulses and stimuli by maintaining intermittent signaling. Specifically, the spinal cord transmits a channel of information and effectively exploits its neural resources via intermittent strategies to produce a sequence of muscle-bone interactions that induce movement.”

By looking to such ideas, they will develop safe and strong reinforcement learning methods to handle teamwork, assign tasks, and manage resources effectively. They will also collaborate with MathWorks to create useful toolboxes and provide internship opportunities.

Tao is an applied and computational mathematician who designs and synergizes mathematical tools to solve practical problems. Recently, he has focused on the applications of these tools to machine learning. Tao works on multiple subareas of machine learning, including deep learning theory, probabilistic methods, generative modeling, and artificial intelligence for science (“AI4Science”). 

"Molei is doing breakthrough work on machine learning and artificial intelligence,” says Mike Wolf, chair of the School of Mathematics. “It is wonderful to see him recognized by Sony, both for his accomplishments so far and also his promise for the future. His unique perspectives, informed by an astonishing deep breadth of understanding of mathematics, have already made him one of the more prominent researchers in this extremely competitive and important field. I know that this award will fuel even more impactful works. We are just thrilled to have Molei on our faculty in the School of Mathematics."

Revolutionizing Generative AI

The award recognizes Tao’s research on the mathematical and algorithmic aspects of diffusion generative modeling, which is considered one of the foundations of modern Generative AI. Using advanced machine learning algorithms, these models have revolutionized the generation of image, video, and 3D content. 

“Exciting products such as ChatGPT, Stable Diffusion, and Sora are generative AI tools, and a good number of them are powered by diffusion models,” explains Tao. “The way the magic works is you basically give a machine learning model a collection of training data, and then the algorithm can generate more content that is similar to the training data. The ability of generating new content is called generative modeling. Diffusion model is one of the latest technologies for generative modeling.”

Tao’s work aims to make diffusion models more versatile and scalable. He hopes to broaden their application and possibly create the next generation of generative modeling tools. 

“The large-scale impact of this research is to make generative AI more accessible, more creative, safer, and more trustworthy,” he adds. 

To learn more about Tao's research, visit his blog or follow him on Twitter at @MoleiTaoMath.

Georgia Tech researchers are making this process a little easier, though. Their new tool, AF2Complex, used deep learning to predict which antibodies could bind to Covid-19’s infamous spike protein. The researchers created input data for the deep-learning model using sequences of known antigen binders. This method correctly predicted 90% of the best antibodies in one test with 1,000 antibodies and was recently published in Proceedings of the National Academy of Sciences. Treating Covid-19 is just the start of its potential. 

“AF2Complex improves therapeutic development,” said Mu Gao, a senior research scientist in the School of Biological Sciences (SBS). “If you have a high-quality model, then you can tinker with the protein sequence and optimize the antibody, making it more suitable for drug development.”

Developing the Deep-Learning Model

The researchers weren’t the first to use deep learning to predict protein structures, but they did considerably expand the model’s capabilities. In 2020, the Nobel Prize-winning DeepMind AlphaFold, an Alphabet project, made breakthroughs using deep learning to predict the protein structures of single proteins. Georgia Tech researchers pushed the model to predict the structures of protein complexes. In 2021, they created the first version of AF2Complex, which could predict interactions between multiple, complex proteins like E. coli. Applying it to human proteins was the next step — but much harder.

“Normally, when you predict protein-protein interactions, the surface area of the protein is quite large, so you could afford to make a few mistakes with an imperfect model,” said Jeffrey Skolnick, a Regents’ Professor and the Mary and Maisie Gibson Chair in SBS and a Georgia Research Alliance Eminent Scholar.  “But an antibody-protein interaction occupies a much smaller interfacial area. Imagine going from hitting a big target anywhere to hitting the bullseye.”

Determining how to predict the antibody-antigen interactions was the biggest challenge. The researchers focused on the Covid-19 virus because it had many complex antigen-binding sequences and epitopes, the specific molecule region that interacts with B- and T-cells to trigger an immune response. Covid-19 also was a widely available dataset, unlike many immunotherapies to which only pharmaceutical companies have access. The Covid-19 database, in effect, offered a rich training environment for the AF2 algorithm. 

Skolnick and Gao used Covid-19 sequences from known antibodies to identify evolutionary relationships and patterns, improving the accuracy of predictions. From there, they applied the AF2 deep-learning model, already trained on a vast amount of protein structure data. The model used sequences to predict how proteins fold and interact, developing a 3D structure of protein complexes. Plus, it could produce 3D structures for more than just one dominant epitope.

The predictions were validated against experimental data, refining the model. With these predicted structures, researchers can do everything from better understanding biological processes to developing new drugs.

Treating the Virus of the Future

The researchers believe deep-learning technologies could revolutionize how we treat future diseases. With infinite resources and time, researchers could manually try every antibody-antigen combination, but no scientist has that. AF2Complex can narrow the focus and get to the treatment sooner. 

“Imagine the virus from hell arises. You could design a series of antibodies using this algorithm, so it cuts down the time for vaccine development,” Skolnick said. “There are no substitutes for a real experiment, but AF2Complex can prioritize which experiments you should do, so you have more shots at the goal.”

The researchers are already collaborating with Emory University to conduct experiments that validate AF2Complex’s findings. They also are pursuing a path to commercialize the model. When the next pandemic starts, we will be better prepared.

The  National Institutes of Health provided, and the Department of Energy and National Science Foundation supported, the main computing resources.

Stopping such cyberattacks requires understanding them first, but this can be challenging. Finding a computing system’s exact deep neural network has many roadblocks. They are often proprietary and, therefore, inaccessible to investigators without considerable legal intervention. Another common problem is that they are updated frequently, making it difficult for investigating researchers to access the most current network iteration. But a new tool from Georgia Tech could unlock the mysterious malware on myriad neural networks in everything from self-driving cars to the IMDB entertainment database. AI Psychiatry (AiP) is a postmortem cybersecurity forensic tool that uses artificial intelligence to recover the exact models a compromised machine runs on and discover where the fatal error occurred.

“We trust self-driving cars with our lives and ChatGPT with our careers, but when those systems fail, how are we going to investigate them?” said Brendan Saltaformaggio, an associate professor with joint appointments in the School of  Cybersecurity and Privacy and the School of Electrical and Computer Engineering (ECE).

AiP can recover the original DL model on both the local network’s memory and the graphics processing unit that trains the network. It can accomplish this without any specific knowledge of the model’s framework, platform, or version. Instead, it recreates the model using what Saltaformaggio refers to as “clues,” or common components in all neural networks. These include weights, biases, shapes, and layers from the model’s memory image — a frozen set of the bits and bytes operating when the model is running normally. The memory image is crucial because it enables AiP to compare it with the model post-attack.

“These models often refine their information as they go, based on their current environment, so an attack might happen as a result of an attacker poisoning the information a particular model is learning,” said David Oygenblik, an ECE Ph.D. student. “We determined that a memory image would capture all those changes that occur during a runtime.”\

Once the model is recovered, AiP can run it on another device, letting investigators test it thoroughly to determine where the flaws lie. AiP has been tested with different versions of both popular machine learning frameworks (TensorFlow and PyTorch) and datasets (CIFAR-10, LISA, and IMDB). It successfully recovered and rehosted 30 models with 100% accuracy. 

“Before our research, you couldn't go to the cyber ‘crime scene’ and find clues because there was no technique available to do that,” Saltaformaggio said. “That's what we are pioneering in the cyber forensics lab right now — techniques to get that evidence out of a crime scene.”

Tools like AiP will allow cyber investigators to see the whole picture immediately. Solving cybercrimes can help prevent future ones, from safeguarding a user’s data to keeping a car on the road. 

AiP is the inaugural winner of GTRI's Graduate Student Fellowship Program. 

Two years since OpenAI introduced ChatGPT, dozens of lawsuits have been filed alleging technology companies have infringed copyright by using published works to train artificial intelligence (AI) models.

Academic AI research efforts could be significantly hindered if courts rule in the plaintiffs' favor. 

Desai and Riedl are Georgia Tech researchers raising awareness about how these court rulings could force academic researchers to construct new AI models with limited training data. The two collaborated on a benchmark academic paper that examines the landscape of the ethical issues surrounding AI and copyright in industry and academic spaces.

“There are scenarios where courts may overreact to having a book corpus on your computer, and you didn’t pay for it,” Riedl said. “If you trained a model for an academic paper, as my students often do, that’s not a problem right now. The courts could deem training is not fair use. That would have huge implications for academia.

“We want academics to be free to do their research without fear of repercussions in the marketplace because they’re not competing in the marketplace,” Riedl said. 

Desai is the Sue and John Stanton Professor of Business Law and Ethics at the Scheller College of Business. He researches how business interests and new technology shape privacy, intellectual property, and competition law. Riedl is a professor at the College of Computing’s School of Interactive Computing, researching human-centered AI, generative AI, explainable AI, and gaming AI. 

Their paper, Between Copyright and Computer Science: The Law and Ethics of Generative AI, was published in the Northwestern Journal of Technology and Intellectual Property on Monday.

Desai and Riedl say they want to offer solutions that balance the interests of various stakeholders. But that requires compromise from all sides.

Researchers should accept they may have to pay for the data they use to train AI models. Content creators, on the other hand, should receive compensation, but they may need to accept less money to ensure data remains affordable for academic researchers to acquire.

Who Benefits?

The doctrine of fair use is at the center of every copyright debate. According to the U.S. Copyright Office, fair use permits the unlicensed use of copyright-protected works in certain circumstances, such as distributing information for the public good, including teaching and research.

Fair use is often challenged when one or more parties profit from published works without compensating the authors.

Any original published content, including a personal website on the internet, is protected by copyright. However, copyrighted material is republished on websites or posted on social media innumerable times every day without the consent of the original authors. 

In most cases, it’s unlikely copyright violators gained financially from their infringement.

But Desai said business-to-business cases are different. The New York Times is one of many daily newspapers and media companies that have sued OpenAI for using its content as training data. Microsoft is also a defendant in The New York Times’ suit because it invested billions of dollars into OpenAI’s development of AI tools like ChatGPT.

“You can take a copyrighted photo and put it in your Twitter post or whatever you want,” Desai said. “That’s probably annoying to the owner. Economically, they probably wanted to be paid. But that’s not business to business. What’s happening with Open AI and The New York Times is business to business. That’s big money.”

OpenAI started as a nonprofit dedicated to the safe development of artificial general intelligence (AGI) — AI that, in theory, can rival human thinking and possess autonomy.

These AI models would require massive amounts of data and expensive supercomputers to process that data. OpenAI could not raise enough money to afford such resources, so it created a for-profit arm controlled by its parent nonprofit.

Desai, Riedl, and many others argue that OpenAI ceased its research mission for the public good and began developing consumer products. 

“If you’re doing basic research that you’re not releasing to the world, it doesn’t matter if every so often it plagiarizes The New York Times,” Riedl said. “No one is economically benefitting from that. When they became a for-profit and produced a product, now they were making money from plagiarized text.”

OpenAI’s for-profit arm is valued at $80 billion, but content creators have not received a dime since the company has scraped massive amounts of copyrighted material as training data.

The New York Times has posted warnings on its sites that its content cannot be used to train AI models. Many other websites offer a robot.txt file that contains instructions for bots about which pages can and cannot be accessed. 

Neither of these measures are legally binding and are often ignored.

Solutions

Desai and Riedl offer a few options for companies to show good faith in rectifying the situation.

• Spend the money. Desai says Open AI and Microsoft could have afforded its training data and avoided the hassle of legal consequences.
  
  “If you do the math on the costs to buy the books and copy them, they could have paid for them,” he said. “It would’ve been a multi-million dollar investment, but they’re a multi-billion dollar company.”
   
• Be selective. Models can be trained on randomly selected texts from published works, allowing the model to understand the writing style without plagiarizing. 
  
  “I don’t need the entire text of War and Peace,” Desai said. “To capture the way authors express themselves, I might only need a hundred pages. I’ve also reduced the chance that my model will cough up entire texts.”
   
• Leverage libraries. The authors agree libraries could serve as an ideal middle ground as a place to store published works and compensate authors for access to those works, though the amount may be less than desired.
  
  “Most of the objections you could raise are taken care of,” Desai said. “They are legitimate access copies that are secure. You get access to only as much as you need. Libraries at universities have already become schools of information.”

Desai and Riedl hope the legal action taken by publications like The New York Times will send a message to companies that develop AI tools to pump the breaks. If they don’t, researchers uninterested in profit could pay the steepest price.

The authors say it’s not a new problem but is reaching a boiling point.

“In the history of copyright, there are ways that society has dealt with the problem of compensating creators and technology that copies or reduces your ability to extract money from your creation,” Desai said. “We wanted to point out there’s a way to get there.”

 “What is happening within the field is revolutionary,” says School of Earth and Atmospheric Sciences Associate Chair and Professor Annalisa Bracco, adding that because many climate-related processes — from ocean currents to melting glaciers and weather patterns — can be described with physical equations, these advancements have the potential to help us understand and predict climate in critically important ways. 

Bracco is the lead author of a new review paper providing a comprehensive look at the intersection of AI and climate physics.

The result of an international collaboration between Georgia Tech’s Bracco, Julien Brajard (Nansen Environmental and Remote Sensing Center), Henk A. Dijkstra (Utrecht University), Pedram Hassanzadeh (University of Chicago), Christian Lessig (European Centre for Medium-Range Weather Forecasts), and Claire Monteleoni (University of Colorado Boulder), the paper, ‘Machine learning for the physics of climate,’ was recently published in Nature Reviews Physics. 

“One of our team’s goals was to help people think deeply on how climate science and AI intersect,” Bracco shares. “Machine learning is allowing us to study the physics of climate in a way that was previously impossible. Coupled with increasing amounts of data and observations, we can now investigate climate at scales and resolutions we’ve never been able to before.”

Connecting hidden dots

The team showed that ML is driving change in three key areas: accounting for missing observational data, creating more robust climate models, and enhancing predictions, especially in weather forecasting. However, the research also underscores the limits of AI — and how researchers can work to fill those gaps.

“Machine learning has been fantastic in allowing us to expand the time and the spatial scales for which we have measurements,” says Bracco, explaining that ML could help fill in missing data points — creating a more robust record for researchers to reference. However, like patching a hole in a shirt, this works best when the rest of the material is intact.

“Machine learning can extrapolate from past conditions when observations are abundant, but it can’t yet predict future trends or collect the data we need,” Bracco adds. “To keep advancing, we need scientists who can determine what data we need, collect that data, and solve problems.”

Modeling climate, predicting weather

Machine learning is often used when improving climate models that can simulate changing systems like our atmosphere, oceans, land, biochemistry, and ice. “These models are limited because of our computing power, and are run on a three-dimensional grid,” Bracco explains: below the grid resolution, researchers need to approximate complex physics with simpler equations that computers can solve quickly, a process called ‘parameterization’.

Machine learning is changing that, offering new ways to improve parameterizations, she says. “We can run a model at extremely high resolutions for a short time, so that we don’t need to parameterize as many physical processes — using machine learning to derive the equations that best approximate what is happening at small scales,” she explains. “Then we can use those equations in a coarser model that we can run for hundreds of years.”

While a full climate model based solely on machine learning may remain out of reach, the team found that ML is advancing our ability to accurately predict weather systems and some climate phenomena like El Niño. 

Previously, weather prediction was based on knowing the starting conditions — like temperature, humidity, and barometric pressure — and running a model based on physics equations to predict what might happen next. Now, machine learning is giving researchers the opportunity to learn from the past. “We can use information on what has happened when there were similar starting conditions in previous situations to predict the future without solving the underlying governing equations,” Bracco says. “And all while using orders-of-magnitude less computing resources.”

The human connection

Bracco emphasizes that while AI and ML play a critical role in accelerating research, humans are at the core of progress. “I think the in-person collaboration that led to this paper is, in itself, a testament to the importance of human interaction,” she says, recalling that the research was the result of a workshop organized at the Kavli Institute for Theoretical Physics — one of the team’s first in-person discussions after the Covid-19 pandemic.

“Machine learning is a fantastic tool — but it's not the solution to everything,” she adds. “There is also a real need for human researchers collecting high-quality data, and for interdisciplinary collaboration across fields. I see this as a big challenge, but a great opportunity for computer scientists and physicists, mathematicians, biologists, and chemists to work together.”

Funding: National Science Foundation, European Research Council, Office of Naval Research, US Department of Energy, European Space Agency, Choose France Chair in AI.

DOI: https://doi.org/10.1038/s42254-024-00776-3

Q: What do you believe will be the most significant breakthrough in AI in 2025? 

A: Predicting breakthroughs is inherently difficult, as they often emerge unexpectedly. Most major advancements stem from years of incremental progress and research that suddenly converge into a solution that surprises us with its effectiveness. Usually, these breakthroughs originate in areas that are not yet widely recognized as promising or trending.

This unpredictability highlights the importance of sustained investment in diverse research directions, including high-risk projects and fundamental science. By prioritizing innovation, companies and the U.S. government can help ensure the groundwork is laid for transformative discoveries in AI.

Q: Which industries will experience the greatest transformation due to AI in the next year?  

A: In the coming year, advancements in AI are poised to transform industries across the board, both within and beyond the tech sector. Sectors such as healthcare, retail (custom service), marketing, law, education, entertainment, and many others are increasingly adopting large language models (LLMs) to enhance data analysis and improve user interactions. These models offer capabilities far surpassing traditional machine learning methods, driving a shift toward more efficient and intelligent systems.

Ethical and Societal Impact 

Q: What are the most pressing ethical challenges for AI development and deployment in 2025? 

A: Safety: One of the most critical concerns is ensuring the accuracy and reliability of AI systems, especially in high-stakes scenarios such as providing medical advice.

Privacy: With AI systems, privacy risks are heightened when users share vast amounts of personal data — such as emails, resumes, and meeting transcripts.

Fairness: Language models must be designed to account for diverse cultural backgrounds, values, opinions, and languages—including dialects and individual linguistic styles. 

Q: How do you see AI influencing education and learning in the coming year?

A: AI will significantly impact education. It can expand access to knowledge through personalized learning tools and make educational materials more widely accessible. Integrating AI into K-12 curriculums will raise public awareness and prepare students for the future. However, educators are concerned about misuse, such as students relying on AI to complete assignments without fully understanding the material.

Q: What role will generative AI play in shaping public discourse and creativity in 2025? 

A: I think Joanna Maciejewska's comment has crystallized it: "I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.

Research and Technology 

Q: What advancements in AI hardware or infrastructure are crucial for 2025?

A: Advancements in AI hardware and infrastructure crucial for 2025 include innovations in GPU and specialized hardware such as TPUs (tensor processing units) or custom AI chips, which can accelerate model training and inference. Equally critical is improving the energy efficiency of these components, as current AI models heavily rely on massive GPU computing and data center infrastructure. 

Q: Which areas of AI research would you say are under-funded or under-explored?

A: Many! While the field sees a high volume of publications, much of the work tends to be surface-level or rushed, with fewer resources dedicated to deep, innovative research. This can lead to the illusion of progress in some areas without truly solving the underlying problems. For instance, despite significant advancements, challenges remain in areas like machine translation. While current systems perform impressively, they often fall short in capturing nuances, cultural context, and idiomatic expressions. Similarly, tools like ChatGPT can assist with writing. However, they still struggle to match a skilled human writer's sophistication, coherence, and creativity. Furthermore, using AI in writing can sometimes result in the loss of personal linguistic style, making outputs less authentic or unique.

Q: How do you predict AI regulations or policies will evolve over the next year? 

A: Over the next year, I anticipate and hope that AI regulations and policies will become more nuanced as policymakers and the public gain a deeper understanding of AI technologies. Ideally, we will see more policymakers, such as U.S. congressional members, with science or engineering backgrounds who can approach AI advancements with a balanced perspective. Such leaders would be better equipped to craft thoughtful regulations addressing privacy, copyright, and bias while avoiding stifling innovation through over-regulation.

Additionally, I hope for a concerted effort to enhance AI literacy among the general public through media, educational initiatives, and public discourse. Greater familiarity with AI technologies can lead to more informed and constructive opinions, helping shape regulations that reflect societal values and priorities. This combination of knowledgeable policymakers and an AI-aware public could foster a regulatory environment that supports innovation while ensuring ethical and

Vision and Advice 

Q: What emerging technologies or approaches in AI excite you the most for 2025?

A: For 2025, I'm most excited about:

• AI-driven personalization: AI adapting to individual language styles, cultural backgrounds, personal knowledge, and real-time context.
• Multilingual and multi-modal models: Enhancing non-English language performance and integrating diverse data types seamlessly.
• Domain-specific foundation models: Specialized models for law, biology, and material science to accelerate breakthroughs.

Additionally, improving large language models in safety, privacy, robustness, efficiency, and better training techniques remains a fundamental focus for better overall performance.

Q: What advice would you offer to students or researchers entering AI in 2025? 

A: My advice for students and researchers entering AI in 2025 is to focus on what truly excites you. Take it step by step—read papers, deepen your knowledge, and stay grounded to avoid feeling overwhelmed. While staying informed about trends is important, aim to be an early adopter of emerging ideas or work on solving long-standing, challenging problems. By pursuing your passion, prioritizing originality, and leveraging your unique strengths, you can forge your path and make a meaningful impact in the field.

Music retrieval apps, for example, employ AI-powered search algorithms. While an update may technically improve the app by decreasing the average error rate of music retrieval, the subtle changes to the retrieval mechanism can cause a subjective performance degradation for the user.

This is known as software regression.

A Georgia Tech School of Electrical and Computer Engineering research team of postdoctoral research Fellow Mohit Prabhushankar, recent Ph.D. graduate Ryan Benkert, and Professor Ghassan AlRegib were the runners-up for the Best Paper Award at the 2024 IEEE International Conference on Big Data in Washington, D.C. Their research focused on minimizing subjective software regression from updates while simultaneously continuing to decrease objective errors.

The paper, “Targeting Negative Flips in Active Learning using Validation Sets,” aimed to train and update AI models using the Negative Flip Rate (NFR) metric.

NFR is a measure used in machine learning to quantify the rate at which a new model makes incorrect predictions on data points that were correctly predicted by a previous model.

The goal for the research is to have low NFR within the models with high algorithm accuracy. They were able to identify certain subsets in the training data that achieved this. The next step is creating a method that can proactively select these subsets.

The paper ultimately proposed to construct a Regression-ordered Subset Estimation (RoSE) using the data to consistently ensure accuracy of the AI model is high while reducing NFR.

Prabhushankar received his Ph.D. in electrical engineering from Georgia Tech in 2021. He is currently a postdoctoral research fellow in the Omni Lab for Intelligent Visual Engineering and Science (OLIVES). He is working in the fields of image processing,
machine learning, active learning, healthcare, and robust and explainable AI.

Benkert received his Ph.D. in electrical engineering from Georgia Tech in 2024. He is currently a deep learning software engineer at NVIDIA in Santa Clara, CA. His research interests are at the intersection of active learning, uncertainty estimation, and performance consistency in neural network learning. Prior to Georgia Tech, he received his B.Sc and M.Sc from the RWTH Aachen University in Germany.

Professor AlRegib is currently the John and Marilu McCarty Chair Professor ECE and directs OLIVES. He and his group work on robust and interpretable machine learning algorithms, uncertainty and trust, and human in the loop algorithms. The group has demonstrated their work on a wide range of applications such as Autonomous Systems, Medical Imaging, and Subsurface Imaging. The group is interested in advancing the fundamentals as well as the deployment of such systems in real-world scenarios. He has been issued several U.S. patents and invention disclosures. He is a Fellow of the IEEE.

College membership honors those, “who have made outstanding contributions to engineering and medicine research, practice, or education,” and “to the pioneering of new and developing fields of technology, making major advancements in traditional fields of medical and biological engineering or developing/implementing innovative approaches to bioengineering education.”

The distinction is among the highest professional distinctions given to medical and biological engineers, comprised of the top two percent of engineers in these fields.

He was nominated and inducted for outstanding contributions to computational neuroengineering, psychiatric neuromodulation, and international leadership in accessible biomedical education.

Rozell's research interests are in computational neuroengineering, an intersection of neuroscience, data science, neurotechnology and computational modeling that advances the understanding of brain function and the design of effective interventions.

His research has a particular focus on advancing our understanding and novel brain stimulation therapies for psychiatric disorders such as treatment resistant depression.

Recently, he was part of a team that identified a unique pattern in brain activity that reflects the recovery process in patients with treatment-resistant depression. This pattern, known as a biomarker, represented a significant advance in treatment for the most severe and untreatable forms of depression.

His work also includes research that takes a creative approach to advancing the understanding of the societal impacts of emerging technologies such as neurotechnology and AI.

Rozell especially takes pride in being a first-generation scholar who is committed to accessibility in scientific communities. In pursuit of this goal, he co-founded and serves on the Board of Directors of Neuromatch, Inc., a global nonprofit increasing access to scientific knowledge.

His scholarly efforts have resulted in many published works in top publications, such as Nature, and a number of awards, including the NSF CAREER Award.

Before joining the ECE faculty in 2008 as an assistant professor, Rozell received a B.S.E. degree in computer engineering and a B.F.A. degree in music in 2000 from the University of Michigan. He then received M.S. and Ph.D. degrees in electrical engineering in 2002 and 2007 from Rice University and was a postdoctoral scholar at the Redwood Center for Theoretical Neuroscience at the University of California, Berkeley.

ATLANTA — March 24, 2025 — Georgia Tech has officially launched Tech AI, a bold new initiative designed to accelerate the real-world impact of artificial intelligence across industry and government. The announcement marks the start of Tech AI Fest, the Southeast’s leading AI event, bringing together leading academics, industry experts, government figures, and students for three days of immersive discussion, creative partnerships, and transformative ideas. 

Now, a Georgia Tech professor and several alumni are helping bring the 1939 classic Hollywood film to what will likely be its largest screen ever: the Las Vegas Sphere's 160,000-square-foot interior screen.

Read more to discover their pivotal role and how generative AI is used to "reconceptualize" the film for the August 28 premiere of The Wizard of Oz at Sphere.

Top South Korean research institutes have enlisted Georgia Tech researchers Sehoon Ha and Jennifer G. Kim to develop artificial intelligence (AI) to help the humanoid assistant navigate hospitals and interact with doctors, nurses, and patients.

Ha and Kim will partner with Neuromeka, a South Korean robotics company, on a five-year, 10 billion won (about $7.2 million US) grant from the South Korean government. Georgia Tech will receive about $1.8 million of the grant.

Ha and Kim, assistant professors in the School of Interactive Computing, will lead Tech’s efforts and also work with researchers from the Korea Advanced Institute of Science and Technology and the Electronics and Telecommunications Research Institute.

Neuromeka has built industrial robots since its founding in 2013 and recently decided to expand into humanoid service robots.

Lee, the group leader of the humanoid medical assistant project, said he fielded partnership requests from many academic researchers. Ha and Kim stood out as an ideal match because of their robotics, AI, and human-computer interaction expertise. 

For Ha, the project is an opportunity to test navigation and control algorithms he’s developed through research that earned him the National Science Foundation CAREER Award. Ha combines computer simulation and real-world training data to make robots more deployable in high-stress, chaotic environments. 

“Dr. Ha has everything we want to put into our system, including his navigation policies,” Lee said. “He works with robots and AI, and there weren’t many candidates in that space. We needed a collaborator who can create the software and has experience running it on robots.”

Ha said he is already considering how his algorithms could scale beyond hospitals and become a universal means of robot navigation in unstructured real-world environments.

“For now, we’re focusing on a customized navigation model for Korean environments, but there are ways to transfer the data set to different environments, such as the U.S. or European healthcare systems,” Ha said. 

“The final product can be deployed to other systems and industries. It can help industrial workers at factories, retail stores, any place where workers can get overwhelmed by a high volume of tasks.”

Kim will focus on making the robot’s design and interaction features more human. She’ll develop a large-language model (LLM) AI system to communicate with patients, nurses, and doctors. She’ll also develop an app that will allow users to input their commands and queries. 

“This project is not just about controlling robots, which is why Dr. Kim’s expertise in human-computer interaction design through natural language was essential.,” Lee said. 

Kim is interviewing stakeholders from three South Korean hospitals to identify service and care pain points. The issues she’s identified so far relate to doctor-patient communication, a lack of emotional support for patients, and an excessive number of small tasks that consume nurses’ time.

“Our goal is to develop this robot in a very human-centered way,” she said. “One way is to give patients a way to communicate about the quality of their care and how the robot can support their emotional well-being.

“We found that patients often hesitate to ask busy nurses for small things like getting a cup of water. We believe this is an area a robot can support.”

The robot’s hardware will be built in Korea, while Ha and Kim will develop the software in the U.S.

Jong-hoon Park, CEO of Neuromeka, said in a press release the goal is to have a commercialized product as soon as possible. 

“Through this project, we will solve problems that existing collaborative robots could not,” Park said. “We expect the medical AI humanoid robot technology being developed will contribute to reducing the daily work burden of medical and healthcare workers in the field.”

Meta announced in February its partnership with the labs of professors Danfei Xu and Judy Hoffman on a novel computer vision-based algorithm called EgoMimic. It enables robots to learn new skills by imitating human tasks from first-person video footage captured by Meta’s Aria smart glasses. 

Xu’s Robot Learning and Reasoning Lab (RL2) displayed EgoMimic in action at ICRA May 19-23 at the World Congress Center in Atlanta.

Lawrence Zhu, Pranav Kuppili, and Patcharapong “Elmo” Aphiwetsa — students from Xu’s lab — used Egomimic to compete in a robot teleoperation contest at ICRA. The team finished second in the event titled What Bimanual Teleoperation and Learning from Demonstration Can Do Today, earning a $10,000 cash prize.

Teams were challenged to perform tasks by remotely controlling a robot gripper. The robot had to fold a tablecloth, open a vacuum-sealed container, place an object into the container, and then reseal it in succession without any errors.

Teams completed the tasks as many times as possible in 30 minutes, earning points for each successful attempt.

The competition also offered different challenge levels that increased the points awarded. Teams could directly operate the robot with a full workstation view and receive one point for each task completion. Or, as the RL2 team chose, teams could opt for the second challenge level.

The second level required an operator to control the task with no view of the workstation except for what was provided to through a video feed. The RL2 team completed the task seven times and received double points for the challenge level.

The third challenge level required teams to operate remotely from another location. At this level, teams could earn four times the number of points for each successful task completed. The fourth level challenged teams to deploy an algorithm for task performance and awarded eight points for each completion.

Using two of Meta’s Quest wireless controllers, Zhu controlled the robot under the direction of Aphiwetsa, while Kuppili monitored the coding from his laptop.

“It’s physically difficult to teleoperate for half an hour,” Zhu said. “My hands were shaking from holding the controllers in the air for that long.”

Being in constant communication with Aphiwetsa helped him stay focused throughout the contest.

“I helped him strategize the teleoperation and noticed he could skip some of the steps in the folding,” Aphiwetsa said. “There were many ways to do it, so I just told him what he could fix and how to do it faster.”

Zhu said he and his team had intended to tackle the fourth challenge level with the EgoMimic algorithm. However, due to unexpected time constraints, they decided to switch to the second level the day before the competition due to unexpected time constraints. 

“I think we realized the day before the competition training the robot on our model would take a huge amount of time,” Zhu said. “We decided to go for the teleoperation and started practicing.”

He said the team wants to tackle the highest challenge level and use a training model for next year’s ICRA competition in Vienna, Austria.

ICRA is the world’s largest robotics conference, and Atlanta hosted the event for the third time in its history, drawing a record-breaking attendance of over 7,000.

The tool, known as PROPEL, combines machine learning with optimization techniques to help manufacturers make better decisions in less time. It was created by researchers at the NSF AI Institute for Advances in Optimization, or AI4OPT, based at Georgia Tech under Tech AI (the AI Hub at Georgia Tech).

The technology is already being tested on real-world supply chain data provided by Kinaxis, a Canada-based company that supplies planning software to global manufacturers in industries ranging from automotive to consumer goods.

Vahid Eghbal Akhlaghi, senior research scientist at Kinaxis and former postdoctoral fellow at AI4OPT and the H. Milton Stewart School of Industrial and Systems Engineering (ISyE) at Georgia Tech, said, “Our industry partner has been instrumental in shaping PROPEL’s capabilities. By validating the approach with real operational data, we ensured it addresses true bottlenecks in supply chain planning.”

"PROPEL represents a leap forward in how we tackle massive, complex planning problems," said Pascal Van Hentenryck, lead researcher, the director of Tech AI and the NSF AI4OPT Institute, and the A. Russell Chandler III Chair and Professor at Georgia Tech with appointments in the colleges of engineering and computing. "By combining supervised and reinforcement learning, we can make near-optimal industrial-scale decisions, an order of magnitude faster."

Traditional supply chain planning problems are typically solved using mathematical models that require immense computing power—often too much to meet real-time business needs. PROPEL, short for Predict-Relax-Optimize using LEarning, reduces this burden by teaching the AI model to first eliminate irrelevant decisions and then fine-tune the solution to meet quality standards.

Reza Zandehshahvar, one of the paper’s co-authors and postdoctoral fellow with the NSF AI4OPT and the H. Milton Stewart School of Industrial and Systems Engineering (ISyE) at Georgia Tech, said the breakthrough lies not just in the AI algorithms but in how they're trained and deployed at scale.

“Many AI models struggle when applied to problems with millions of variables. PROPEL was built from the ground up to handle industrial complexity, not just academic examples,” Zandehshahvar said. “We’re seeing real improvements in both solution speed and quality.”

 In trials using Kinaxis’ historical industrial data, PROPEL achieved an 88% reduction in the time needed to find a high-quality plan and improved solution accuracy by more than 60% compared to conventional methods.

While many AI methods in supply chain rely on simulated data or simplified models, PROPEL’s performance has been validated using real-world scenarios, ensuring its reliability in high-stakes operational settings.

The Georgia Tech team says PROPEL could benefit industries that manage large, multi-tiered production networks, including pharmaceuticals, electronics, and heavy manufacturing. The researchers are now exploring partnerships with additional companies to deploy PROPEL in live environments.

Access the abstract on arXiv.

School of Electrical and Computer Engineering (ECE) Assistant Professor Amirali Aghazadeh and Ph.D. student Daniel Saeedi have developed AstroAgents, an AI system that analyzes mass spectrometry data — detailed chemical compositions from meteorites and Earth soil samples — to generate novel hypotheses about the origins of life on the planet. 

What sets AstroAgents apart is its use of agentic AI. Unlike traditional AI systems that perform fixed tasks, this agentic system is designed to pursue a scientific goal. It draws from astrobiology literature, interprets complex data, and proposes original ideas that researchers can investigate further. 

Their paper, recently featured in the journal "Nature", is opening new possibilities for how scientists explore questions that have remained unanswered for decades. 

In a special Q&A, Aghazadeh and Saeedi explain how AstroAgents analyzes space chemistry, what it’s revealing about the possible origins of life on Earth, and what they hope to explore next.

READ THE Q&A

 The National Science Foundation (NSF) has awarded Georgia Tech and its partners $20 million to build a powerful new supercomputer that will use artificial intelligence (AI) to accelerate scientific breakthroughs. 

Christopher Rozell, Julian T. Hightower Chaired Professor in the School of Electrical and Computer Engineering, will serve as the inaugural executive director of Georgia Tech’s new Institute for Neuroscience, Neurotechnology, and Society (INNS). 

Georgia Tech faculty and student experts advancing this field were in action in June at the globally renowned CVPR conference from IEEE and the Computer Vision Foundation. Georgia Tech was in the top 10% of all organizations for lead authors and the top 4% for number of papers. More than 2000 organizations had research accepted into CVPR's main program.

Watch the video and hear from Tech experts about what’s new and what’s coming next. Featured students include College of Computing experts Fiona Ryan, Chengyue Huang, Brisa Maneechotesuwan, and Lex Whalen.

These researchers in computer vision are showing how they are extending AI capabilities with image and video data.

HIGHLIGHTS:

- College of Computing faculty, from the Schools of Interactive Computing (IC) and Computer Science (CS), represented the majority of Tech's faculty in the CVPR papers program (8 of 10 faculty).

- IC faculty Zsolt Kira and Bo Zhu each coauthored an oral paper, the top 3% of accepted papers. IC faculty member Judy Hoffman coauthored two highlight papers, the top 20% of acceptances.

- Georgia Tech is in the top 10% of all organizations for number of first authors and the top 4% for number of papers. More than 2,000 organizations had research in the main program.

- Tech experts were on 30 research paper teams across 16 research areas. Topics with more than one Tech expert included:

• Image/video synthesis & generation
• Efficient and scalable vision
• Multi-modal learning
• Datasets and evaluation
• Humans: Face, body, gesture, etc.
• Vision, language, and reasoning 
• Autonomous driving
• Computational imaging

AfriMed-QA, developed by researchers from Georgia Tech and Google, could reduce the burden on African healthcare systems. 

The researchers said people in need of medical care file into overcrowded clinics and hospitals and face excruciatingly long waits with no guarantee of admission or quality treatment. There aren’t enough trained healthcare professionals available to meet the demand.

Some healthcare question-answer chatbots have been introduced to treat those in need. However, the researchers said there’s no transparent or standardized way to test or verify their effectiveness and safety.

The dataset will enable technologists and researchers to develop more robust and accessible healthcare chatbots tailored to the unique experiences and challenges of Africa. 

One such new tool is Google’s MedGemma, a large-language model (LLM) designed to process medical text and images. AfriMed-QA was used for training and evaluation purposes.

AfriMed-QA stands as the most extensive dataset that evaluates LLM capabilities across various facets of African healthcare. It contains 15,000 question-answer pairs culled from over 60 medical schools across 16 countries and covering numerous medical specialties, disease conditions, and geographical challenges. 

Tobi Olatunji and Charles Nimo co-developed AfriMed-QA and co-authored a paper about the dataset that will be presented at the Association for Computational Linguistics (ACL) conference next week in Vienna.

Olatunji is a graduate of Georgia Tech’s Online Master of Science in Computer Science (OMSCS) program and holds a Doctor of Medicine from the College of Medicine at the University of Ibadan in Nigeria. Nimo is a Ph.D. student in Tech’s School of Interactive Computing, where he is advised by School of IC professors Michael Best and Irfan Essa.

Focus on Africa

Nimo, Olatunji, and their collaborators created AfriMed-QA as a response to MedQA, a large-scale question-answer dataset that tests the medical proficiency of all major LLMs. That includes Google’s Gemini, OpenAI’s ChatGPT, and Anthropic’s Claude, among others.

However, because MedQA is trained solely on the U.S. Medical License Exams, Nimo said it is not adequate to serve patients in underdeveloped African countries nor the Global South at-large.

“AfriMed-QA has the contextualized and localized understanding of African medical institutions that you don’t get from Med-QA,” Nimo said. “There are specific diseases and local challenges in our dataset that you wouldn't find in any U.S.-based dataset.”

Olatunji said one problem African users may encounter using LLMs trained on MedQA is that they may advise unfeasible treatments or unaffordable prescription drugs.

“You consider the types of drugs, diagnostics, procedures, or therapies that exist in the U.S. that are quite advanced. These treatments are much more accessible, for example in the US, and Europe,” Olatunji said. “But in Africa, they’re too expensive and many times unavailable. They may cost over $100,000, and many people have no health insurance. Why recommend such treatments to someone who can’t obtain them?”

Another problem may be that the LLM doesn’t take a medical condition seriously if it isn’t predominant in the U.S.

“We tested many of these models, for example, on how they would manage sickle-cell disease signs and symptoms, and they focused on other “more likely” causes and did not rank or consider sickle cell high enough as a possible cause,” he said. “They, for example, don’t consider sickle-cell as important as anemia and cancer because sickle-cell is less prevalent in the U.S.”

In addition to sickle-cell disease, Olatunji said some of the healthcare issues facing Africa that can be improved through AfriMed-QA include:

• HIV treatment and prevention
• Poor maternal healthcare
• Widespread malaria cases
• Physician shortage
• Clinician productivity and operational efficiency

Google Partnership

Mercy Asiedu, senior author of the AfriMed-QA paper and research scientist at Google Research, has dedicated her career to improving healthcare in Africa. Her work began as a Ph.D. student at Duke University, where she invented the Callascope, a groundbreaking non-invasive tool for gynecological examinations

With her current focus on democratizing healthcare through artificial intelligence (AI), Asiedu, who is from Ghana, helped create a research consortium to develop the dataset. The consortium consists of Georgia Tech, Google, Intron, Bio-RAMP Research Labs, the University of Cape Coast, the Federation of African Medical Students Association, and Sisonkebiotik.

Sisonkebiotik is an organization of researchers that drives healthcare initiatives to advance data science, machine learning, and AI in Africa.

Olatunji leads the Bio-RAMP Research Lab, a community of healthcare and AI researchers, and he is the founder and CEO of Intron, which develops natural-language processing technologies for African communities.

In May, Google released MedGemma, which uses both the MedQA and Afri-MedQA datasets to form a more globally accessible healthcare chatbot. MedGemma has several versions, including 4-billion and 27-billion parameter models, which support multimodal inputs that combine images and text.

“We are proud the latest medical-focused LLM from Google, MedGemma, leverages AfriMed-QA and improves performance in African contexts,” Asiedu said. 

“We started by asking how we could reduce the burden on Africa’s healthcare systems. If we can get these large-language models to be as good as experts and make them more localized with geo-contextualization, then there’s the potential to task-shift to that.”

The project is supported by the Gates Foundation and PATH, a nonprofit that improves healthcare in developing countries.

Co-founded by Daniel Forcade and Hanna Bodnar, recent graduates of Georgia Tech’s Master of Science in Analytics program, Politheon is designed to overcome the limitations of standard artificial intelligence in providing businesses and other organizations with accurate and actionable information about legislative activity.

Bodnar credits the team's collaboration with Associate Professor Omar Asensio’s Data Science and Policy Lab in the Jimmy and Rosalynn Carter School of Public Policy for helping shape the platform.

"Collaborating with Professor Asensio’s lab was pivotal," Bodnar said. "As engineers, we had to expand our perspective beyond the technical implementation and deeply understand how public policy researchers and practitioners interpret legislative data. That interdisciplinary feedback helped us design a system that is both technically rigorous and policy-aware."

Forcade agreed, saying it took the combined resources of CREATE-X and the collaboration with Asensio’s lab to make Politheon what it is.

“CREATE-X gave us the business foundation to build and scale, while our collaboration with Professor Asensio’s lab helped us strengthen the scientific rigor behind the system. In policy, it's incredibly important to have testing, validation, and empirical grounding behind what you build.”

When it comes to understanding the potential impact of sometimes obliquely written legislation, precision and insight are vital. Publicly available large language models often struggle in these environments, sounding authoritative but often hallucinating in place of facts and failing to reason out the hidden impacts of legislation. Politheon, however, offers a potential solution, Asensio said.

"This is a very exciting use case for agentic AI in the context of evidence-informed policy," he said.

The project originally started as the final project for Bodnar and Forcade’s analytics program. Forcade said their instructors encouraged them to apply to CREATE-X to take the project further.

Forcade said CREATE-X liked the idea but asked them to talk to more experts. Forcade and Bodnar then reached out to Asensio.

Asensio was enthusiastic and invited them to present at his lab, where his team spent two and a half hours offering Forcade and Bodnar intensive constructive feedback. The duo has been collaborating with the lab ever since.

Asensio noted that this kind of cross-pollination is an embedded feature of his lab.

"We often start with data or policy solutions to guide technical development, and not the other way around," Asensio said. "This means our technologists learn to do causal inference and policy impact evaluation, and our policy scholars learn to code and train models and algorithms as part of their work."

That focus on critical evaluation aligns seamlessly with the founders' technical training.

"My background in mathematics and Georgia Tech’s Analytics program gave me a strong foundation in statistical modeling and machine learning systems," Bodnar said. "The program emphasizes not just building models but evaluating them rigorously. That mindset shaped how we designed Politheon, especially how we validate outputs and measure accuracy in a space where precision really matters."

The platform is already demonstrating its capabilities. Recent agent outputs include a large-scale scan of more than 25,000 Oregon bills, drawn from a broader searchable database of over 1.6 million state and federal bills, identifying emerging trends in artificial intelligence regulation. The system has also delivered validated, cross-jurisdictional analysis of “buy-now-pay-later” legislation in New York and Congress, with findings reviewed by senior government affairs professionals, tracing how the issue emerged and how it evolved over time.

The startup recently secured $100,000 in funding which helped build complete, and near real-time, data coverage across the federal government and  all U.S. states.

“The raise enabled us to bring in the live data stream,” Forcade said. “With real-time coverage in place, we’re now advancing pricing discussions and pilot rollouts with multiple organizations.”

Ultimately, the platform is designed to provide clarity amid the noise of modern governance.

"Policy moves quickly, and missing a compliance date or legislative shift can be costly," Bodnar said. "Our goal is to surface what’s relevant, explain why it matters, and provide clear citations to the original bills so teams can make informed decisions with confidence.”

Jennifer Kim and Mark Riedl recently received a $500,000 NSF grant to develop large language models (LLMs) that provide strength-based job coaching for autistic job seekers. 

The two Georgia Tech researchers work with Heather Dicks, a career development advisor in Georgia Tech’s EXCEL program, and other nonprofit organizations to provide job-seeking resources to autistic people.

Dicks said the average job search for people with autism can take three to six months in a good economy. It can take up to 18 months in a bad one. However, the new LLMs from Georgia Tech could help to reduce stress and fast-track these job seekers into employment.

Kim is an assistant professor who specializes in human-computer interaction technology that benefits neurodivergent people. Riedl is a professor and an expert in the development of artificial intelligence (AI) and machine learning technologies.

The team’s goal is to identify job-search pain points and understand how job coaches create better employment prospects for their autistic clients.

“Large-language models have an opportunity to support this kind of work if we can have more data about each different individual strength,” Kim said.

“We want to know what worked for them in specific settings at work, what didn’t work, and what kind of accommodations can better help them. That includes how they should prepare for interviews, how they can better represent their skills, how they can address accommodations they need, and how to write a cover letter. It’s a broad range.”

Dicks has advocated for neurodivergent people and helped them find employment for 20 years. She worked at the Center for the Visually Impaired in Atlanta before coming to Georgia Tech in 2017.

She said most nonprofits that support neurodivergent people offer career development programs and many contract job coaches, but limited coach availability often leads to long waitlists. However, LLMs could fill this availability gap to address the immediate needs of job seekers who may not have access to a job coach.

“These organizations often run at a slow pace, and there’s high turnover,” Dicks said. “An AI tool could get the job seeker quicker support. Maybe they don’t even need to wait on the government system.

“If they’re on a waitlist, it can help the user put together a resume and practice general interview questions. When the job coach is ready to work with them, they’re able to hit the ground running.”

Nailing the Interview

Dicks said the job interview is one of the biggest challenges for people with autism.

“They have trouble picking up on visual and nonverbal cues — the tone of the interview, figuring out the nuances that a question is hinting at,” she said. “They’re not giving the warm and fuzzy vibes that allow them to connect on a personal level.”

That’s why Kim wants the models to reflect a strength-based coaching approach. Strength-based coaching is particularly effective for individuals with autism. Many possess traits that employers value. These include:

• Close attention to detail
• Strong technical proficiency
• Unique problem-solving perspectives

“The issue is that they don’t know how these strengths can be applied in the workplace,” Kim said. “Once they understand this, they can communicate with employers about their strengths and the accommodations employers should provide to the job seeker so they can successfully apply their skills at work.”

Handling Rejection

Still, Kim understands that candidates will need to handle rejection to make it through the search process. She envisions LLMs that help them refocus their energy and regain their confidence after being turned down.

“When you get a lot of rejection emails, it’s easy to feel you’re not good enough,” she said. “Being constantly reminded about your strengths and their prior successes can get them through the stressful job-seeking process.”

Dicks said the models should also be able to provide feedback so that candidates don’t repeat mistakes.

“It can tell them what would’ve been a better answer or a better way to say it,” Dicks said. “It can also encourage them with reminders that you get 100 noes before you get a yes.”

You’re Hired, Now What?

Dicks said the role of a job coach doesn’t end the moment a client is hired. Government-contracted job coaches may work with their clients for up to 90 days after they start a new job to support their transition.

However, she said, sometimes that isn’t enough. Many companies have probationary periods exceeding three months. Autistic individuals may struggle with on-the-job training or communicating what accommodations they need from their new employer. 

These are just a few gaps an AI tool can fill for these individuals after they’re hired.

“I could see these models evolving to being supportive at those critical junctures of the probationary period being over or the one-year job review or the annual evaluation that everyone dreads,” she said.

Dicks has an average caseload of 15 students, whom she assists in landing jobs and internships through the EXCEL program.

EXCEL provides a mentorship program for students with intellectual and developmental disabilities from the time they set foot on campus through graduation and beyond.

For more information and to apply, visit EXCEL’s website.

Georgia Tech’s Qi Tang is building machine learning (ML) models to accelerate nuclear fusion research, making it more affordable and more accurate. Backed by a grant from the U.S. Department of Energy (DOE), Tang’s work brings clean, sustainable energy closer to reality.

Tang has received an Early Career Research Program (ECRP) award from the DOE Office of Science. The grant supports Tang with $875,000 disbursed over five years to craft ML and data processing tools that help scientists analyze massive datasets from nuclear experiments and simulations.

Tang is the first faculty member from Georgia Tech’s College of Computing and School of Computational Science and Engineering (CSE) to receive the ECRP. He is the seventh Georgia Tech researcher to earn the award and the only GT awardee among this year’s 99 recipients.

More than a milestone, the award reflects a shift in how nuclear research is done. Today, progress depends on computing and data science as much as on physics and engineering.

“I am honored and excited to receive the ECRP award through DOE’s Advanced Scientific Computing Research program, an organization I care about deeply,” said Tang, an assistant professor in the School of CSE. 

“I am grateful to my former colleagues at Los Alamos National Laboratory and collaborators at other national laboratories, including Lawrence Livermore, Sandia, and Argonne. I am also thankful for my Ph.D. students at Georgia Tech, whose dedication and creativity make this award possible.”

[Related: New Faculty Applies High-Performance Computing, Scientific Machine Learning Interests to Studies in Plasma Physics]

A problem in nuclear research is that fusion simulations are challenging to understand and use. These simulations generate enormous datasets that are too large to store, move, and analyze efficiently.

In his ECRP proposal to DOE, Tang introduced new ML methods to improve the analysis and storage of particle data.

Tang’s approach balances shrinking data so it is easier to store and transfer while preserving the most important scientific features. His multiscale ML models are informed by physics, so the reduced data still reflects how fusion systems really behave.

With Tang’s research, scientists can run larger, more realistic fusion models and analyze results more quickly. This accelerates progress toward practical fusion energy.

“In contrast to generic black-box-type compression tools, we aim at preserving the intrinsic structures of the particle dataset during the data reduction processes,” Tang said. 

“Taking this approach, we can meet our goal of achieving high-fidelity preservation of critical physics with minimum loss of information.”

Computing is essential in modern research because of the amount of data produced and captured from experiments and simulations. In the era of exascale supercomputers, data movement is a greater bottleneck than actual computation.

DOE operates three of the world’s four exascale supercomputers. These machines can calculate one quintillion (a billion billion) operations per second.

The exascale era began in 2022 with the launch of Frontier at Oak Ridge National Laboratory. Aurora followed in 2023 at Argonne National Laboratory. El Capitan arrived in 2024 at Lawrence Livermore National Laboratory.

With Tang’s data reduction approaches, all of DOE’s supercomputers spend more time on science and less time waiting for data transfers.

“Qi’s work in computational plasma physics and nuclear fusion modeling has been groundbreaking,” said Haesun Park, Regents’ Professor and Chair of the School of CSE. 

“We are proud of Qi and what this award means for him, Georgia Tech, and the Department of Energy toward leveraging computation to solve challenges in science and engineering, such as sustainable energy."

Previous Georgia Tech recipients of DOE Early Career Research Program awards include:

Itamar Kimchi, assistant professor, School of Physics

Sourabh Saha, assistant professor, George W. Woodruff School of Mechanical Engineering

Wenjing Lao, associate professor, School of Mathematics

Ryan Lively, Thomas C. DeLoach Professor, School of Chemical & Biomolecular Engineering

Josh Kacher, associate professor, School of Materials Science and Engineering

Devesh Ranjan, Eugene C. Gwaltney Jr. School Chair and professor, Woodruff School of Mechanical Engineering

Georgia Stoica is one of 38 Ph.D. students worldwide researching machine learning who were named a 2025 Google Ph.D. Fellow.

Stoica is designing AI training methods that bypass fine-tuning, which is the process of adapting a large pre-trained model to perform new tasks. Fine-tuning is one of the most common ways engineers update large-language models like ChatGPT, Gemini, and Claude to add new capabilities. 

If an AI company wants to give a model a new capability, it could create a new model from scratch for that specific purpose. However, if the model already has relevant training and knowledge of the new task, fine-tuning is cheaper.

Stoica argues that fine-tuning still uses large amounts of data, and that other methods can help models learn more effectively and efficiently.

“Full fine-tuning yields strong performance, but it can be costly, and it risks catastrophic forgetting,” Stoica said. “My research asks if we can extend a model’s capabilities by imbuing it with the expertise of others, without fine-tuning?

“Reducing cost and improving efficiency is more important than ever. We have so many publicly available models that have been trained to solve a variety of tasks. It’s redundant to train a new model from scratch. It’s much more efficient to leverage the information that already exists to get a model up to speed.”

Stoica said the solution is a cost-effective method called model merging. This method combines two or more AI models into a single model, improving performance without fine-tuning.

On a basic level, Stoica said an example would be combining a model that is efficient at classifying cats with one that works well at dogs.

“Merging is cheap because you just take the parameters, the weights of your existing models, and combine them,” he said. “You could take the average of the weights to create a new model, but that sometimes doesn’t work. My work has aimed to rearrange the weights so they can communicate easily with each other.”

Through his Google fellowship, Stoica seeks to apply model merging to create a cutting-edge vision encoder. A vision encoder converts image or video data into numerical representations that computers can understand. This enables tasks such as image or facial recognition and generative image captioning.

“I want to be at the frontier of the field, and Google is clearly part of that,” Stoica said. “The vision encoder is very large-scale, and Google has the infrastructure to accommodate it.”