UMD Center for Machine Learning

The University of Maryland Center for Machine Learning uses powerful computing tools to tackle challenges in big data, computer vision, health care and finance.

[UMD Center for Machine Learning @ml_umd]

Led by Mohammad Hajiaghayi (@MTHajiaghayi) and Reza Ghanadan, researchers from @UofMaryland & @MIT will develop AI-driven “co-authors” capable of helping mathematicians break down extremely complex problems and uncover new insights.

https://t.co/7AGMZSy3nB

[UMIACS @umiacs]

Can we make AI forget harmful data without erasing its entire intelligence?

Researchers @UofMaryland and the Max Planck Institute for Software Systems (@mpi_sws_) have developed a new AI “unlearning” method to address this hurdle in LLMs. The method can selectively remove sensitive data—like private information, copyrighted text or misinformation—without damaging the model’s reasoning abilities.

Led by @FeiziSoheil and @RezaeiKeivan, the team’s method works like a precision rollback tool for AI memory, allowing developers to reverse the influence of specific data instead of retraining massive models from scratch.

The research was presented at #ICLR2026 (@iclr_conf) in Rio de Janeiro. Read more: https://t.co/IWLcpGNPi6

[Furong Huang @furongh]

AI can write code, pass exams, and generate videos.

But ask a robot to pour almonds into a bowl, and it may still fail.

Why?

One reason: robots are often using visual encoders trained for the internet — not for action.

Our new work asks: do robots have the wrong eyes?

[UMIACS @umiacs]

Researchers @UofMaryland are advancing robotics with a new initiative designed to enable humanoid systems to perform complex, real-world household tasks with greater autonomy and reliability.

Built on @NVIDIAAI infrastructure, the project integrates breakthroughs in trustworthy machine learning, sequential decision-making and generative AI to create robotic systems that can reason, adapt and act in dynamic home environments.

The effort is led by Furong Huang (@furongh) and Tom Goldstein (@tomgoldsteincs) of @umiacs, which will install and maintain the new computing infrastructure in its high-performance data center.

Learn more: https://t.co/IO46sBESgX

[Furong Huang @furongh]

Excited that our project HomeGraph in collaboration with @tomgoldsteincs has been selected for the NVIDIA Academic Grant Program!

We’re building tool-native GR00T humanoid robots using a unified scene–skill graph for long-horizon household autonomy.

Grateful for NVIDIA’s support with RTX PRO 6000 Blackwell GPUs and Jetson AGX Thor to push this research forward.

Special shoutout to @ruijie_zheng12, now at NVIDIA’s GEAR Lab, who played a major role in early GR00T work while in my lab. Excited to continue collaborating.

Looking forward to building on our partnership with @DrJimFan and @yukez and advancing the future of generalist robotics.

Research supported by the NVIDIA Academic Grant Program.

#NVIDIAGrant @NVIDIAAIDev

[Mohammad Hajiaghayi (IG@mhajiaghayi,YT:hajiaghayi) @MTHajiaghayi]

My 21st Ph.D. student, Suho Shin, just defended his thesis Game Theory & joins Stanford and MIT for postdocs. All 21 of my students’ theses focus on challenging deep theoretical foundations applied to auctions, games, big data, networks, algorithms, distributed & intelligent sys

[UMIACS @umiacs]

Tom Goldstein (@tomgoldsteincs) was honored at the 2026 Maryland Research Excellence Celebration for his influential work in machine learning and AI security. With more than 9,000 Google Scholar citations annually and an h-index exceeding 90, he ranks among the most highly cited scholars in his field. Read more about the event: https://t.co/8DT78rHN2g

[TRAILS @trails_ai]

TRAILS is launching 11 Broader Impact Awards to expand access to trustworthy AI. These grants—up to $25K apiece—support projects @UofMaryland, @MorganStateU and @GWtweets to empower diverse stakeholder communities to engage with and influence the future of AI. https://t.co/qyKkTa6w6m

[UMIACS @umiacs]

An ultra-low-power spatial sensing system designed for miniature mobile robots, developed by @umiacs-affiliated researchers, was recently featured @CACMmag: https://t.co/wZs11Mw3ab

Co-authored by Yang Bai (@FighterYangbai8), Nakul Garg (@nakulgarg22) and Nirupam Roy(@NeeeRooo9), the paper explains how they designed the system to meet the strict power and hardware constraints of micro-robot platforms.

Bai completed her Ph.D. @umdcs and is now a machine learning scientist @Apple. Garg, who also earned his Ph.D. @umdcs, is now an assistant professor @RiceECE. Roy is an associate professor @umdcs with an appointment @umiacs.

[UMIACS @umiacs]

Research by computer science Ph.D. student Seungjae Lee (@JayLEE_0301) will enable robots to learn not only from their own physical experiences, but also from the vast reservoir of human activity captured online.

https://t.co/pfMrVZj6xM

[UMD Center for Machine Learning @ml_umd]

Today is #GivingDayatUMD! Gifts to the UMD Center of Machine Learning support the center’s activities, enabling cutting-edge developments in machine learning research and realizing the world-changing application of those discoveries.

👉Give here: https://t.co/P67YhgxAy8

[TRAILS @trails_ai]

It's not too late to register for #TRAILSCon2026 at George Washington University on Wednesday, March 4. This year, TRAILSCon will explore real-world approaches to evaluating AI performance, risk, and impact. Learn more: https://t.co/rNoAAgUWc6

[UMD Center for Machine Learning @ml_umd]

#GivingDayUMD is just two days away! Help #ScienceTerps move fearlessly forward—save the date for March 4.

Gifts to @ml_umd will enable cutting-edge developments in machine learning research.

https://t.co/5odAndnHag

[GW Engineering @gwuengineering]

How do we measure AI’s performance, risk & real-world impact? 🤖

Find out at TRAILSCon 2026, March 4–5 at GW. Panels, roundtables & workshops focused on responsible AI design & governance.

🔗 Register: https://t.co/HK1IPjQrqw
#TRAILSCon2026 #AI #ResponsibleAI #FutureOfTech

[Furong Huang @furongh]

Back when working on FLARE: Robot Learning with Implicit World Modeling
📄 https://t.co/ekqFw5fF4X

We realized something important:

👉 Co-training is not just a trick. It’s a scaling law for robotics.

By aligning latent future representations, FLARE showed that mixing robot demonstrations with human egocentric video unlocks surprising generalization — even to unseen objects with minimal robot data.

That insight stayed with us.

Now Ruijie has graduated from our lab and joined NVIDIA GEAR Lab — one of the frontier labs in modern robotics.

And they’re taking this idea further.

Why is co-training powerful?

• Robot data provides precise action grounding
• Human video provides massive visual diversity
• Latent alignment bridges embodiment gaps

You don’t need perfect action labels.
You need the right representation.

The next generation of VLAs will not just react — they will anticipate.

Proud former advisor moment 🚀

#Robotics #WorldModels #VLA #EmbodiedAI #DiffusionModels

[UMIACS @umiacs]

"I think we are entering into a new era of security," Tom Goldstein, a professor of computer science and director of the Center for Machine Learning at the University of Maryland, told Yahoo Finance.

https://t.co/RMFRIGla63

[UMIACS @umiacs]

A multi-institutional team—which includes @UofMaryland researchers Yiannis Aloimonos and students Aadi Palnitkar and Arjun Suresh—has created a rhino-detection system that pairs satellite data with AI to boost conservation efforts.

https://t.co/NVKC99IXAh

[UMIACS @umiacs]

After nearly 30 years of working in the @umiacs business office, Yerty Valenzuela is retiring. As director of research program administration, Yerty has been a cornerstone of UMIACS operations, supporting research programs and managing key accounts with dedication and expertise. Congratulations to Yerty on her retirement!

[ECE at UMD @eceumd]

Professor Uzi Vishkin Receives 2026 IEEE Computer Society Charles Babbage Award: Recognized for his seminal contributions to the parallel random-access machine theory and as the inventor of fundamental work-efficient parallel algorithms. https://t.co/6gS3WvcW0a

[Sanghamitra Dutta @Sangha26Dutta]

Presenting today at #NeurIPS2025 : Can explainability help with model compression? — A new strategy of few-shot distillation of LLMs that cleverly use counterfactual explanations as boundary pegs.
Achieves superior performance using only half the original samples + Theoretical guarantees using Hausdorff distance

📄 Paper: https://t.co/I1jC2EnIWK
Code: https://t.co/bxeDl5hi05

For more details, feel free to stop by our poster:
🗓 Wed, Dec 3, 2025
⏰ 11:00 AM – 2:00 PM PST
📍 Exhibit Hall C,D,E — # 2711