NYU Center for Data Science

Eero P. Simoncelli, a founding member of CDS and Professor of Neural Science, Mathematics, Data Science, and Psychology, was elected to the National Academy of Sciences.

He was among 120 new members recognized for “distinguished and continuing achievements in original research,” according to the academy’s announcement.

Simoncelli studies the representation of sensory information in brains and machines, examining how neurons encode visual information, how those representations shape perceptual capabilities, and how these principles can be used to build computational systems for processing visual data.

https://www.nasonline.org/news/2026-nas-election/

In adversarial imperfect-information games like poker, simple deep reinforcement learning was long thought to fail without help from specialized algorithms.

A new paper presented at ICLR 2026 — a multi-institution collaboration spanning NYU, UT Austin, UC Berkeley, MIT, and CMU — found that with proper tuning, generic policy gradient methods like PPO match or beat those specialized approaches.

Co-led by Max Rudolph (UT Austin), Nathan Lichtlé (UC Berkeley), and Sobhan Mohammadpour (MIT), with CDS-affiliated Assistant Professor of Civil and Urban Engineering Eugene Vinitsky among the senior authors, the team ran over 7,000 training runs across five large games, including 3x3 Dark Hex and Phantom Tic-Tac-Toe.

They also released the first broadly accessible exact exploitability benchmarks, so other researchers can verify the result for themselves.

https://arxiv.org/abs/2502.08938

Reevaluating Policy Gradient Methods for Imperfect-Information Games In the past decade, motivated by the putative failure of naive self-play deep reinforcement learning (DRL) in adversarial imperfect-information games, researchers have developed numerous DRL algorithms based on fictitious play (FP), double oracle (DO), and counterfactual regret minimization (CFR). I...

CDS welcomed admitted PhD students to campus for a visit that included a Lightning Talks session.

Faculty and current PhD students shared quick looks at their work across a wide range of topics.

CDS Assistant Professor (starting Sept 2026) Jaume Vives i Bastida presented work on new doubly-robust estimation methods.

CDS Assistant Professor Grace Lindsay discussed studying attention at the interface of biology and AI.

CDS Faculty Fellow Mateo Dulce Rubio spoke about continuous learning in humanitarian demining.

CDS Assistant Professor Qi Lei covered learning from weak signals and the road to self-improving AI.

CDS PhD students Yilun Kuang, Wentao Wang, and Uriel Martinez Leon rounded out the session with talks on joint-embedding predictive architectures, rapid word learning through meta in-context learning, and statistical optimal transport.

The cheap, fast head CT scans used in emergency rooms can do more than detect strokes and bleeds.

CDS PhD student Haoxu (Howard) Huang and CDS-affiliated Associate Professor Narges Razavian, along with CDS PhD alum student Weicheng (Jack) Zhu and colleagues, developed FM-HCT, a 3D foundation model trained on 361,663 head CT scans that outperforms commercial alternatives from Google and others.

The model identifies Alzheimer’s disease and related dementias with strong accuracy from CTs, opening a path to earlier detection in emergency rooms and underserved communities where MRI access is limited.

Published in Nature Biomedical Engineering.

https://nyudatascience.medium.com/head-ct-foundation-model-outperforms-commercial-alternatives-detects-diseases-beyond-hemorrhage-b6fa8b2d10dd

Head CT Foundation Model Outperforms Commercial Alternatives, Detects Diseases Beyond Hemorrhage The cheap, fast scans that emergency rooms use to check for strokes and brain bleeds may also detect Alzheimer’s disease and related…

Relying too heavily on algorithms for answers could eventually cause human progress to plateau.

University of Cambridge researcher Katherine M. Collins, former CDS Faculty Fellow Umang Bhatt, and CDS Faculty Fellow Ilia Sucholutsky mapped the effects of artificial intelligence on human knowledge.

They applied an evolutionary theory known as Rogers’ paradox to test what happens when individuals stop seeking out new information and default to social learning.

Their simulations showed that the widespread availability of cheap artificial intelligence systems does not improve our collective understanding of the world over the long term.

When the researchers factored in the potential for cognitive decline, the population’s knowledge equilibrium dropped below the original baseline.

Read more about how the team is rethinking the future of human-machine interaction.

https://nyudatascience.medium.com/will-relying-on-ai-stop-human-progress-a-new-study-tests-an-evolutionary-paradox-6a94a1950b78

Will Relying on AI Stop Human Progress? A New Study Tests an Evolutionary Paradox Overusing artificial intelligence for answers could eventually drag society’s collective knowledge down to a state worse than if the…

CDS MS student Nikolas Prasinos had his first abstract accepted for digital poster presentation at ISMRM 2026, the annual meeting of the International Society for Magnetic Resonance in Medicine.

Combining MRI with EEG lets researchers see both where and when brain activity happens — useful for studying conditions like epilepsy and Alzheimer's. But the MRI scanner produces electrical interference that drowns out the EEG signal, making it hard to read what the brain is actually doing.

Prasinos and colleagues trained a deep learning model to clean up that interference, recovering brain and heart signals that closely match recordings taken when the scanner is off.

The work, "Deep Residual Learning for Artifact Suppression in Simultaneous Sodium MRI–EEG Acquisition," was conducted with collaborators at NYU Langone's Center for Biomedical Imaging, the Center for Advanced Imaging Innovation and Research (CAI²R), the Comprehensive Epilepsy Center, and the Alzheimer's Disease Research Center, under the supervision of NYU Langone Assistant Professor of Radiology Yongxian Qian.

NYU Langone Research Scientist Ying-Chia (Amy) Lin presented the poster at ISMRM 2026 on May 14 in Cape Town.

The second edition of the Workshop on Scientific Methods for Understanding Deep Learning (Sci4DL) drew a packed room throughout the day at ICLR 2026 in Rio de Janeiro.

The workshop was co-organized by CDS PhD alumni Zahra Kadkhodaie and Sanae Lotfi, CDS Instructor Florentin Guth, CDS-associated Professor Eero Simoncelli, and collaborators from UPenn, UCL, the University of Amsterdam, Imbue/UC Berkeley, and Pacific Northwest National Lab.

The workshop brought together researchers from signal processing, computer science, and physics to examine deep learning through the lens of the scientific method.

Speakers included CDS Silver Professor Julia Kempe, alongside David Bau (Northeastern), Jeremy Cohen (Flatiron Institute), Richard Baraniuk (Rice & OpenStax), and Matthieu Wyart (Johns Hopkins & EPFL).

https://scienceofdlworkshop.github.io/2026/

CDS Silver Professor of Computer Science, Mathematics, and Data Science Julia Kempe co-authored new research with Charles Arnal, Vivien Cabannes, Taco Cohen, and Remi Munos on making reinforcement learning for large language models more computationally efficient.

The paper explored how AI systems can reuse past training data instead of constantly generating new information from scratch. The researchers found that this approach reduced computing costs while maintaining strong model performance.

The work addressed a growing challenge in AI research as training advanced language models becomes increasingly expensive.

Accepted to the ICML '26 conference.

https://arxiv.org/abs/2604.08706

Efficient RL Training for LLMs with Experience Replay While Experience Replay - the practice of storing rollouts and reusing them multiple times during training - is a foundational technique in general RL, it remains largely unexplored in LLM post-training due to the prevailing belief that fresh, on-policy data is essential for high performance. In thi...

How do you study a coral reef without disturbing it?

CDS Capstone students Amaan Mansuri, Vishwa Raval, and Shravan Khunti built a computer vision system that identifies reef fish in real time from underwater robot footage — surgeonfish, parrotfish, and grunts — shot off the coast of Barbados.

The system serves a dual purpose:

For tourists, it powers a virtual diving experience that reduces foot traffic on fragile ecosystems while widening access for non-swimmers and people with disabilities.

For marine biologists, the same pipeline automates species detection across hours of underwater footage, replacing manual frame-by-frame labeling with real-time identification.

Working with the UNDP Accelerator Lab for Barbados and the Eastern Caribbean, the team’s prototype was selected by the Japan Cabinet Office for the 2025 Japan SDGs (Sustainable Development Goals) Challenge.

https://nyudatascience.medium.com/underwater-ai-cds-capstone-team-builds-fish-detection-system-for-sustainable-caribbean-tourism-28f2db55b92a

Underwater AI: CDS Capstone Team Builds Fish Detection System for Sustainable Caribbean Tourism By 10 meters underwater, the red channel of visible light has essentially disappeared, leaving behind the blue-green tint familiar from any…