Ramin Zabih
Ramin Zabih

Ramin Zabih

Professor of Computer Science, Cornell University and Cornell Tech
Research Scientist, Google DeepMind

I split my time between Cornell and Google DeepMind, and I served as the Executive Director of arXiv from 2022 through 2026. My research interests have always been in computer vision; my main focus now is bringing techniques from causal inference into it: in particular, asking when a generative model that produces convincing video can be trusted to answer "what if" questions, and what to do when it cannot.

Current research: causal inference meets computer vision

Imagine a robot that kicks a ball in a crosswind. When a gust pushes the ball left, the controller almost always kicks right; it tries a left kick in a left gust about once in a thousand gusts. A video model trained on this robot's footage will render the usual outcome beautifully, but when a planner asks it what happens if I kick left in a left gust?, it has barely seen that situation and quietly makes something up. The planner then picks the imagined kick. The model is an excellent video generator and a poor decision-maker, and nothing in its training loss tells you which situations it is guessing about.

Most of my current work, with Michele Santacatterina and my students, grows out of this picture. The tools come from causal inference, where "what would have happened under a different treatment" is the central question and ideas such as overlap, importance weighting and doubly robust estimation have been developed to answer it carefully; the objects are the large video models of modern computer vision. We study counterfactual world models: generative models that are asked not "what usually happens next" but "what would happen if I did something else." Three questions organize the work:

The recipe, "reweight, check, collect," applies well beyond video: to off-policy evaluation in reinforcement learning, to treatment-effect estimation in medicine, and to any setting where a learned simulator is used to choose among actions it has rarely seen. Related projects in the group look at how rare the data can get before the certificate fails, what happens when a model is trained on its own outputs, and how to choose which experiments to run in the real world when imagination is not enough.

A note for prospective PhD students

I work with students at both Cornell campuses (Ithaca and Cornell Tech in New York City). If you are considering a PhD with me, the thing to know is that my main interest is applying techniques from causality to computer vision, along the lines described above.

The work mixes modern generative modelling (diffusion and video transformers, trained at real scale) with ideas from causal inference and reinforcement learning (importance weighting, doubly robust estimation, off-policy evaluation). You do not need to know all of this already, but you should be comfortable with probability and want to write careful, falsifiable experiments as much as you want to train large models. A typical first project takes one piece of the recipe (one diagnostic, one fix, one certificate) and tests it end to end, from a controlled simulator to real robot or driving video.

Applications go through the regular Cornell CS PhD admissions process, which covers both campuses; if this direction is what you want to work on, say so in your statement and name me. I read those applications closely. I generally cannot answer individual emails about admission, but a short note pointing me to something you have built or written is welcome.

Students

Current PhD students: Jingsen Zhu, Khiem Duc Pham, Ali Behrouz and Michelle Shu (on leave).

Postdoc: Raphael Kim

Alumni: PhD: Richard Strong Bowen, Akshay Bhat, Alex Fix, Charles Herrmann, Jing Huang, Vera Kettnaker, Junhwan Kim, Vladimir Kolmogorov, Ashish Raj, Gurmeet Singh, Olga Veksler, Chen Wang and Jie Zhu. Postdoc: Yuri Boykov. Undergraduates: Brian Cody, Chris Danis, Devin Kennedy, Justin Miller, Greg Pass, Brian Rogan and Justin Voskuhl.

Background: computer vision

My research has always focused on computer vision and its applications; in the past I was interested in medical imaging, but I am no longer actively working on that topic. Much of my work concerned early vision problems such as stereo and motion, and my group developed new algorithms for these problems based on graph cuts. That line of work received the Helmholtz Prize (ICCV 2011 and 2013) and the Koenderink Prize (ECCV 2012) for contributions that have stood the test of time. I am a Fellow of the ACM and of the IEEE. A short list of selected vision papers is here; Google Scholar is the complete record.

Open access and service

I have spent more than a decade bringing open access to the computer vision community. I founded and am president of the Computer Vision Foundation, the non-profit that co-sponsors CVPR, ICCV and WACV with the IEEE Computer Society and provides the open-access versions of their papers; this work received the 2016 Everingham Prize. From 2022 until recently I led arXiv as its Executive Director.

I was Editor-in-Chief of IEEE Transactions on Pattern Analysis and Machine Intelligence (2009-2012) and chaired the PAMI-TC (2013-2015). I was general co-chair of CVPR 2013, 2020 and 2024, ICCV 2025 and ECCV 2018, program co-chair of CVPR 2007, and an area chair for CVPR, ICCV and ECCV many times. With Gerard Medioni I founded and run the International Workshop on Computer Vision, which has met every two years since 2008.