Robert Geirhos

Robert Geirhos is a computer vision researcher who studies why neural networks see the world the way they do by comparing them with human vision. As a graduate student at the University of Tübingen, he was lead author of the 2019 ICLR paper showing that ImageNet-trained CNNs favor texture, and in 2020 he first-authored the perspective "Shortcut Learning in Deep Neural Networks". His 2022 data pruning paper with Ben Sorscher, Shashank Shekhar, Surya Ganguli and Ari Morcos won an Outstanding Paper Award at NeurIPS, and he received the 2023 ELLIS PhD Award. After joining Google Brain in Toronto in 2022, he coauthored the 2025 paper "Video models are zero-shot learners and reasoners", which showed that Veo 3 solves vision tasks it was not explicitly trained for. He is a Staff Research Scientist at Google DeepMind in Zürich, where he pursues visual intelligence with video models such as Veo and Omni.

Now at: Google DeepMind

Timeline

2019

Texture bias in ImageNet CNNs

As a graduate student at the University of Tübingen, in the labs of Matthias Bethge and Felix Wichmann, he is lead author of the ICLR 2019 oral paper "ImageNet-trained CNNs are biased towards texture". It shows that ImageNet-trained CNNs strongly favor recognizing textures, and that a ResNet-50 trained on Stylized-ImageNet learns a shape-based representation with a much better fit to human behavior. Quanta Magazine reports on the work in July 2019.
2020

The shortcut learning perspective

He is first author of "Shortcut Learning in Deep Neural Networks", written with Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge and Felix Wichmann and published in Nature Machine Intelligence. The perspective traces many of deep learning's problems to one cause, shortcut learning, and offers recommendations for model interpretation and benchmarking.
July 2021

A FAIR research internship

In July 2021, he submits his doctoral thesis and joins FAIR at Meta AI as a summer PhD research intern, working with Ari Morcos and Surya Ganguli.
2022

Dissertation award, then Google Brain

The University of Tübingen recognizes his dissertation, "To err is human? A functional comparison of human and machine decision-making", with its 2022 dissertation award. In August 2022, he joins Google Brain in Toronto.
November 2022

A NeurIPS Outstanding Paper

"Beyond neural scaling laws: beating power law scaling via data pruning", which he coauthors with Ben Sorscher, Shashank Shekhar, Surya Ganguli and Ari Morcos, receives an Outstanding Paper Award at NeurIPS 2022. The paper shows that error could fall faster than a power law as datasets grow if a high-quality data pruning metric ranks which examples to discard, and it benchmarks ten pruning metrics on ImageNet.
2023

The ELLIS PhD Award

The ELLIS Society names him a recipient of its 2023 PhD Award for work on how machines see the world compared with human visual perception, and he receives it in December 2023 at the ELLIS General Assembly. He continues his research as a Research Scientist at Google DeepMind in Toronto.
September 2025

Veo 3 as a vision model

In September 2025, he coauthors "Video models are zero-shot learners and reasoners" as one of its senior authors. The paper shows that Veo 3 can solve tasks it was not explicitly trained for, including segmenting objects, detecting edges and editing images, and argues that video models are on a path to becoming unified, generalist vision foundation models.
As of October 2026

Staff Research Scientist in Zürich

He is a Staff Research Scientist at Google DeepMind in Zürich, where his research focus is solving visual intelligence with video models such as Veo and Omni. In October 2026, he has an open student researcher position for 2027.