Dumitru Erhan — Google DeepMind

Now

At Google DeepMind, I am a Senior Research Director and co-lead of Gemini Omni. I was also a lead of Veo and Phenaki, focusing on generative video foundation models and world models that learn how the physical world behaves.

Gemini Omni (2026)

Veo 3 (2025)

Phenaki (2022)

FitVid / SV2P (2018+)

Google Brain

While in Google Brain, I focused on video prediction and model-based reinforcement learning—training agents entirely inside learned world models (SimPLe, SV2P, FitVid). We also explored simulation-to-reality domain adaptation with Domain Separation Networks and PixelDA.

Foundations

Earlier at Google, I worked on foundational vision architectures and core behaviors—including Inception (GoogLeNet), SSD, adversarial examples, and Show and Tell. I also spent a few years on the early Google Photos team, building visual search models ahead of its 2015 launch.

Previously

I did my PhD at the University of Montreal (now Mila) on deep learning with Yoshua Bengio, studying deep representations (feature visualization, unsupervised pre-training). Before Google, I spent 2011–2012 at Yahoo Labs working on search ranking.

Selected Papers

The through-line: understanding representation learning → interpreting deep networks → scaling visual recognition/detection → connecting vision and language → discovering failure modes → making perception fast → learning under imperfect supervision/domain shift → predicting video/world dynamics → generative video → unified multimodal generation.

A recurring theme has been taking research all the way from new ideas and benchmarks to systems and products used at scale.

Going Deeper with Convolutions (CVPR 2015)
Introduced the 22-layer Inception architecture; won ImageNet 2014 and the CVPR 2025 Test of Time Award (Longuet-Higgins Prize).
SSD: Single Shot MultiBox Detector (ECCV 2016)
Eliminated region proposals to enable fast, single-stage real-time object detection; won the ECCV 2026 Test of Time Award.
Intriguing Properties of Neural Networks (ICLR 2014)
Discovered adversarial examples and unlocalized representations, helping establish adversarial robustness as a major area of study.
Show and Tell: A Neural Image Caption Generator (CVPR 2015)
First end-to-end deep neural network translating visual scenes to natural language descriptions; tied for 1st place in the COCO 2015 challenge.
Phenaki: Variable Length Video Generation from Open Domain Text (ICLR 2023)
Pioneered causal spatio-temporal video tokens and masked transformers to synthesize variable-length video stories from sequential text prompts.
Model-Based Reinforcement Learning for Atari (SimPLe) (ICLR 2020)
Showed that agents can learn complex behaviors with high sample efficiency by training entirely inside learned video prediction world models.
Domain Separation Networks (NeurIPS 2016) & PixelDA (CVPR 2017)
Disentangled shared vs. domain-private representations, and introduced generative pixel-level domain adaptation from simulation to reality.
Why Does Unsupervised Pre-training Help Deep Learning? (JMLR 2010)
Landmark empirical study demonstrating that pre-training acts as an unusual regularizer, conditioning optimization toward superior basins of attraction.
Visualizing Higher-Layer Features of a Deep Network (2009)
Introduced activation maximization for visualizing receptive-field preferences, an early approach to neural network feature visualization.

A complete list of 100+ publications can be found on Google Scholar and ResearchGate.

Awards & Competitions

ECCV 2026 Test of Time Award (announcement): for SSD: Single Shot MultiBox Detector
CVPR 2025 Test of Time Award (Longuet-Higgins Prize): for Going Deeper with Convolutions
ICLR 2024 Test of Time Runner-Up (announcement): for Intriguing Properties of Neural Networks
MS COCO Captioning Challenge 2015: Tied for 1st place with Show and Tell (results)
ImageNet (ILSVRC) 2014: Won 1st place in both classification and object detection with GoogLeNet (results)

Products & Press

Gemini Omni (2026): Gemini app, Flow, YouTube Shorts · Google Launch · TechCrunch
Veo 3 (2025): Gemini app, Flow, YouTube Shorts (230M+ videos generated) · Google Launch · TechCrunch · The Verge
Phenaki (2022): foundational video model powering Google’s generative video roadmap · Google Research · Ars Technica
Google Photos (2015): visual search in Photos (1.5B+ monthly users) · Google Announcement · The Verge
Image Captioning (2014): automated descriptions in Google Photos & Cloud Vision · Google Research · The New York Times
GoogLeNet / Inception (2014): visual search in Google Images, Photos & YouTube · Google Research · MIT Technology Review