Dumitru Erhan

Dumitru Erhan researches visual understanding and generative media. At AI Engineer World’s Fair 2026, he was introduced as a senior research director at Google DeepMind and co-lead of Gemini Omni. His collaborative research spans how deep networks learn representations, how machines recognize and describe images, and how video models represent multiple possible futures.

Early research and why pre-training works

Erhan grew up in Chișinău, Moldova, and earned a bachelor’s degree in electrical engineering and computer science at International University Bremen in 2004. His master’s research applied collaborative filtering and multi-task learning to drug discovery in collaboration with AstraZeneca. He subsequently completed doctoral research in deep learning at the University of Montréal with Yoshua Bengio.

A central question in that work was why unsupervised pre-training helped deep networks learn. In research published with his collaborators in 2010, the team examined how an initial learning phase without task labels affected later supervised training. Their experiments supported a regularization explanation: pre-training steered optimization toward solutions that generalized better beyond the training examples. They also investigated how depth, model capacity, and the amount of available data changed its effects.

From search to visual understanding

At Yahoo Labs in 2011–2012, Erhan worked on federated search, improving the ranking of different kinds of results within Yahoo’s search page. At Google, he worked on visual understanding, including neural models for Google Photos, object detection, image captioning, and visual question answering. He later pursued video generation and prediction within Google Brain, connecting video modeling to model-based reinforcement learning.

• Recognition efficiency and reliability: His contributions to image recognition addressed both computational efficiency and the limits of apparent accuracy. Erhan co-authored the Inception research behind GoogLeNet, which advanced classification and detection in the 2014 ImageNet challenge. The architecture increased network depth and width through processing at multiple scales while holding the computational budget constant. His co-authored research on adversarial examples exposed a different problem: carefully constructed image changes too small for a person to notice could cause misclassification, and some perturbations transferred between networks trained on different data subsets. Successful recognition on ordinary examples did not guarantee dependable behavior under small, deliberate changes.
• Image captioning and object detection: Erhan also co-authored Show and Tell, which connected computer vision with language generation. Instead of returning an object label, its recurrent generative model learned to produce a sentence describing an image. The Single Shot MultiBox Detector, or SSD, tackled object localization and classification within one network. It scored default boxes at different scales and aspect ratios, then adjusted them to fit objects. Predictions from feature maps at multiple resolutions helped detect objects of different sizes; removing a separate proposal-generation stage made detection faster and easier to integrate into other systems.

Predicting possible video futures

• Representing uncertainty: Video prediction introduced another challenge: an observed scene can have several plausible continuations. Erhan’s co-authored SV2P research used sampled latent variables to represent that uncertainty, allowing stochastic video prediction to produce different possible futures rather than a single deterministic outcome. The team evaluated prediction both from observations alone and conditioned on actions, making the work relevant to systems that must anticipate the consequences of decisions.
• Predicting higher-level representations: In hierarchical video-prediction research, Erhan and his collaborators jointly trained an encoder, a predictor, and a decoder. The predictor advanced a higher-level representation through time, and the decoder used that representation alongside the first frame to construct future images. This approach reduced dependence on manually annotated structures such as human joint positions.
• Efficiency and generalization: His co-authored FitVid research examined architectural efficiency and overfitting in video prediction. Convincing outputs could come from repeating training examples, so visual quality alone could not establish generalization. The team investigated image augmentation as a way to reduce overfitting while improving prediction.

Generative media and the details that make it usable

Google DeepMind credits Erhan among the research and engineering contributors to Veo 3, its joint audiovisual generation model. Gemini Omni, introduced in May 2026, accepts combinations of text, images, video, and audio references to generate video and support conversational editing. These capabilities extend visual modeling into a task that requires both change and continuity: carrying out an edit while preserving the parts of a video the user wants to keep.

At the World’s Fair 2026 generative media panel, Erhan joined Shane Gu and Nicole Brichtova in discussing the gap between attractive media and useful results. The panel examined how human preferences can favor sharp, saturated outputs without establishing realism, and how expert feedback and customer workflows reveal failures missed by general evaluations. Its examples included reproducing a pattern across different rug sizes, preserving the physical scale of earrings in a virtual try-on, and matching a brand’s particular colors. Those problems make the practical stakes of generative media concrete: an appealing image or video must also preserve the details that make it usable.