Photo of Dumitru Erhan discovering AI


I’m Dumitru Erhan, I live in San Francisco, California and I work at Google. Some things I enjoy are cycling, cooking, kitchen gadgets, spending time with our two cats, raising our son, and of course neural nets.


At Google, I am working on algorithms that would detect and recognize what is happening in an image or video. I'm in the Google Brain team. Current focus areas are scalable high quality detection, fast single-shot detection, image captioning, visual question-answering, as well as structured outputs for image understanding. Formerly, I was part of Google Photos, and built neural network models that understand all the images and photos on the web. Here's an overview from 2013 and 2015.



My up-to-date list of publications can be found on my Google Scholar or ResearchGate, which do a better job than me at indexing them.

News (ish)

  • 29 Dec 2015  —  A new paper on SSD: Single Shot MultiBox Detector, showing that it is possible to get high-quality detections for 200-classes in a single conv-net pass through an image, now on arXiv.
  • 29 Dec 2015  —  We refreshed our paper on Scalable High-Quality Object Detection and it's on arXiv.
  • 26 Feb 2015  —  We contaminated one of our models with the half of the validation set. We still report state of the art on the ImageNet detection challenge, but it's not as optimistic as we previously claimed. Up-to-date numbers and paper on arXiv.
  • 17 Nov 2014  —  Google Research Blog post on our Show and Tell: A Neural Image Caption Generator work. Coverage in NY Times, BBC.
  • 18 Aug 2014  —  The GoogLeNet team, of which I am proud to be part of, won a couple of challenges in the annual ImageNet competition. Official results and New York Times article.
  • 23 Jun 2014  —  Invited talk at the 2014 CVPR workshop on Biometrics in Columbus, OH. Slides available here.
  • 27 May 2014  —  Needlessly pessimistic coverage of our Intriguing properties of neural networks paper by slashdot, ZDNet.
  • 21 Jun 2013  —  Ian Goodfellow, myself and Yoshua Bengio organized the ICML 2013 Challenges in Representation Learning workshop in Atlanta. In addition to talks and papers (see workshop page) we organized 3 Kaggle challenges, in which hundreds of teams participated.


  • Download my (outdated) résumé
  • Email:
  • My wife, a post-doctoral fellow at the Diabetes Center of UCSF.


I'm on Facebook, Google+, LinkedIn, and Twitter.