Profile cover photo
Profile photo
Dumitru Erhan
715 followers -
Neural network aficionado.
Neural network aficionado.

715 followers
About
Posts

Post has shared content
Very proud to be open-sourcing TensorFlow, Google's newest Deep Learning framework! TensorFlow is both a production-grade C++ backend, which runs on Intel CPUs, NVidia GPUs, Android, iOS and OSX, and a very simple and research-friendly Python front-end that interfaces with Numpy, iPython Notebooks, and all the familiar Python-based scientific tooling that we love. TensorFlow is what we use every day in the Google Brain team, and while it's still very early days and there are a ton of rough edges to be ironed out, I'm excited about the opportunity to build a community of researchers, developers and infrastructure providers around it. Try it out!
Add a comment...

Post has shared content
Today we've announced the Google Brain Residency Program, a one year program for people with computer science background who are interested in Deep Learning to join us at Google, study machine learning and conduct research with our team.
Details, timeline and application process: http://g.co/brainresidency
Add a comment...

Post has shared content
Google regains the Imagenet (ILSVRC 2012) classification mantle, but just barely: 4.82% top-5 classification error vs 4.94% for Microsoft. Best human accuracy on this task is about 5% error. 

What's really cool is the batch normalization idea in the paper. It is not only a huge speed improvement, but also a notable quality improvement. This should become the new standard way of training deep nets. Worth noting is that the Inception net we trained is much cheaper than the recent Microsoft or Baidu counterparts. 

Congratulations +Sergey Ioffe  and +Christian Szegedy on this awesome work. 
Add a comment...

Post has shared content
Progress in object detection, with pretty graph on the Google Research blog. If you want to chat about this, hit me up at NIPS this week.
Increased accuracy and reduced computation time in Object Detection

This summer, the GoogLeNet team reported top results (http://goo.gl/Y7qLLa) in the 2014 Large Scale Visual Recognition Challenge (ILSVRC), with ~2X improvement over the previous year’s best results. However, the quality of these results came at a high computational cost: processing each image took about two minutes on a state-of-the-art workstation.

In the recent paper, Scalable, High Quality Object Detection (http://goo.gl/Oy1Vr4), Google Engineers Christian Szegedy, +Dumitru Erhan, Dragomir Anguelov, and Software Engineering Intern/University of Michigan PhD student Scott Reed detail advances in object detection that have resulted in a 23% relative gain in mean average precision and 140X reduction in computational resources over the previous GoogLeNet results. Learn how in the Google Research blog, linked below.
Add a comment...

We have some new results improving the state of the art on the ImageNet detection challenge by quite a large margin: from ~0.45 mAP to 0.557 mAP. In a nutshell, we improve our multibox bounding box regression method, which by itself obtains state of the art results, while being faster. Combined with selective search proposals, we get even better. More details in our arXiv submission: http://arxiv.org/abs/1412.1441
Add a comment...

Our team at Google is looking for summer interns. We have a number of exciting projects, mostly in computer vision and related areas: be it classification/detection (http://googleresearch.blogspot.com/2014/09/building-deeper-understanding-of-images.html), image description (http://googleresearch.blogspot.com/2014/11/a-picture-is-worth-thousand-coherent.html) or yet-to-be-announced projects, and we aim to push the state of the art at everything that we do. Fundamentally, we are also an applied team so we end up launching models that are used by hundreds of millions of users (image search, YouTube etc).

Ideal candidates will have strong coding (C++/Python) skills and good machine learning understanding. Prior computer vision experience is not required. We aim to publish as much as possible and as fast as possible so a publication is always a possibility of a successful internship.

(+Dragomir Anguelov, +Andrew Rabinovich, +Christian Szegedy, +Samy Bengio, +Oriol Vinyals, +Alexander Toshev, +Xiangxin Zhu 
Add a comment...

One of the launches to which our team contributed. Fun stuff, check it out! http://www.wired.com/2014/05/google-photo-stories/
Add a comment...

Post has attachment

Post has shared content
Big names and fun topics, will try to show up as well :)
Looking forward to seeing how people improve Gradient Descent for neural networks. IPAM's 5-day Stochastic Gradient Methods workshop is in 2 weeks.

https://www.ipam.ucla.edu/programs/sgm2014/
Add a comment...

Post has shared content
Lately the ImageNet competitions have become quite fun, especially since both the number but also the quality of the submissions has risen considerably! We may even participate too :)
Check out the new and improved ImageNet challenge this year! The object detection dataset will be significantly expanded, plus there are a few more changes: http://image-net.org/challenges/LSVRC/2014/ +Jonathan Krause +Jia Deng +Alex Berg +Fei-Fei Li 
Add a comment...
Wait while more posts are being loaded