Pierre Sermanet
Pierre Sermanet's posts
Post has shared content
TensorFlow in the Cloud!
Very excited to be launching our new Cloud ML service. Run arbitrary TensorFlow graphs on GCE, tightly integrated with the rest of Google's Cloud APIs.
Bonus: brand new Speech APIs that provide real-time recognition in the many languages we support!
Very excited to be launching our new Cloud ML service. Run arbitrary TensorFlow graphs on GCE, tightly integrated with the rest of Google's Cloud APIs.
Bonus: brand new Speech APIs that provide real-time recognition in the many languages we support!
Post has shared content
Excited to show you our toys ... and the friendly robots that help us pick them! This is a very exciting collaboration between my team and [X]. The field of computer vision has been making giant progress these past few years. As researchers, we're getting to the point where we have to look beyond pure perception, and what we can do with these newfound visual abilities. Robotics is a natural next step in that journey, linking perception and action to perform end-to-end tasks seamlessly in a robust and scalable manner. Check out the videos!
Post has shared content
More robot mistreating! :(
New Boston Dynamics video:
Post has shared content
Now you can detect objects in your images the same way +Google Photos does. Learn how to get started with Cloud Vision API: http://goo.gl/iePJsA
Post has shared content
Stunning auto-generated images.
After the release of DCGAN (and code), which is a neural network that can generate natural images (collaboration between Alec Radford, Luke Metz and me), people from the community have generated their own manga characters, flowers and now their own letters from reading chinese books. Fun stuff. All these generations are the hallucinations of a neural network and are not real.
Our paper: http://arxiv.org/abs/1511.06434
Our code:
https://github.com/Newmu/dcgan_code (Theano)
https://github.com/soumith/dcgan.torch (Torch)
https://github.com/mattya/chainer-DCGAN (manga)
http://genekogan.com/works/a-book-from-the-sky.html (chinese characters)
https://twitter.com/vintermann/status/675599478494208000 (flowers)
Looking for more fun stuff from the community.
Our paper: http://arxiv.org/abs/1511.06434
Our code:
https://github.com/Newmu/dcgan_code (Theano)
https://github.com/soumith/dcgan.torch (Torch)
https://github.com/mattya/chainer-DCGAN (manga)
http://genekogan.com/works/a-book-from-the-sky.html (chinese characters)
https://twitter.com/vintermann/status/675599478494208000 (flowers)
Looking for more fun stuff from the community.
‹
›
12/15/15
6 Photos - View album
Post has shared content
Congratulations to Microsoft Research Asia for winning this year's ImageNet challenge!
Crossing 60% mean average precision on detection is a big result. In part, it suggests that we might soon be at the performance level users would actually consider ok if we showed detection results on a user-facing UI. This kind of milestone would dramatically change what we can do with computer vision technologies today.
Our team only entered in the classification competition and got 3.581% error, a tad behind the best 3.567%. In case you're wondering about our abysmal 19% localization error figure: we didn't even try ;-) This year's rules said we had to enter bounding box coordinates, so we just threw in the best proposal from a class-agnostic localizer.
Crossing 60% mean average precision on detection is a big result. In part, it suggests that we might soon be at the performance level users would actually consider ok if we showed detection results on a user-facing UI. This kind of milestone would dramatically change what we can do with computer vision technologies today.
Our team only entered in the classification competition and got 3.581% error, a tad behind the best 3.567%. In case you're wondering about our abysmal 19% localization error figure: we didn't even try ;-) This year's rules said we had to enter bounding box coordinates, so we just threw in the best proposal from a class-agnostic localizer.
Post has shared content
We've just posted our latest pretrained ImageNet classifier on the TensorFlow web site. Enjoy!
Post has shared content
http://arxiv.org/abs/1512.00567
~3.5% top-5 (17.2% top-1) classification error on ImageNet (ILSVRC2012 validation) with the latest Inception model.
~3.5% top-5 (17.2% top-1) classification error on ImageNet (ILSVRC2012 validation) with the latest Inception model.
Post has shared content
We've just released our Cloud Vision API. You can now get rich image annotations, fast, and at scale!
Post has attachment
Are GoogLeNet's deep dog dreams early signs of brain disease?
"According to one small study published in 2011, for example, the early stages of Parkinson's disease are characterised by alterations in the content of dreams, particularly the presence of animals [...]"
http://www.theguardian.com/science/neurophilosophy/2013/may/22/dreaming-of-animals-and-other-warning-signs-of-neurodegeneration
Joke aside, in addition to regular and sufficient sleep, you should sleep on your side to maximize your brain's waste disposal (if you are a mouse, may also apply to humans) as it may reduce chances of brain diseases on the long term:
http://www.theguardian.com/science/neurophilosophy/2015/aug/22/how-to-optimise-your-brains-waste-disposal-system
"According to one small study published in 2011, for example, the early stages of Parkinson's disease are characterised by alterations in the content of dreams, particularly the presence of animals [...]"
http://www.theguardian.com/science/neurophilosophy/2013/may/22/dreaming-of-animals-and-other-warning-signs-of-neurodegeneration
Joke aside, in addition to regular and sufficient sleep, you should sleep on your side to maximize your brain's waste disposal (if you are a mouse, may also apply to humans) as it may reduce chances of brain diseases on the long term:
http://www.theguardian.com/science/neurophilosophy/2015/aug/22/how-to-optimise-your-brains-waste-disposal-system
Wait while more posts are being loaded