Geoffrey Hinton: "Does the Brain do Inverse Graphics?"
HTML-код
- Опубликовано: 30 сен 2024
- Graduate Summer School 2012: Deep Learning, Feature Learning
"Does the Brain do Inverse Graphics?"
Geoffrey Hinton, University of Toronto
Institute for Pure and Applied Mathematics, UCLA
July 12, 2012
For more information: www.ipam.ucla....
This is simply amazing. You can see how the future is being created in his mind.
Unfortunately, it is being created only in his mind... Seems to have little or nothing to do with the nature of reality.
what's the meaning of 'compute the handedness of a coordinate transformation'?
"Is the image horizontally mirrored?"
16:23 As a person who use 3d modelling tools a lot, at the first 100ms I immidiately got irritated that the left R is both symmetric and rotated without thinking about it :D (there are very few that kind of motives/patterns that requires both at the same time)
So, we're onto learning some form of affine transformation invariance, rather than just translation invariance?
(I'm not sure if this is 100% possible, though)
jhui.github.io/2017/11/03/Dynamic-Routing-Between-Capsules/
blog.acolyer.org/2017/11/14/matrix-capsules-with-em-routing/
Going to pore over the papers referenced in the URLs, here:
Dynamic Routing Between Capsules: arxiv.org/abs/1710.09829
(Under review) : Matrix Capsules with EM Routing: openreview.net/pdf?id=HJWLfGWRb
Would a capsule network will be more resistant (compared with current CNNs) to adversarial attacks that add noise to an image that, for example, make a panda look like a gibbon?
This will make computer vision so much better if it can be scaled to be used real-time for real objects. Being able to a have recognition of objects from any angle, any distance, and with massive distortions would be great.
Do google "You Look Only Once", "Single Shot Detectors" and "Faster RCNN's". They come pretty close(they basically do the job well) , but still have some pitfalls
this is just awesome ... but could anyone tell me about how this thing will revolutionize the ml as we know it ? in which tasks ?
There is one very important property of CapsNet that nobody talk. It is vectorize output of recognized pattern. Vectors can be put into the same address bus to search for close objects like in Word2Vec algorithm. Or Couple Vectors can be added to a new one creating new Abstract object! I think that would a realrevolution since it allows to create something that does not exist by combining existing ideas. Like human creativity.
Indeed ! But I doubt that with vanilla CNN we will not able to derive some kind of vectorized output as well ?