Historically, very very interesting discussion in the comments (Oct 15, 2012)

Historically, very very interesting discussion in the comments (Oct 15, 2012)
"""
Geoffrey Hinton: I predicted that some vision people would say that the task was too easy if a neural net was successful. Luckily I know Jitendra so I asked him in advance whether this task would really count as doing proper object recognition and he said it would, though he also said it would be good to do localization too. To his credit, Andrew Zisserman says our result is impressive.

I think its pretty amazing to claim that a vision task is "just too easy" when we succeed even though some really good vision people tried hard at it and failed to do nearly as well. I also think that trying to discredit a system that gets about 84% correct by saying you could get 0.5% correct by chance is a bit desperate.
"""
Yann LeCun: The vision community as a whole has been skeptical about deep learning and feature learning, but there was the same kind of skepticism from the ML community until 4 or 5 years ago, and from the speech recognition community until 2-3 years ago. Thankfully, things are changing quickly now: deep learning is the hottest topic in speech recognition these days.
[...]
Thankfully, results on standard benchmarks have a way of quieting down theological arguments. Regretfully, they also have a way of killing new ideas in the womb, before they had a chance to grow and prove their worth. That's why brand new techniques in a particular community often come from other communities, where they have had a chance to grow before before confronting the real world. That's why much of the deep learning work was initially published at ICML and NIPS before making it to vision conferences. That's why it's been a struggle getting deep learning papers accepted at CVPR, ECCV and ICCV until now (it was a struggle to get them accepted at NIPS and ICML for a while too).
"""
Alex Krizhevsky: To me, our result demonstrates that a model which makes few prior assumptions about the visual world can do well when given lots of training data and lots of computing power (relatively speaking).

(Of course, to +Ilya Sutskever, this was obvious from the start and needed no demonstration.)
"""
Yann LeCun: The difference between pooling layers in convolutional nets (particularly max-pooling at the second stage) and deformable part models is actually very small when you think about it.
"""
https://plus.google.com/+YannLeCunPhD/posts/JBBFfv2XgWM

AlexNet

Originally shared by Yann LeCun

Alex Krizhevsky's talk at the ImageNet ECCV workshop yesterday made a bit of a splash. The room was overflowing with people standing and sitting on the floor. There was a lively series of comments afterwards, with Alyosha Efros, Jitendra Malik, and I doing much of the talking.

The three of us plus David Forsyth and Vitto Ferrari will be on what promises to be a lively panel discussion at the end of the afternoon today in the ECCV Workshop on "Parts and Attributes".

I'm also giving a talk at 2:30 in the ECCV workshop on "Higher-Order Models and Global Constraints in Computer Vision".

Labels: Oleg Zabluda

- posted by Oleg Zabluda @ 5:08 PM | permalink | | | technorati | bloglines | feedster