4 ms·
ImageNet Large Scale Visual Recognition Competition
- jbarrow 12y agoI love looking at the results and seeing the annual improvement and breakout strategies. Deep learning has really come into its own during the last eight years. People interested in learning about some of the techniques employed (CNNs and deep networks) are encouraged to take Geoff Hinton's Coursers course [1], which does a really good job of covering modern techniques from one of their pioneers. [1] https://www.coursera.org/course/neuralnets https://www.coursera.org/course/neuralnets
- rjtavares 12y agoHow much are results improving? What's the general long term trend?
- dave_sullivan 12y agoThe trend is depreciating returns with current architectures (after truly breakthrough results in the first place). It's a good time for companies to update their ML pipelines to use deep learning. Performance does seem to scale with hardware though, so as hardware improves (as well as the software), these models will see some more gains without any real technology change. But I should also point out deep learning doesn't work for everything. Random forests or SVMs work about as well on certain (I would argue "more linear") datasets. Deep learning is for when RFs and SVMs aren't giving the results you need (or at all). It doesn't negate all things that came before, nor should it. One thing to consider: As far as I can tell, there's an order of magnitude more people working on these types of problems with an order of magnitude more money being thrown at it and an order of magnitude increase in "collaborative efficiency" due to Internet and open source/research. I think breakthroughs are thought of best as a probability distribution, but I think the probability of further breakthrough successes has increased considerably in just the last few years. There are reasons for optimism and pessimism, as usual.
- m_ke 12y ago15% in 2012, 11% last year and 6% this year. We might also see late submissions beat the this years top result in the next few months.
- theschwa 12y agoIs there any place to view Hinton's lectures? I've been waiting for that Coursera course to open again since I barely missed out on the first one, but I haven't seen it run since.
- Houshalter 12y agoAll of the lectures are there, you just need to join the course to see them.
- cr4zy 12y agoLooks like BVLC's Caffe and R-CNN are getting a lot of mindshare. Also interesting to see that the creator of Caffe, Yangqing Jia, was on team GoogLeNet and didn't appear to have used either. Edit: Last years results with no Caffe (just Decaf, its predecessor) http://image-net.org/challenges/LSVRC/2013/results http://image-net.org/challenges/LSVRC/2013/results
- dave_sullivan 12y agoCuda convnet 2 was recently released, it's faster than caffe it seems. Caffe is very cool though, I've been playing with it a bit. https://code.google.com/p/cuda-convnet2/ https://code.google.com/p/cuda-convnet2/
- cr4zy 12y agoHey Dave! (Craig here) nice, achieved 6x speed up on an 8 gpu box (http://arxiv.org/pdf/1404.5997v2.pdf http://arxiv.org/pdf/1404.5997v2.pdf). Thanks for the pointer and hope to see you soon.
- dave_sullivan 12y agoHey Craig, nice to hear from you, I hope things are going well. That is a nice bump! The data and model parallelism they added is very cool. We'll be hosting a meetup soon, it's been a while, hope to see you there.
- jiayq 12y agoHi cr4zy! Thanks for mentioning me in the post. At google we have our own large-scale deep learning framework (the Google Brain) that is well constructed and fits the google infrastructure well, so it is understandable that we did not use Caffe at google for the challenge. Mathematically all libraries are alike though, so one is more or less free to choose whatever tool that comes most handy. (hint: you probably want to try out Caffe, or join Google :)) On the detection front we actually adopted the idea of R-CNN quite a lot, so I guess in the context of research, we are standing on the shoulder of existing great research results! Please stay tuned for our upcoming blog post about GoogLeNet and a more detailed introduction of the method at ECCV.
- contingencies 12y agoScience is but a perversion of itself unless it has as its ultimate goal the betterment of humanity. - Nikola Tesla Does this not nearly amount to "population-scale mass surveillance algorithms"? Do people not feel this is accelerating negative social impacts of technology? Is it merely a coincidence that winning teams include many from countries criticized for their totalitarian social contracts: Hong Kong University of Science and Technology, National University of Singapore, Microsoft Research China, Southeast University (China), Chinese Academy of Sciences? There's also a presence from Holland. Oh, and guess who won the category "with additional training data"? Google. Come on people, we can do better than this! SHAME SHAME SHAME.
- Houshalter 12y agoMachine vision has numerous applications, better surveillance being only one of them. Mainly self driving cars, and significantly better AI for robotics and some other automated systems. The organizations you mention are just places that invest in or do a lot of research in AI.
- Houshalter 12y agoThe winner got 6.7% in top-5 classification error. 2013 was 11%, 2012 was 15%. This may be better than exponential, since there is an upper limit of 0% error, and every additional percentage point is harder than the last. Exponential improvement in this area, and it's a pretty significant field that will affect everyone. The results will likely help improve neural networks in other areas as well, as NNs are extremely general purpose. Here is an example from a system from 2012 in the same competition to give an example of what this is: http://mappingignorance.org/fx/media/2013/04/Deep-learning-5.png http://mappingignorance.org/fx/media/2013/04/Deep-learning-5... A bunch of images were scraped from Flickr and then given a single label by a human. The computer gets 5 guesses to guess what label the human used. As you can see, even when the computer fails, it's guesses are usually semantically very good. This specific system got an error rate of 17%. You can explore the images of ImageNet here: http://image-net.org/explore http://image-net.org/explore. Here are some image recognition systems online you can play with: * Toronto Deep Learning Demo: http://deeplearning.cs.toronto.edu/ http://deeplearning.cs.toronto.edu/ * clarifai: http://www.clarifai.com/ http://www.clarifai.com/ * AlchemyVision: http://www.alchemyapi.com/products/demo/alchemyvision/ http://www.alchemyapi.com/products/demo/alchemyvision/ Note these aren't related to the winner of this competition and probably are no where near as good. Just examples of this kind of system you can play with.