ImageNet Large Scale Visual Recognition Competition
image-net.org
image-net.org
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.
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.
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... 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.
Here are some image recognition systems online you can play with:
* Toronto Deep Learning Demo: http://deeplearning.cs.toronto.edu/ * clarifai: http://www.clarifai.com/ * 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.
Edit: Last years results with no Caffe (just Decaf, its predecessor) http://image-net.org/challenges/LSVRC/2013/results
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.
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.
The organizations you mention are just places that invest in or do a lot of research in AI.