Big News of the week: Google has trained a Neural Network with a billion internal nodes, using aabout 16000 computers......well, to detect cats on internet!
At least that's how media touts it. NYTimes published an article on Google's recent feat here
. The article says that Google's training set consists of millions of images from youtube videos, their neural network "taught itself" to identify objects in the images and "roughly doubled" the accuracy of detection.
A widespread publicity of academic paper in a non-academic paper (aka NEWS) is always exciting. In particular, this paper is simply amazing given its results and scale of operation and richly deserves the media coverage it has received. I tried to look at the ICML publication, now available on Google Research page
. One of the most remarkable aspects of the research is the fact that it uses unsupervised learning algorithm
with an unlabeled data
of 10 million 200X200 pixel images. This means that although most of the object-detection algorithms take labeled data, (i.e.images with objects outlined and identified by someone ---either users or mechanical turks or researchers themselves), this research has no such outline and identification to start with. Naturally, the algorithm has to be complex. Just how complex is it? --- It has a 9-layer Neural Network with a billion connections and took 3 days to find the parameters of the model.
What about the results? How accurate is the algorithm? First and foremost, as I mentioned, the algorithm is able to detect faces without any help --- with no labels
and it has increased accuracy of any existing algorithm by a staggering 70%. It can now recognize objects in 20000 images with the accuracy of, well, 15.8%. Now 15.8 is a fairly small number. But that shows us how far we really are from developing truly intelligent systems. Yet, this paper should prove to be a milestone in achieving that lofty goal. Hopefully! In the meanwhile, here's to the amazing team (consisting of some of the great names in research and industry: Jeff Dean and Andrew Ng) - amazing work!
P.S: I have just finished listening to all video lectures by Prof. Ng on Coursera. Highly recommended for those who are excited by this stuff!