Saturday, June 30, 2012

Google AI loves cats

Last week the media was full of stories about a new Google X Lab project that has created an AI which seems to love cats. I couldn't avoid the story because friends kept emailing me links to it - this story in the Financial Times is fairly typical
    Google built a huge neural network using 16,000 computer processors to see what it would learn when exposed to 10 million clips randomly selected from YouTube videos. There are basically two types of machine learning; supervised when you say "here's an example of X," "here's an example of Y" and "here's another example of X," and unsupervised learning where there is no instructor. Google's system was unsupervised, it just looked at all the YouTube clips and tried to find interesting patterns. It did - cats! Google's system can now look at a YouTube clip and tell you, with some certainty, if there is a cat in it or not.
   Before we leap to conclusions that AIs like cats or want pets first consider an old experiment conducted with neural nets for the Pentagon. They wanted to find Russian tanks in spy photos; so using supervised learning they showed a neural net hundreds of photos, some of which had tanks in and some which did not. After training the neural net could, again with some confidence, identify tanks in photos it had never previously seen. Success they thought. Later they discovered that most of the photos they had of tanks were taken on cloudy days, whereas most of the photos of countryside without tanks were sunny days. The computer had learnt to see if it was cloudy or sunny - the tanks were a coincidence. Google's network may recognize cats but not as we do.
   Actually the cats were a by-product, Google system can recognise 20,000 different things in the YouTube clips.