The immense processing power of Google’s global computing network and the brainpower of its secretive Google X research labs remain largely hidden from a curious world. But this week we got a glimpse of what the company’s great minds, human and electronic, are thinking about – cats.
Google scientists presented an impressive paper at the international machine learning conference in Edinburgh, which demonstrated the company’s ambitions in artificial intelligence as well as the strength of its computing resources. They built an electronic simulation of the brain, ran it on 16,000 computer processors in Google data centres and discovered what it would learn when exposed to 10m clips randomly selected from YouTube videos.
Unprompted, the computer brain learnt to identify the feline face. That might seem a trivial accomplishment, demonstrating little more than the obsession of cat owners with posting videos of their pets. But in fact Google has made a significant advance in artificial intelligence, a research field that has long promised much but delivered little to computer users.
Standard machine learning and image recognition techniques depend on “training” the computer first with thousands of labelled pictures, so it starts off with an electronic idea of what, say, a cat’s face looks like. Labelling, however, requires a lot of human labour and, as the Google researchers say, “there is comparatively little labelled data out there.”
Google needs to master what it calls “self-taught learning” or “deep learning”, if it is to extend its search capabilities to recognise images among the vast volume of unstructured and unlabelled data. That would enable someone, for example, who owned an unidentified portrait painting by an unknown artist to submit a photograph of it to a future Google – and stand a reasonable chance of having both the scene and the painter identified through comparison with billions of landscape and art images across the internet.
The study presented this week is a step towards developing such technology. The researchers used Google data centres to set up an artificial neural network with 1bn connections (most networks used in machine learning research have just 1m to 10m connections) and then exposed this “newborn brain” to YouTube clips for a week, without labelling or identifying data of any sort.
The network taught itself to identify human faces and bodies and, to the scientists’ surprise, cats. The results may be significant for fundamental brain research, because recognition ability was not distributed evenly across the network. Individual artificial neurons developed specific skills, such as spotting cats. This supports the controversial neuroscience theory that the brain contains “grandmother cells” that learn to recognise particular categories of people or even individuals such as a grandparent.
But Google will be more interested in the practical applications for its corporate ambitions in artificial intelligence. Andrew Ng and Jeff Dean, the leaders of the neural study, say the project is not just about images: “We’re actively working with other groups within Google to applying this artificial neural network approach to other areas such as speech recognition and natural language modelling.”
Over the past few years the company’s founders, Sergey Brin and Larry Page, have indicated their fascination with AI and their ultimate aim to give Google Search an artificial intelligence exceeding that of any human brain. Then you could ask any question in any language, and Google would search all the words, sights and sounds of the internet to come up with as much or as little relevant information as you requested.
Anyone who has used Google’s automatic translation service to turn a foreign language web page into English will know how much help it still needs. To test it for this article, I requested an English translation of the French Wikipedia entry for chat (cat).
Here is a verbatim sequence from what I got back: “It also means more familiarly by the cat pussy and pussy by pussy. This term, dating back from 1560, comes from mine, popular name of the cat in Gallo-Romance. This word is the origin of the expression ‘at the crack of dawn’, which means ‘good morning’.”
Admittedly Google can do much better with its more specialised machine translation services, such as one it launched this year with the European Patent Office. But even the most advanced technology for understanding, generating or translating human language by computer lags behind far behind the predictions of AI pioneers 30 or 40 years ago of where we would be in the early 21st century.
The world of computing has been celebrating this week’s the centenary of the birth of Alan Turing, who set the best-known test for AI back in 1950: if an intelligent person cannot tell whether he or she is conversing with another human or with a computer, then the computer has achieved artificial intelligence. There is little sign of any machine passing the Turing test in the near future.
Yet Google has grand ambitions for accelerating the arrival of AI. Although the company reveals very little about its research plans, Ng and Dean say they are working to scale up substantially their neural artificial neural network, which may already be the largest in the world.
There is a long way still to go before any artificial network comes close to the human brain with its 80bn neurons, 100tn connections and a biological way of processing information utterly different to computer programs running on silicon-based hardware. Even a cat’s brain is far beyond our capabilities.
Several other industrial and academic projects aim to simulate human thinking by computer. But Google, with its feline thinking, is well placed in the race for the cream of artificial intelligence.
Get alerts on Alphabet Inc when a new story is published