News from the world of maths: Machine learning and clever 8-year olds

Monday, March 31, 2008

Machine learning and clever 8-year olds

This week, Plus is attending the Edinburgh International Science Festival — an educational charity aiming to engage society in the value of science and technology, placing a particularly strong emphasis on providing experiences of science that are inspiring and confidence building. The Festival is internationally well-regarded and one of the biggest science events in the UK. Its timing suggests we are now well into science festival season, with roughly another 10 city-wide events to be held around the country, in addition to the recently run Cambridge and Brighton Science Festivals. In 2007, Plus attended the BA York Science Festival — you can read more on the Plus Blog.

There is a comprehensive range of talks from some of the world's leading scientists, and last evening Plus attended a talk by Chris Bishop, Deputy Director of Microsoft Research in Cambridge, on Machine Learning. Compared to humans, machines are very bad at recognising some patterns — for instance, recognising that a photo of a cow is actually a cow and not a dog or a shadow or even a spoon — they simply do not innately know these things and must be taught. Whilst humans are very good at understanding messy handwriting — a subtle task — computers must learn. Bishop's work involves developing computers that can learn and so become smarter in completing their tasks. To read more about the mathematics involved, see the Cambridge team's introduction to applied mathematics for machine learning.

Central to machine learning is the Turing Test. English mathematician Alan Turing — instrumental in breaking the Enigma code during World War II — is widely regarded as one of the founders of modern computer science — see the Plus article What computers can't do. In a 1950 paper, he described what he called the Imitation game, now known as the Turing Test, in which a person (in a separate room) tries to distinguish between human and computer test subjects by asking them each a series of questions. If the person can't distinguish between the computer and the human, the computer is deemed to be intelligent. There is a (currently unclaimed) prize of $100,000 for a computer that can pass the Turing Test. Bishop talked about the Forza Motorsport X-box game, in which you race against other players or against the computer. The computer car can be taught to drive like a human — you can create a "Drivatar" which will learn from the way you drive and then drive like you. The word Drivatar comes from avatar, which in gamer-speak is the player's representation within the game. Participants in trials have said that they cannot tell the difference between a human playing and a computer, and so within a very limited context, the Drivatar passes the Turing Test.

Bishop also talked about other developments in machine learning, from Spam email detection — see the Plus article CAPTCHA if they can — to websearches and modern developments in HIV detection. New software is being developed to look for patterns in the HIV protein code. This code evolves over time, and so searching for the same string of amino acids each time is likely not to work. He also discussed developments in the way computers play the games Chess and Go. He stated that whilst the IBM Deep Blue chess-playing computer was able to beat world chess champion Garry Kasparov in 1997, it was not because of any particular advances in machine "intelligence" — the success was largely due to advancements in computer memory and speed that allowed the computer to analyse the game very thoroughly. If at each turn you can make around 30 possible moves, and then there are around 30 possible replies by your opponent, and then in turn another 30 possible moves you can make, and so on, then a human could never hope to analyse each and every scenario. However, Deep Blue was able to look at the full set of possible moves and then make the best one (although it did have the intelligence to look within a set of smart moves). With the game Go however, at each turn there are around 200 possible moves, meaning that it is impossible currently to completely analyse each scenario before playing a move. Bishop is working with computers that are learning by playing against humans.

I have also attended Wonderama and the mathematical puzzles at the Museum of Scotland. Wonderama is designed for the whole family and hosts many free activities and events for budding scientists, aspiring surgeons, wannabe adventurers and trainee archaeologists. Thankfully, my time around science festivals means that I wasn't outdone by 8-year olds in the mathematical puzzle department. Actually, that is not completely true...

Tonight, I am eagerly awaiting the Edinburgh Medal Address by Professor Chris Rapley, Director of the Science Museum, on the dangers of climate change. The Edinburgh Medal is awarded each year to those within science whose professional achievements are judged to have made a significant contribution to the understanding and well-being of humanity.

For more information and a comprehensive program, see the Festival website. The event runs until April 6.

posted by westius @ 1:57 PM