Last week the top researchers in Artificial Intelligence (AI) gathered in Edinburgh to analyse the state of their subject. The topics under discussion ranged from robotic exoskeletons, to what tool-using crows can teach us about our own brains. Impressive results were reported in several fields, with previously intractable problems dropping like flies. Yet true machine intelligence seems as much of a dream as ever.
Over a thousand scientists came from around the world to attend the prestigious week-long International Joint Conference in AI (IJCAI). They were an eclectic mix of computer scientists, mathematicians, and psychologists, plus a few philosophers. Also in attendance were a pair of Sony's latest robots, the amazingly cute QRIOs.
The mood was optimistic, but AI has an unfortunate history of hope, hype, and dismal failure. Each new innovation promises to be the key breakthrough, then turns out to have a rather more limited scope. Underneath its turbulent public fortunes though, the subject has developed relatively smoothly. Important 'core' technologies have advanced dramatically, and continue to do so. Admittedly a lot of the progress is due to the beneficent effects of Moore's law (the rule-of-thumb that says computing power increases exponentially over time). There have also been great advances in the 'core' algorithms that underlie much of AI.
An example of this is the work of young Turkish scientist Zeynep Kiziltan, who received an award for her ground-breaking (if rather technical) work on what is called the Constraint Satisfaction Problem (CSP). This is an important field that should in turn lead to improvements in areas such as the automatic verification – and even design – of computer hardware and software.
The conference also honoured British scientist Geoff Hinton for his work on neural networks (programs that imitate collections of nerve cells). Neural networks were invented by Marvin Minsky, who – disappointed with logic-based techniques – looked to the brain for inspiration. Neural networks are good at pattern recognition (for example, they are used for hand-writing recognition).
Neural network programming has in turn influenced neuroscience. Computer neural networks are currently far simpler than the networks found in real brains. A typical neural network program might have a few thousand cells – whilst an ant's brain has a quarter of a million. Nevertheless, neuroscientists use them to model sections of the brain, trying to match computer simulations to experimental recordings.
Professor Hinton was involved in several key developments, including Boltzmann machines and the back-propagation algorithm. In recognition of this, he was presented with an award for Research Excellence.
The latest panacea in AI is to use statistical methods. With a solid basis in mathematics, they are largely driving out 'woollier' techniques such as fuzzy logic (where statements can be more-or-less true). They are applied - sometimes with more hope than insight - to everything from speech recognition to forensic science. The buzzwords of the day are Bayesian methods and Markov models. Given that the Reverend Bayes was an 18th century British church minister and Russian mathematician Andrey Markov died in the 1920s long before the computer, one might wonder what everyone's been doing all this time. In fact, many of the actual techniques depend on recent advances in statistics.
This work also capitalises on the increasing availability of raw data – the lifeblood of statistical methods. Perhaps the most important new area is biology. Modern biology generates vast amounts of data – particularly in genetics and bio-chemistry. This is has spawned the hot new topic of bio-informatics. It is an exciting cutting edge field; the potential is huge, and it's likely to be very profitable. It causes respectable academics to drool into their laptops, and funding bodies to go weak at the knees.
The Big Picture
The mainstream AI community is focused on specific technical problems and applications. It is an approach which has been very successful. By contrast, attempts to solve the 'big problems' of intelligence have typically sunk without a trace. However, it may now be time to return to the bigger picture. Aaron Sloman of Birmingham University is launching an ambitious new project. Called CoSy, it will use a substantial fraction of the EU's research budget (a cool €7m) to address the bigger questions of general reasoning and meaning. The result will be a series of new robots that try to tie together the different strands of AI into one coherent system.
Professor Sloman is realistic about the likely outcome: "We don't promise any results. We assume that [human-like thinking] is far beyond the current state of the art and will remain so for many years. But we are asking important questions."
Robots: cute but dumb
The climax of the conference was the appearance of Sony's adorable QRIO robots. Looking like child astronauts, they can balance and dance with motor control so smooth and fluid as to make grown men weep.
In other respects, QRIO is a little disappointing. His walking marks an improvement in robot technology, but it is still much more of a shuffle than a stride. There are people doing more impressive robot control. For example, Japan's Riken institute, who have taught their robot to juggle three balls. But when it comes to building lovable robots, no-on can touch Sony for cuteness (and no, they're not on sale yet).
QRIO is perhaps symptomatic of AI. He is undeniably impressive – yet unable to perform even the simplest human tasks. Machines still find day-to-day life challenging. This is partly because of life's unpredictability. Computers like perfect order; the real world rarely complies. A more fundamental problem is that they lack the rich understanding of the world that we have by nature. Strangely, beating Kasparov at chess is easier for a machine than holding a conversation about last night's telly.
Through our flexible understanding of the world and our ability to think creatively, we can solve the complex many-faceted problems of daily life. Our brains enable us to scale mountains, make websites, fall in love, build cities, and send men to the moon. QRIO can recognise faces and say "Hello". AI is still a young subject. It has a long way to go, but people are constantly pushing the boundaries. Meanwhile the only true intelligence remains our own. ®