Extracting knowledge from data

In the early years of AI, many researchers were overly optimistic. They envisioned a not-so-distant future where software could routinely and flawlessly translate between different languages, where computers could have intelligent conversations with people (the Turing test), where anthropomorphic robots would be a part of our society, and so on. We're clearly not there yet: researchers have underestimated the difficulty of creating intelligent machines, and many AI programs still grapple with everyday human tasks. 

So has AI failed? On the contrary, there's a lot more AI in the software you use daily than you would expect, for example in Google's search engine or in your spam filter. De Raedt reassures us: 'Although researchers have abandoned the overly ambitious science fiction expectations of AI, there has been tremendous progress in various subfields of AI, such as robotics, machine learning, vision, natural language processing, and so on. Some AI research also branched off and is not anymore known under the AI umbrella. For instance, think about bioinformatics, where many data analysis tasks are performed using machine learning techniques. We don't want to imitate human intelligence anymore, but we want to create software that takes the right actions in a specific environment to achieve a useful goal.'

Challenging domains

This is also a fundamental difference between human and artificial intelligence: while we humans can adapt to different environments, all current intelligent agents are limited to a specific environment. 'We have many good systems for specific and even very complex environments, but from the moment they leave their target environment, they are useless,' De Raedt says. De Schreye adds: 'AI research always begins in a rather limited domain, and gradually you want to expand this domain, but if you expand the domain too much, the problem you are researching changes substantially.' 

That's why most well-known advances in AI have a limited domain, of which chess is the best known example. Recently, there has been a lot of progress in much more challenging domains. There's the Robot Scientist, a robot that makes scientific discoveries all by itself, the DARPA Urban Challenge, were unmanned vehicles have to drive autonomously in traffic, and the RoboCupRescue project, where robots go on a search and rescue mission. De Raedt believes that these research projects will advance enough in the coming decade to solve their tasks satisfactorily: 'The tasks AI is able to solve become more and more spectacular.'

A tale of two approaches

There have always been two big 'schools' in AI research: one that reduces intelligence to logic, rule-based systems and symbol manipulation, and another one that tackles all problems using sub-symbolic and statistical methods. 'By the 1980s, progress in symbolic AI seemed to stall, while in the last 10 to 15 years statistical methods have brought a lot of progress in information retrieval, data mining, machine translation and natural language processing,' De Schreye maintains. 'Nevertheless, I have some issues with this shift from rule-based to statistical systems: it's relatively easy now to create systems that work 95% of the times, but do we want this? Are these systems robust enough?'

Both professors think that this is why we will have to combine both approaches if we want to continue making progress. 'I see this happening now,' De Raedt says, 'the low-level sensory tasks in robotics and machine vision and the high-level tasks such as reasoning are going to converge in the next 10 years. In our own division, we are working on the integration of logic with statistical methods.'

An explosion of data and processing power

According to De Raedt, one relatively new but really helpful aspect is the explosion of data in our information society: 'More than ever we have access to a wealth of data. If you apply data mining on a large enough dataset, you can extract a lot of information. This is the power of Google: they use relatively simple algorithms that return very useful results because their search spiders have crawled so much data.'

De Schreye adds: 'I have seen many times that a company launched clever software using the newest technology but had to shut down the project after a year. You almost need a postdoctoral researcher to keep such advanced systems running. So smarter algorithms are not always better. It's much easier to just throw a lot of raw computing power at a problem, using algorithms that are simple to understand and to adapt.'

However, De Schreye warns that this doesn't mean that we can tackle all problems with enough computing power: 'Running simple algorithms on more and more powerful computers generally improves the results for specific domains. But when you move on to integrated AI systems, this approach breaks down. For instance, people like Ray Kurzweil use Moore's law to argue that computers will soon have the same processing power as human brains and hence become intelligent. But he is talking about intelligence in a broad domain, and I don't believe simple algorithms suffice for this. You don't get intelligence for free.'

Useful applications

When asked about the future, the two researchers are reluctant to predict which AI applications we can expect in the next decade. De Schreye says thoughtfully: 'Ten years ago, no one could predict the technologies that have the most impact on our lives now.' De Raedt, though, has one prediction: 'Software will adapt more and more to our habits. This will turn our smartphones into autonomous assistants. For example, when you're planning a meeting, your smartphone could use data from the past to suggest a suitable date. I think these kinds of systems can be handy for elderly people too: when something unusual happens, the system can evoke an alarm, for instance when the person enters the wrong bus or when he falls. There is research going on in this domain, for example the CALO project (Cognitive Assistant that Learns and Organizes) in the US, and I think we'll see the first applications in ten years.'