Fascination Machine Learning

Ever since machines were built, people have wondered whether one could build 'learning machines'. These machines improve their behaviour over time with experience on specific tasks (just as humans do). So game playing machines (such as e.g. chess) learn when they win more games with experience, expert systems for e.g. medical diagnosis learn when they get to diagnose their patients better, robotic systems learn when they achieve their goals faster, ...

There at least two reasons for the interest in machine learning. First, although there is no consensus on the nature of intelligence, it is generally accepted that something is not intelligent unless it is capable of learning. Thus if one is interested in intelligente machines (or artificial intelligence) one should also be interested in learning machines. Second, learning machines would come be very practical and would simplify our lives to a great extent. Imagine that instead of programming our computers or robots explictly, we could simply train them to perform these tasks.  This would be much more flexible and less cumbersome.

Although we are still far away from machines that learn like we humans do, a lot of progress has been made in the past few decades.

Research on machine learning started with the seminal work by Arthur Samuel, who developed a learning checker playing program  during the 60ies. The systems learned by replaying games among grandmasters and by playing against itself. After a lot of training it performed quite well (certainly given the machines available in those days). Today, the same type of technique is incorporated in the TD-gammon system by G. Tesauro  that plays at the world-champion level, and it is also being applied to more complex games such as chess.

ML in Freiburg

Machine learning in Freiburg focusses on so-called inductive learning. Induction is the process whereby one generalizes speicfic observations into general laws. For instance, if you see two swans and they are both white you may be tempted to believe that all swans are white. One difficulty with induction is that you can never be certain about your conclusions. (Indeed, the next sway you might observe could be black and would falsify your theory).

Inductive learning is at the heart of the field of data mining which aims at discovering new 'nuggets' of knowledge in large databases. E.g. if you have data about the symptoms and deseases of different patients, you might be interested in finding general rules that relate symptoms of patients to the disease they suffer from. Such rules are useful when diagnosing new patients.Data mining systems are able to induce such rules from databases. If these rules are previously unknown to the medical doctors, they also constitute a new piece of scientific knowledge. Hence one sometimes also talks of machine discoveries.