Robots as disaster relief workers could save human rescue teams from having to undertake dangerous operations. And as nursing assistants they could help to cope with the problems of an ageing society with more and more people needing assistance. It will be a few years yet before they are able to undertake such tasks, however.
After all, two-legged robots today cannot move autonomously across an uneven floor – their motoric skills do not adapt quickly enough to unfamiliar terrain. If the machines learned as well as insects, not to mention human beings, a rocky path at least would no longer present a problem. The Max Planck ETH Center for Learning Systems aims to equip them with this ability to learn.
“We not only want to solve application problems, such as teaching a two-legged robot how to move on uneven ground,” says Bernhard Schölkopf, a Director at the Max Planck Institute for Intelligent Systems in Tübingen and one of two Co-Directors of the Center in addition to Thomas Hofmann from ETH Zurich. “We first want to understand what constitutes the intelligence of living beings which enables them to organize perception, learning and action and to act successfully in a complex environment.”
The researchers then want to use the insights from these fundamental investigations to further develop the methods of machine learning. These methods are already in use today to detect statistical regularities in large sets of data. But they are always limited to specific tasks. A method for reliably recognizing faces on images, for example, does not help a robot to practise moving steadily over any type of terrain. “The learning ability of humans in particular is largely independent of the specific task, in contrast,” explains Schölkopf. “If we have a better understanding of how what has been learned can be transferred to different tasks, we could possibly develop artificial systems which learn like living beings.”
The general principles of learning should then not only impart intelligence to robots, but also to the software which analyzes large volumes of data, for example. Computers should no longer determine only statistical relationships in large sets of data, but also causal ones. They should autonomously estimate the effect of genetic modifications in data about the genetic code or protein interactions; these are causal relationships about which even medical professionals still have no knowledge to date.
The Max Planck ETH Center, which is the home of the collaboration between researchers from Tübingen, Stuttgart and Zürich, builds on an existing cooperation between the Max Planck Institute for Intelligent Systems and the ETH Zurich in the field of machine learning. Its objectives are not only scientific collaboration, but also the joint use of research infrastructure and the training of doctoral students. Joint summer schools and workshops will be organized via the Center. The Center will receive total funding of five million euros in the first five years, and this will be contributed equally by the Max Planck Society and the ETH Zurich.