Wednesday, September 16, 2009

Machines that learn


I once played ball with the SONY robot dog Aibo. It was a surreal experience, but nothing that struck me as philosophically interesting. Just massive computer algorithms and some wonderful engineering. Aibo shows every evidence of being the product of an Intelligent Designer.

The Japanese have been the front-runners in this sort of thing -- electromechanical versions of Vaucanson's duck. But here is a picture of a rather more interesting project -- a European robot called iCub. The idea here is not to simulate human activities, but to investigate human learning.

iCub is about the size of a three-and-a-half year-old child. Its brain is also that of toddler. That is, the programmers gave iCub the basic skills of perception and data processing that might come hardwired in a human infant: visual, auditory and tactile senses that are filtered to determine what sensations are most relevant to the task at hand; a comparison of sensations with what has previously been experienced; the use of prior experience to decide "what if" for present actions. iCubs have been (and will continue to be) distributed to researchers who want to know how children learn motor skills and even language. Neurologists will be closely following the results.

The 20th century was the century of physics (quantum and relativistic) and molecular biology (genomics and proteomics). The two most exciting areas of scientific research in the present century will be -- IMHO -- developmental biology (how a single cell develops reliably into a platypus or a hummingbird) and neurology (how the brain enables us to navigate through a complex world of physical objects and abstract ideas).

Will solving one or the other problem require a modification of the materialist paradigm that has prevailed in science since the Scientific Revolution? I doubt it. I wish I were going to be around to see if a descendant of iCub can be taught to understand and communicate in human language, starting only with basic sensory input and data processing algorithms.