Teaching Robots

For a long time, the conventional approach to teachng robots has been a fundamentally wrong one. Often described as a 'dictionary' approach, the assumption was that you could pre-load a robot or AI with a bunch of terms, and their definitions in a 'look-up table' structure, and then associate those items with the sensory signature of each item and presto, you've got a robot or AI that understands what "give me the heavest ripe apple" should mean.

Fundamentally, though, this meant that other than handing you objects as a stationary, screwed-down, demo platform, you would never get anywhere near a serviceable, adaptable and useful 'product'. Recently, there seems to be an understanding that the appropriate solution will involve teaching in the same way humans (and other animals) learn. Vocabulary must be learned from in-context repetitive instruction. (e.g. Deb Roy at MIT media Lab). Trial and failure approaches are going to be the only ones that work in the foreseable future. At some point, perhaps, a virtual environment may allow for rapid trial-and-error training, but that process of building up linkages in the neural net approach is going to be fundamental to machine intelligence. Then again, the term "neural net" has been co-opted to mean some specific and boringly mundane things these days. Let's instead refer to learning in this way as a "biological associations network building"

HEY! That spells band - honest, I didn't pick the words to spell a snappy acronym.. although most acronyms do come about that way.

There was recently some discussion on this topic on Canada's "Quirks and Quarks" a science radio show on public broadcaster CBC. (See more at: ( CBC Q & Q Site )