What we have not learned about "learning"?
Speaker: Berthold Klaus Paul Horn , MIT CSAILContact:
Date: November 15 2005
Time: 1:00PM to 2:00PM
Location: 32-G449 (Kiva)
Host: Metin Sezgin, MIT CSAIL
Metin Sezgin, 617 253 2663, email@example.comRelevant URL: http://projects.csail.mit.edu/dangerous-ideas/dangerous/www/
If a system does learn how to do something, what have *we* learned?
Often retrospective analysis of "learning" systems yield insights,
but not the expected ones.
Hoping a system will "learn" how to do something can be a way to avoid
doing the hard science necessary to solve the problem properly.
And in the end often not much may be learnt by the researcher,
since typically the learned state of the system is inscrutable.
Garbage in => garbage out applies to "learning" as much as it does to
computing in general. Extreme efforts to advance the algorithms and
mathematics underlying "learning" may not pay off as well as simple
efforts to provide better inputs by understanding vision better.
I hope to stir matters a bit in this short talk by exposing some
nuggets of discontent of this type. While I'm at it:
Tool building is a comfortable "displacement activity" engaged in
frequently in order to avoid working on the hard problems. It may be
helpful to review previous fashions in research that had similar
aspects. Such fashions seem to appear in a cyclical fashion and most
frequently infect the youth that has not yet been exposed to the
previous epidemic. Are you ready to be immunized?
See other events that are part of Dangerous Ideas Seminar Series Fall 2005
See other events happening in November 2005