News for hackers. Stuff that matters.
Re:Artificial Intelligence Is Already Weirdly Inhuman by rcthompson on Thursday August 06, 2015 @07:24PM (#10020104)
I think the "static is a cheetah" example just highlights that the neural net is not identifying the best features with which to identify a cheetah. Or alternatively, if, during training, the neural net was only fed pictures of nature with or without cheetahs in them, then what it's really telling you is not the probability that a picture contains a cheetah, but rather the conditional probability that a picture contains a cheetah given that it is a picture of nature. In other words, that picture of static is most likely well outside the domain of the training set, so classifying it involves a large extrapolation, with all the attendant amplification of errors.

Perhaps what we need is a classifier that can tell when a picture is significantly outside of its training experience and say "I've never seen anything like that before" instead of giving an arbitrary classification.

There are no comments on this comment yet.