A
technique for
machine learning, similar in flavor to
neural networks. A
boosting system is a
classifier which determines whether or not an
input is a member of a
category (or, more generally, how good a
member of the category it is, ala
fuzzy logic.) Boosting works by considering each predictor out of a huge
space of possible predictors, and choosing the best one. It then compares its performance against correct performance and chooses the next best predictor to make it perform better, and so on.
Because you have to have a correct categorization for it to compare to, boosting is an instance of supervised learning and thus suffers from all the philosophical and implementational problems of other supervised learning systems.
Boosting systems are inherently linear but get around the general linear separability problem by including inputs which are combinations of features. From a cognitive science point of view, this is probably an egregious hack. From a pragmatic point of view, it's a great time-saver and very effective.