Re: Corpora: robustness in statistical methods

Osborne M. (osborne@let.rug.nl)
Mon, 13 Sep 1999 11:03:45 +0200 (METDST)

One possible thread would be to look at the work done on noise-tolerant
learning: variants include classification noise (you mislabel your
training set) and malicious noise (you corrupt the examples,not just their
labels). The general idea here is to see how well you can learn given
arbitrary amounts of errors in the training sets.

Here are a few pointers to get going:

http://www.toc.lcs.mit.edu/~sed/research/IPL96-abs.html

http://www.research.att.com/~mkearns/

Miles Osborne