Clustering (or extracting mutual information) is necessary to perform some kind of improvement, as things you find mutual information between provide the definition of what constitutes an improvement. If you have a concept that is being detected by a certain feature detector, how do you improve that feature detector? Each modification applied to it slightly alters the set of things that it detects. How do you know which change is a good change? One way is to find other concepts that correlate with this one, and move the concept towards the mutual information between these concepts, bias the concept towards the states that light more feature detectors in approximately the same situations. It's also possible to have a predefined reference concepts with which to search for mutual information, which is in a way the case for reinforcement learning, so that you adjust other concepts in correspondence with the standard. In other cases, the goal is specified by many correlated constraints, which includes direct instructions and the probability distribution of exemplars from the target domain, which is the case in semi-supervised learning. Focusing on mutual information of multiple clusters corresponds to finding the clusters of denser-than-average probability in the space of these features. Why clustering is so effective is another question.
-- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.com/ ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244&id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
