Bayesian Concept Learning

by Kevin Murphy

Consider the problem of learning to understand the meaning of a word, such as ``dog''. Presumably, as a child, one's parents point out positive examples of this concept, saying such things as, ``look at the cute dog!'', or ``mind the doggy'', etc. It is very unlikely that they provide negative examples, by saying ``look at that non-dog''.

However, most machine learning methods require negative as well as positive examples. In this talk, I summarize work by Josh Tenenbaum that shows how one can learn concepts from small amounts of positive only data. This provides a very simple illustration of the power and elegance of Bayesian methods.

Back to the LCI Forum page