Today, I attended Tom Mitchell's talk. This was perhaps the best research I have seen in 6-7 years. If you have a moment, I encourage you to take a look at his paper summarizing this work.
In short, they found that when a person thinks about some noun, some parts of that person's brain get increased blood flow. It turns out that:
1. The parts of YOUR brain that increase blood flow are so similar to the parts of MY brain that increase blood flow, that I can build a machine learner to learn what brain states map to what nouns based on YOUR brain and then be able to predict what nouns I am thinking about based on pictures of MY brain.
2. By looking at what verbs are most commonly used (on the Internet) in combination with a set of nouns, taking proportional linear combinations of brain pictures for certain nouns allows correct predictions of brain pictures for other nouns.