Training Humans

I am an amateur in the realm of neural networks (or any other classifiers in AI), but I have read a bit. It seems that in general, training goes like this:

  1. Provide each of the networks in the set with some inputs.
  2. Ask the networks for their outputs.
  3. Record how well each network answered the question.
  4. Tweak the networks according to how well they performed.
  5. Repeat until satisfied.

The tweaking step can be done in many ways. I'm familiar with evolutionary algorithms where the weak networks die and are replaced with new ones, and ones where calculus is used to find local maxima in the performance level of networks over all possible networks. Whatever the method, the idea is that feeding past performance into the training should allow better networks to be generated for the next round of testing.

How well do these algorithms converge on one that works? A better computer scientist than I could give you mathematical bounds, but I think an interesting comparison could be made against humans that were trained in the same manner:

  1. Provide each of the people in the group with some inputs.
  2. Ask the people for their outputs.
  3. Record how well each person answered the question.
  4. Tell the people how well they performed, and allow them to discuss.
  5. Repeat until satisfied.

A simple experiment might involve showing a group of people a series of colors and asking them for a number. They have no idea what the expected correlation is – perhaps blue should be 0, red should be 1, green should be 2, etc. Could they "figure it out" in fewer rounds than neural networks?

The discussion part is the magic. Humans can train on their own, but 10 people can cooperate. I imagine a group of people agreeing to cast a wide net by each guessing a different number within a range in order to "feel out" what the space of appropriate answers is. They could then develop a strategy and test it, repeating if it doesn't work out. They could designate a few people to test possible strategies while others keep guessing.

A single person could do the same thing, but it would take much longer, and they would be unlikely to be able to strategize as well as several diverse people.

I wonder whether new training algorithms could be developed by observing the ways the best groups strategize.

Interestingly, I imagine that a group of people who were repeatedly subjected to the experiment would get very good at cooperation and creating strategies. Repeating the experiment sounds like a great way of evolving the method of self-training, which sounds valuable in itself.

Posted Feb 13, 2008, in the night.

Comments

Click here to view the comments on this post.