Q & A on Specialized Learning in Primates

Why Specialized Learning?

Here is a naïve view of learning: The best kind of learning system is one that can learn anything. The smarter an agent is, the more it can learn. If artificial intelligence really worked, you wouldn't have to program an agent, it would just learn, like kids do.

Why is that naïve? Because learning is a form of search. It's a search for the right parameters for an intelligent system so that the system works well. The more possible parameters that have to be searched, the longer the search will take. In fact, unless you had an infinitely fast computer (and you can't have one) searching every possible parameter would take forever.

There are a few other reasons that view is naïve. First, kids can't learn everything -- neither can adults. Can you memorize a phone book? Second, kids already have a lot of information built into them, that they have inherited genetically. For an artificial agent, all that information would need to be programmed (or evolved, but it took kids 4 billion years to evolve!) Third, we do spend a lot of time teaching kids, which is kind of like programming. And fourth, kids take years to learn enough so they can have jobs. Would you like to buy a program and then wait 16 years before it starts working?

By specializing learning, we can give an agent enough information that it can learn quickly and reliably exactly what it needs to know. This information is called bias. So bias is a good thing for learning, even though it may make some things harder to learn.

Why Study Primates?

All animals are highly biased --- they learn some things very well, but other things not at all. But primates learn more than most animals. Consequently, they might be a good model of how to learn a lot even within a specialized framework.

Also, understanding primate learning is intrinsically interesting. Especially because people are primates. Why not study human learning? First, plenty of other people are doing that already. Second, adults learn a great deal of information culturally, because they have language. Infants don't have language, but develop so quickly it's hard to tell when they are learning and when they are developing. So adult non-human primates might be a little easier to model.

Does it Make Sense to Look for a General Mechanism for Specialized Learning?

Maybe not, it does seem like something of a contradiction. But we know there are general principles underlying how the brain learns, such as conditioning. On the tube project, we may have found a slightly higher-level general purpose mechanism for at least one kind of specialized learning: learning to perform new tasks with established skills.

Can you justify working on primates when you aren't studying life and death matters?

This question is a little weird since I actually do all my research on computers, not animals, but I do get asked it. Further, I do work with `real' primatologists, and I am not hypocritical enough to say that I think I am better than they are. So let's get to the point.

This question implies that it might be OK to work with primates if it was the only way to save a human life, e.g. if you were doing medical research. Well, the World Health Organization (WHO) has figured out that the leading cause of ill-health in the world is war. This isn't just because of bombs & shooting, it's mostly because of the disease and starvation that accompanies the mass migration of people trying not to get killed. Note that not only people get killed. There's a good chance our closest surviving relations, the bonobos, will be totally wiped out by the wars in Central Africa. This is an inexpressible tragedy as we are just beginning to understand the social order of a species far more peaceful than either humans or chimpanzees. (Obviously, the millions of human deaths and the losses of human culture that have resulted from that war are also incomprehensibly tragic, but no one ever asks me whether it's OK to do AI research with human subjects.)

So I would argue that research in the behavioral, political and social sciences is a matter of life or death. I can't see justification for animal suffering for cosmetics or in factory farms (for food.) I'm not even sure about the huge numbers of animals put down because they are unwanted pets every year (many times more animals are killed for this reason than for science.) And I do think that primatologists, like developmental psychologists, have deeper ethical obligations than most scientists for taking the best care possible of their subjects and a higher standard of justification for their research programmes. But I don't have difficulty working with primatologists who do good science.

page author: Joanna Bryson
last updated: 27 January 2005

Back to the Primate Learning Page