What has really changed our thinking over the past thirty years?
We have probably learned more about children, especially very young children, than in all of recorded history. We now know that babies know much more about the world than we would ever have thought possible. They have ideas about other human beings, about objects and about the world 鈥� right from the time they are born. And these are fairly complex ideas, not just reflexes or responses to sensations.
Even more interesting is the knowledge that, from the beginning, extremely powerful learning capacities of several different kinds are also in place. Newborn babies have an initial 鈥渢heory鈥� about the world and the inferential learning capacities to revise, change and rework those initial ideas on the evidence they experience from the very beginning of their lives. Those capacities are much more powerful than can be explained by traditional ideas about learning: they involve much more than association and conditioning. Thanks to ideas from developmental psychology, computing and cognitive science, we are just starting to explain what those inferential learning mechanisms might be like. It is very, very exciting.
Advertisement
Traditional theory pictures babies as blank slates on which experience writes, or as having a neurologically determined part that is shaped by evolution, and a cultural, socially determined part that is shaped by learning. Your ideas sound very different.
Yes. Remember Ulysses鈥� boat? Ulysses spent decades on his epic voyages, and as he travelled he had to continually repair and rebuild his boat. At the end of the voyage, perhaps nothing remained of his original craft. Our development may be similar: we rewrite ourselves as we grow. Babies are like little scientists, continually getting data and overthrowing theories that no longer fit the new evidence. By the time we are adults, we will be completely different from how we began. We may bootstrap our way into the future.
Over the past three decades we have become extremely good at charting the changes in early life. We can now say with some certainty, here is what a baby is like at birth and here is the knowledge they have. Here is what is there at one year, at three years and so on. That鈥檚 a huge accomplishment, but it has really been 鈥渘atural history鈥�. Now we are at a turning point where we can begin to come up with better ideas about how the learning mechanisms might work.
Your new ideas about learning mechanisms have been built on all that research into key developmental problems 鈥� that 鈥渘atural history鈥� as you call it. Let鈥檚 turn to just one part you have been involved in: how babies come to know about other minds. What are some of the key findings?
An elegant story has emerged from exploring the natural history of the mind. The first big surprise came from Andy Meltzoff at the University of Washington in Seattle, who I鈥檝e worked with since the 1970s. He shocked everyone when he showed that newborn babies can imitate facial expressions. The earliest study was at just 42 minutes after birth. He found that a newborn baby could imitate him pulling faces and poking his tongue out, for example. That means the baby can map what it sees in another human鈥檚 face onto its own face, even though that baby is far too young to recognise itself in a mirror. If you asked a philosopher what intuitive skills a baby would need to come into the world with they would never think of this, but it does have obvious survival value.
Babies at birth can also distinguish human faces and voices from other sights and sounds, and prefer them. Within a few days, they recognise familiar faces, voices and even smells, and prefer them to unfamiliar ones. Within the first 9 months, they can tell the difference between happiness and sadness and anger, and which tone of voice goes with which expression. All this makes sense as the baby鈥檚 initial world is that of its carers. At this stage babies are still reacting directly to the people around them.
How does it change as babies grow older?
One-year-olds have a radically new understanding of people. They start understanding that other people鈥檚 actions, emotions and perceptions can be directed at a separate external world. They will look where other people point and even know how they should feel about something by seeing how other people feel. They can work out what to do with objects by looking at what other people do with them. They modify what they do, such as reaching out for an object, according to whether a parent reacts with smiles or disapproval. By 18 months, there鈥檚 another big change as babies begin to understand that other people鈥檚 desires and actions can differ from their own.
One of my former students, Betty Repacholi, did a beautiful experiment illustrating this. She showed babies two bowls of food, one full of crackers that the babies loved and the other with raw broccoli that they hated. Betty would taste them and make a face to show which one she liked and which she didn鈥檛. Then she would put the bowls near the babies and hold out her hand for the baby to give her some. If her face showed she liked crackers, the babies handed them over. But if she showed that she liked broccoli not crackers 14-month-olds, with their innocent assumption that everyone must love the same things, still gave her crackers. Eighteen-month-olds are wiser and can understand that other people can have different desires. They gave Betty the broccoli, even though they despised it.
So by 18 months, babies have begun to understand that different people may want different things. Have they grasped that other people think differently about the world?
No. There is a big difference between desires, which are directed at things, and understanding that other people can have different beliefs. Beliefs are very different from the objects in the world, as they are hidden away in people鈥檚 minds. The emergence of so-called theory of mind, which enables understanding of other people鈥檚 beliefs, is another radical change in a child鈥檚 development and it does not take place for a couple more years. The now classic experiment is to show children a familiar box of candies. They assume that there is candy inside it, but it鈥檚 actually a trick because when they open it, it is full of pencils. They are astonished. We ask the child simple questions: what did you think was inside the box? What will your friend think is inside it? This lets us explore whether children understand about other people鈥檚 beliefs. Four-year-olds will get the answers right, but children just turning three will not. They will say that everyone will know there are pencils in the box. They will even say that they thought there were pencils in the box.
Even though those children were astonished when they opened the box and found it full of pencils?
Yes. It is very counterintuitive and quite shocking when you see it for the first time. I have done experiments which suggest that three-year-olds seem unable to remember how they learned about something. They may have very good memories of everyday events but they don鈥檛 seem to understand their own minds any better than they understand the minds of the people around them.
Wouldn鈥檛 it also mean that children of this age can鈥檛 really tell lies?
Yes. To deceive people or to recognise that you are being deceived you need to be able to understand the difference between what they believe and what you believe, and how to create false beliefs in the mind of another person. Two or three-year-olds are scarcely able to lie. They may understand lying as a strategy for getting out of a tough corner but not how it really works. A three-year-old standing on the other side of the street which she has been expressly forbidden to cross will say: 鈥淚 didn鈥檛 cross the street.鈥� They are bad liars just because they don鈥檛 seem to understand what it takes to make someone believe something Real lies only begin to emerge at the age of four.
What about autistic children?
Children with autism don鈥檛 develop this way. They have trouble imitating facial expressions, they don鈥檛 point or follow people pointing in the same way as ordinary children do, and they don鈥檛 understand false beliefs like the candy box example until much later than usual. Children with autism seem to lack the fundamental presupposition that they are like other people and other people are like them. This unquestioned first principle, this axiom of our everyday psychology, is paradoxically part of what allows most children to discover all the differences between themselves and others. The difficulties of children with autism suggest there is an innate foundation for our understanding of mind. On the other hand, the extended unfolding of different types of knowledge about the mind suggests that we have to build on that foundation.
Your analogy of babies being little scientists who possess the capacity for discovery, continuous change and theory rebuilding seems to run pretty deep.
Yes, our 鈥渢heory theory鈥� is simply the theory that children are born with intuitive theories of the world, analogous to scientific theories, which change in ways that are similar to scientific theory change. All these big changes in an infant鈥檚 development can be thought of as theory changes: that is, changes in the way the baby understands what causes what. That, of course, leads to the big questions: how is the child鈥檚 knowledge of the causal structure of the world represented? What learning mechanisms allow this to happen? I think we have found an immensely exciting candidate in Bayes nets.
We have heard a lot about Bayesian statistics in recent years. Are Bayes nets the same sort of thing?
Bayes nets are not the same as Bayesian statistics but both come from the insights of Thomas Bayes, the 18th-century English clergyman. The interesting thing about Bayes nets is that they search out causes rather than mere associations. They give you a single representational structure for dealing both with things that happen and with interventions 鈥� things you observe others doing to the world or things you do to the world. This is important because there is something really special about the way we treat and understand human action. We give it a special status in terms of our causal inferences: we think of human actions as things that you do that are designed to change things in the world as opposed to other events that just take place. Bayes nets work perfectly because they treat interventions differently from mere associations.
What do they look like?
A Bayes net is a little like a map. It enables you to construct something that represents all the relationships between causes. Its overall structure enables you to infer hidden causes or to ask what happens if we change some of the events on the map.
We can draw a Bayes net as a so-called causal graph. There are lots of little arrows representing the causal influences of one thing on another, along with their probabilities. Then there are some formal rules on the overall distribution of probabilities that determine how one thing changes if you change another thing.
This doesn鈥檛 sound very much like how a brain works.
We aren鈥檛 saying that you are juggling little graphical maps with arrows and equations in your head the whole time, rather that causal knowledge and learning can be represented this way at a computational level. You can compare this to how the visual system works. Your brain can work out 3D representations of objects by comparing geometric disparities in the 2D images formed on your retina. It can do that because your visual system knows the right transformations to perform on those images, although it can occasionally be fooled by illusions. You don鈥檛 actually know all that geometry or how these transformations work. In the same way, Bayes nets provide the principles that allow you to see what causes what in the world 鈥� it鈥檚 like the geometry of causality. No other representation or learning mechanisms allows us to do that.
Can you give a simple example?
I鈥檒l try! Here is a simple experiment I have done with another former student, David Sobel. We play a game with three-year-old children to investigate their causal reasoning. We have a box that we call a 鈥渂licket detector鈥� that lights up and plays music when certain objects, called 鈥渂lickets鈥�, are placed on top of it. The aim of the game is to find out which objects are blickets. Obviously, if you place just one object on top of the box and it lights up, then that object is a blicket. Any simple associative learning theory would predict that. But with more objects we can make more complex problems.
Take these two examples. Put blocks A and B together on the box and the machine goes. A alone and the machine does not go. Children infer that B must be a blicket. Put blocks A and B together and the machine goes. Then A alone and the machine also goes. Children correctly infer that A is a blicket but crucially they also say that they think that B is less likely to be a blicket.
This kind of change, where views about an object (object B) are gained from an experience in which that object did not appear is characteristic of human learning. It is very hard to explain this using simple associative learning theories because object B was not present to associate anything with. Bayes nets take care of all these problems with ease 鈥� they can assess the entire causal structure.
This sounds fascinating, but other attempts to describe how humans acquire knowledge have also looked incredibly persuasive.
Yes, scientists can get infatuated. When connectionism came along, people thought it was the answer to everything. Now we know that networks can be good at learning but not very good at getting you the structured knowledge that you need to make inferences. Bayes nets enable you to take evidence and infer complex structure based on that evidence. And Bayes nets are just one example of a whole set of new ideas about learning. Cognitive scientists are finally really starting to understand how observing patterns of evidence in the world can lead to genuinely new knowledge. I might be wrong, but this time I think it is the real thing. I think this is where the action will be for the next decade. It鈥檚 wonderful!
