91av

Even after decades of research into artificial intelligence, machines still don’t think like human beings. Marvin Minsky, the discipline’s founding father, refuses to give up hope. His solution is to make machines more emotional – and feelings, he says, are simpler to model than rational thought. He talks to Amanda Gefter about the need for emotional machines, the inner workings of the human brain, and the future of AI.

Many people are disappointed at the lack of progress in AI since the 1980s. Why so little headway?

In the early years of computing, we found it easy to program machines to solve problems that people regarded as difficult, such as designing efficient aeroplane wings, playing chess, or diagnosing heart attacks. But none of those programs could do the things that people regard as relatively easy – such as making a bed, babysitting or understanding a story from a children’s book.

It is much the same today. Each program has only one specialised skill, and when anything happens that isn’t expected the computer produces absurd results or gets stuck in an endless loop.

In contrast, humans rarely get totally stuck because we have many different ways to deal with each situation or job. So whenever your favourite method fails, you can usually find a different approach. For example, if you get bored with one particular job, you can try to persuade someone else to do it or get angry with those who assigned it to you. We might call such reactions emotional, but they can help us deal with the problems we face.

You call your new book The Emotion Machine. Is that because you’re convinced that computers need emotions to help them think in the same way as people?

Yes and no. The goal of the book is to try to explain what gives people their unique resourcefulness so that we can make our machines more versatile. We all grow up with the idea that emotions and thinking are quite different things, that thinking is basically simple because it is mainly a matter of rational logic, whereas emotions are far more complex and mysterious. I take the opposite view: that emotional states are usually simpler than most of our other ways to think.

Do you have an example?

When someone gets angry, we can see that some of their mental resources switch off. They abandon some long-range plans and goals, and become less cautious and thoughtful. This frees them to be stronger and to think on their feet, making it easier for them to intimidate others. Similarly, a person who has fallen in love might describe a sweetheart as “unbelievably beautiful” or “incredibly intelligent”, which suggests the speaker has turned off the critical abilities that are normally used to recognise someone’s deficiencies.

Is there such a thing as pure rational thought devoid of emotion?

The traditional view of emotions is that they add extra features to thoughts, just like adding colour to a black-and-white drawing. This makes emotions seem very mysterious because we can’t imagine what those extra features are. However, if we regard each emotional state as suppressing some of our usual mental activities, much of the mystery disappears. Perhaps this is why we have hundreds of different words for emotional states, but we have very few terms for describing our everyday ways to reason and think.

The trouble comes from our failure to recognise that there’s no such thing as pure rational thought, because our thinking is always influenced by our current ambitions and biases. Besides, we take common-sense thinking for granted. It works so well that we feel no need to ask how we represent and retrieve the knowledge required for such thinking.

“There’s no such thing as pure rational thought”

Why is it so difficult to give computers common sense?

Every normal person learns millions of fragments of information. Also, each person accumulates many thousands of useful processes during their life. In contrast, most computer programs have access only to some specialised knowledge about one particular subject. However, this is changing. There are now several ambitious projects that are collecting large amounts of everyday knowledge and trying to find effective ways to organise and apply it.

How could such knowledge be applied to help us make machines more like people?

The Emotion Machine suggests an overall scheme that I call the “Critic-Selector model of mind”. The idea is to think of a brain or machine as a system that contains many different kinds of structures and processes: let’s call these “resources”. Then each of our different “ways to think” – whether emotional or intellectual – is simply what happens in that brain or machine when a certain set of resources is active.

Now let’s also add some other resources: Critics, which can recognise some types of problems we face, and Selectors, which can switch different sets of resources on and off. Each set of resources can make us think about things in a different way. We call some of these different ways of thinking “emotional” and others “intellectual”.

For example, if a problem seems too difficult, a Critic can turn on a way of thinking that can try to split the problem into parts, change the way you describe it, solve a simpler version first, find an analogy for the problem or even ask someone else for help.

How much of our lack of knowledge about the way our minds work results from our use of ambiguous language?

Most of our popular words for psychology date from before we developed our ideas about how machines could use diverse kinds of information. Today’s computers use many different ways of representing and storing a huge variety of types of knowledge and processes. Accordingly, my book suggests that each of our traditional words used in psychology refers to perhaps a dozen kinds of machinery.

Is “consciousness” one of these words?

I agree with writers such as Daniel Dennett that there is no single meaning for “consciousness”. In fact, one chapter of The Emotion Machine suggests more than a dozen ways in which a particular thought process can know about what other processes are doing. Furthermore, if those processes have different goals, then that person won’t have a single, unified self, but will have multiple “sub-personalities” that may need to compete for the resources they need.

When do you foresee us having sophisticated AI? What will be the major forces driving the development of AI in years to come?

AI researchers have developed many techniques for solving various types of problems, but few of them have tried to come up with schemes that combine multiple ways of thinking. I hope that these new ideas about minds will encourage students to further develop them and influence neuroscientists to design new kinds of experiments to test them. Progress is likely to be slow, though, because most students today have to base their careers on developing practical applications. There is not enough support for basic research in this area.