ÐÓ°ÉÔ­´´

You are walking down the street and you see a woman you are certain
you know, but you just can’t ‘place’ her. You stand there trying desperately
to decide whether you should approach her and say hello. Is it someone you
know personally, someone who works in a shop you frequent, or maybe just
someone you’ve seen on TV? With luck she will walk on, and you will be spared
embarrassment. Perhaps later you realise that it was a colleague whom you
rarely see outside work.

Here is another example. You are talking to a man, perhaps before a
meeting, and you know exactly who he is. This is the fellow you met last
year who works in Newcastle and drives a Golf. The trouble is, you can’t
remember his name. While you are trying to recall it, rejecting many possibilities
to yourself in the process, you hope you won’t have to introduce him to
a colleague.

These are common – and madden-ing – experiences. But what is actually
happening here? In the first case, it seems that the brain’s mechanism for
recognising a person’s face is functioning incompletely. This sort of partial
blockage is intriguing because our memory for faces is, in general, remarkably
accurate: we have little difficulty identifying thousands of people on sight.
In the second case, the process of facial recognition is complete – it is
the name that eludes. Both types of breakdown provide valuable clues to
psychologists trying to establish how the brain recognises faces and retrieves
names.

Over the past decade considerable progress has been made in this field.
But the sheer complexity of the human nervous system makes it a very hard
task. Although neuroanatomy offers some useful evidence, it seems unlikely
that we will fully understand the recognition process by examining the brain
itself. One of the primary reasons for this is that the present state of
technology makes it impossible to trace the activity in all the brain cells
involved when we look at a person. And even if this were possible, it is
not clear that the problem would then be solved: would we really understand
how a particular program worked in a computer if we were able to trace its
effects through all the computer’s electronic components?

These difficulties led psychologists to adopt an ‘information processing’
approach, in which the brain is studied in terms of what it does rather
than in terms of where things happen. With this approach, we can view the
functioning of the brain as the activity of a set of modules, each of which
does a particular job. If we can uncover the jobs these modules do, and
how information is passed between them, we will begin to grasp the functioning
of this most complex organ.

One of the most useful tools in this approach is the errors people make
in recognising each other, because the ways in which a system fails may
highlight how it works. In the mid-1980s Andrew Young, a professor of psychology
now at the University of Durham, set out, with his colleagues, to establish
whether there is any pattern to such errors. He asked 22 people to keep
a diary of all the mistakes they made in recognising people over seven weeks.
The results showed an interesting pattern. People frequently made three
types of error: they would simply fail to recognise someone as a person
they knew at all; they would recognise someone as familiar, but be unable
to recall the context in which they knew them; or they would fully recognise
the person, but be unable to retrieve their name. Interestingly, no one
ever reported a fourth type of error, in which they recognised a person
as familiar and knew their name, but didn’t know anything else about them.
As Young says, we never find ourselves saying: ‘I know that’s Dustin Hoffman,
but who is he?’

Results such as this led Young, together with psychologist Vicki Bruce
of the University of Nottingham, to propose a hierarchical model of face
recognition. They suggested that the system works like this: first, you
see a face and recognise it as familiar or not. If you recognise it, you
may then go on to access everything you know about that person – a body
of information called identity-specific semantics. Finally, and only if
you have succeeded in passing the previous two stages, you begin the process
of retrieving the person’s name. This model seems to tally with the errors
that people make, but it has a further, and surprising, implication. For
some reason, the brain seems to store information about people’s names separately
from all other information about them. Why has the brain evolved to separate
the different types of information in this way?

The ‘separate store for names’ hypothesis draws support from sources
other than errors. Psychologists often use evidence from studies of reaction
time. In these, people are shown pictures and are asked to make decisions
about them as quickly as possible, while researchers measure their reaction
times. Using this technique, Young and his colleagues have shown that the
hierarchical model seems to fit the data. In their study, subjects were
shown a picture of a famous face – either politician or pop star – and asked
either to name it, or to judge the person’s profession. Naming took longer
than judging the occupation, even when subjects had seen all the photographs
before the experiment and knew which ones would turn up – although not the
order in which they would be shown.

In other studies, researchers have found that subjects display the same
pattern when presented with information other than professions, and with
different tasks. If people were asked to push a button in response to the
question, ‘Are these two people both American?’ or ‘Are these two people
both called John?’, the matching of names took more time than the matching
of nationality. These reaction-time studies seem to suggest that we access
personal information before we access a person’s name: a pattern consistent
with the hypothesis that names are stored separately.

A limitation with studies of this type is that it is hard to control
for other factors which influence reaction time. For example, we tend to
retrieve words which turn up frequently in language faster than those which
occur rarely. This makes it hard to compare names with descriptive words
such as ‘athlete’ or ‘British’. Kathryn McWeeny and her colleagues at the
University of Lancaster got around this problem by using the same words
sometimes as names and sometimes as other personal information.

In McWeeny’s experiment, people were shown pictures of unfamiliar faces
and given a name and an occupation for each, which they were then asked
to memorise. The researchers chose names which were also descriptive of
occupations, such as ‘Baker’ and ‘Cook’, so that they could use them in
either category of information: some subjects would learn about a person
called ‘Mr Baker’, while others would learn about a person whose occupation
was ‘baker’. What they found was compelling. It is harder to learn a name
than to learn an occupation, even when the same words are used. So, it takes
us longer to learn that a person is called Mr Baker than to learn that a
person is a baker. Again, McWeeny’s evidence seems to support the hypothesis
for a separate store for names.

Neuropsychology provides the final source of evidence. ÐÓ°ÉÔ­´´s have
known for over a century that people suffering from brain damage caused
by an injury, tumour or stroke can exhibit specific patterns of breakdown.
It is possible to lose certain functions without losing others. Recently,
Brenda Flude and her colleagues at the University of Lancaster came across
a patient whose behaviour seemed to have some bearing on the issue of name
retrieval. The patient, EST, was a well-educated man in his 60s who had
suffered from a brain tumour which had been surgically removed. He was tested
on his ability to recognise people. When shown photographs of famous and
of unfamiliar people, he was able to sort them into two piles – familiar
and unfamiliar – with almost perfect accuracy. He could also tell the researchers
quite a lot about these people – for example, that one was a past prime
minister, one an actor in cowboy films, and so on. But EST was almost totally
unable to retrieve the names of any of these people. It seems that, somehow,
EST experiences a block at the stage between retrieval of personal information
and retrieval of names. To date, no one has found a brain-damaged patient
who shows the complementary deficit – an ability to retrieve names but no
other personal details. So, on the basis of present evidence, we again appear
to have a case for the three-stage hierarchical system of face recognition
proposed by Bruce and Young.

One odd aspect of this hypothetical mechanism is that the brain’s compartmentalising
of information about people must have evolved very recently. In our own
and many other cultures, names derive from personal information. So, as
we all know, people are called Smith and Cooper precisely because their
ancestors were once blacksmiths and barrel-makers, or coopers. It was only
recently that European names ceased to have these associations. Faced with
this change, how does the brain ‘know’ to allocate ‘Mr Baker the name’ to
one store of information and ‘the baker’ to another?

Over the past year, Bruce and I have developed a computer model of the
processes involved in recognising people. The aim was not to produce a machine
that can recognise faces – a problem many researchers are tackling with
limited success – but to try to understand how the human brain might do
the job. Our approach has been to write computer programs which function
in ways suggested by existing theories of face recognition. This is a technique
used by many scientists, for example meteorologists, who feed computer programs
with data about climatic conditions and make predictions from the outcome.
The method has allowed us to experiment with the programs and see what happens
under different circumstances.

Our experiments have suggested how the brain might deal with names and
other information about people. What is intriguing about this computer model
is that it produces ‘behaviour’ which fits in with everything we know about
face recognition, but does not need the problematic separate store for people’s
names.

We wrote the program using a method called ‘interactive activation and
competition’ or IAC, a way of building models which two American researchers,
James McClelland and David Rumelhart, introduced in the early 1980s to study
reading. Our model is made up of a set of simple ‘units’, each representing
different aspects of the brain’s recognition system for people. The units
are clustered according to the job they do. They are also interconnected
in ways that represent the ‘flow of information’ involved when we recognise
a face. At the core of the model there are two clusters of units, called
‘person identity nodes’ and ‘semantic information units’. The idea is that
there is one identity node for each person we know. When we see a face,
or recognise a person by other means, such as hearing their voice on the
telephone, that person’s unit becomes active. This activation signals that
the person is familiar, but does not prompt any further details. The more
active the unit, the easier will be the decision that the person is familiar.
The second stage – retrieval of further information – is tackled by a second
cluster of units, each of which represents something you know about that
person. So there will be units such as ‘British’, ‘dead’ and ‘athlete’.
When any of these units reaches a certain level of activation, the appropriate
information is retrieved.

The crucial aspect of the model is that each ‘person’ unit is linked
to the information cluster. So the ‘person’ unit representing Neil Kinnock
is connected to units such as ‘politician’, ‘Welsh’, ‘married to Glenys’
and so on. These connections enable information to flow in both directions:
if the Kinnock person unit becomes active, it then triggers the ‘politician’
information unit, and vice versa.

Using this system, we have experimented with various ways in which the
program could model the retrieval of people’s names. The obvious thing to
do was to follow the present theories and add a third cluster of units representing
names. But this turned out to be unnecessary. We started with a database
of people and ensured that the program knew a certain number of things about
each, such as profession, nationality, home address and name. We made the
assumption that people store names in the same way that they use them. For
example, we might store the names ‘John Kennedy’, ‘John Lennon’, ‘John Smith’
and ‘John’ – the last being someone, such as a close relative, whom we know
so well that we don’t normally use their surname. We decided to include
these names with the general cluster of information units, and watch what
happened.

When we connected the person units to the information units, we noticed
that while certain details of personal information are common to many people,
names are usually unique. So, we can know many politicians, many people
who are American, and many who are dead, but in general we know only one
John Lennon. This means that the unit representing ‘American’ is connected
to many person units, whereas the unit representing ‘name is John Lennon’
is connected to only one.

If we activate a person unit – by seeing a face or, perhaps, hearing
a voice – we can then observe how the information units behave. What happens
is that the information unit representing a name always becomes active more
slowly than other information units, and that it always ends up being activated
less than the other units.

Why does this happen? Imagine that Neil Kinnock is recognised as a familiar
person, that is, his person node becomes very active. As a result, all the
information units connected to this unit are triggered. Now, some of these
information units – for example, ‘politician’ – are connected to other person
units – such as Roy Hattersley, David Owen and John Major. As information
can flow in both directions, these person units start to gain in activation.
The activation is not sufficient to trigger a decision that the people represented
are familiar; but it is sufficient for information to be passed back again
to the ‘politician’ unit. As a consequence, all information which is shared
will become very active more quickly than information which is unique. Here
we seem to have an alternative model of name retrieval: what makes names
comparatively hard to retrieve is not that they are stored separately from
other information, but that they are generally unique.

This account appears to fit the data. The slower rise of activation
in the name units is consistent with the fact that it takes people longer
to retrieve names than other information. The lower activation of name units
is consistent with the fact that we seem to be blocked on names even while
we are able to retrieve other information – if we are going to forget anything,
it will be information on units with the lowest activation. And, returning
to EST : if we assume that he is generally less able to make decisions based
on activation in the information units, we would expect him to lose names
first because these, along with other unique information, are the hardest
to retrieve. Further, using this computer model, it is impossible to conceive
how shared information could be lost from memory while names could still
be retrieved. This is encouraging because such a pattern does not seem to
occur in people with brain injuries.

Although our model offers one explanation for difficulties with names,
it is not the only possible solution. Gillian Cohen of the Open University
has recently suggested an alternative. She proposes that names are hard
to retrieve not because they are unique, but because they are generally
meaningless.

To test this idea, Cohen chose a set of names, some of which were also
words for meaningful professions – such as Mr Cook – and some of which were
not – such as Mr Ryman. She then taught her subjects about a set of people,
each with a name and occupation. When the subjects learned about a person
with a meaningful profession but a meaningless name – such as Mr Ryman who
is a cook – names were, as usual, harder to recall. But when the person
had a meaningful name and a meaningless occupation – Mr Cook who is a ryman
– the name was easier to recall than the profession. These results support
Cohen’s hypothesis. Perhaps we are just better at recalling meaningful things,
and the apparent difficulty with names is a side effect of this ability.
It is going to be hard to disentangle Cohen’s theory from our own; a person
who has the profession ‘ryman’ has both a meaningless and a unique occupation.

Whatever emerges as the truth, computer models of the kind we have developed
are a useful tool for unraveling the complex functioning of the brain. They
not only help to clarify our thinking about the implications of theories;
they can also suggest new areas of study. Similar approaches are currently
being used in trying to understand many other aspects of human cognition,
such as understanding of language, memory and picture recognition. The danger
lies in dismissing these processes as trivial because we generally find
them so easy.

Mike Burton is a lecturer in psychology at the University of Nottingham.