¿ìè¶ÌÊÓƵ

Computers can perform a broad range of tasks that involve reasoning,
learning, planning and other functions usually associated with human intelligence.
Does that make these machines intelligent? Some scientists and philosophers
argue that mechanical intelligence is not the same as mind: computers can
never be conscious of their thoughts and actions and so lack an essential
part of human intelligence. Not everyone agrees: the assertion that machines
could be conscious unites two leading researchers who have otherwise disagreed
on the best way to push forward the frontiers of machine intelligence.

Opinion has long been divided as to how best to achieve machine intelligence.
Igor Aleksander, professor of neural systems engineering at Imperial College,
London, belongs to the school which wants to develop intelligent machines
by mimicking the way the human brain is built. He believes that a number
of attributes of consciousness can now be ‘captured’ in neural networks
– computer systems which mimic the workings of the brain. Marvin Minsky,
director of the Artificial Intelligence Laboratory of the Massachusetts
Institute of Technology in the US is identified with an opposing school,
which favours creating complex computer software that mimics the characteristics
of human intelligence. Like Aleksander, Minsky believes that machines could
be conscious – possibly even more conscious than humans.

Thinking about thinking

Minsky is one of the founding fathers of artificial intelligence. He
began his career in 1951 shortly after the first electronic computers appeared.
At the time, the British mathematician and computer pioneer Alan Turing
was publicly speculating that, by the turn of the century, it would be generally
accepted that machines could think. Minsky says he had been ‘thinking about
thinking’ since high school. He started investigating models of networks
made up of simple processing units modelled on brain nerve cells, or neurons,
but gradually became disillusioned with their potential.

In 1969, Minsky and a colleague at MIT, Seymour Papert, published a
book, called Perceptron, pointing out the limitations of the main neural
network model of the time, the ‘perceptron’. In their view, intelligence
could never emerge from such models of the human brain – what they called
the ‘bottom-up’ approach. Instead, they argued for a ‘top-down’ approach
to artificial intelligence – imitating human intelligence by programming
computers to process information by manipulating symbols that represented
knowledge and rules. The book put a brake on research on neural networks
for more than a decade, and Minsky’s top-down approach became synonymous
with the term artificial intelligence, or AI.

The paths of Aleksander and Minsky crossed briefly in the late 1960s,
when Aleksander worked at Minsky’s AI laboratory at MIT. Here Aleksander
studied the role of feedback in complex systems, a concept which was later
to become central to the learning process in neural networks. But while
Minsky’s views caused the tide of research in the 1970s to flow strongly
towards symbolic processing, Aleksander was one of the few researchers who
held firm to the belief that neural models offered the best approach to
computation and artificial intelligence. He moved to Brunel University,
London, where in 1981 with colleagues Bruce Wilkie and John Stonham he built
Wisard, the first large-scale neural network for commercial use.

In the brain, neurons are connected in complex networks where input
from, say, the nerve cells of the eye, generates patterns of activity in
particular centres of the brain, such as the primary visual cortex. Wisard
had a quarter of a million artificial neurons – simple processing elements
based on memory chips – linked by adjustable connections (see Box in ‘Breaking
Chomsky’s rules’, ¿ìè¶ÌÊÓƵ, 30 January). It demonstrated that neural
networks could be trained relatively quickly to perform certain tasks that
had proved difficult or impossible for symbolic processing systems – in
particular the tasks of recognising shapes and patterns.

Giving machines vision is a task AI researchers seriously underestimated.
Minsky, for instance, is reported to have given the problem of machine
vision to a student at MIT in the 1960s to solve as a summer project. Writing
programs to analyse images captured with cameras, however, has proved enormously
difficult.

Neural networks are not programmed. Instead they ‘learn’ to solve problems
from examples. For instance, if a network is to recognise faces, it is first
presented with a number of examples of faces. Images of the faces are captured
by a camera, converted to digital data and presented individually to the
network as input. The input will trigger patterns of activity, or ‘states’,
in the network, which will result in an output. The connections between
the artificial neurons are adjusted by the learning program so that a particular
input – an individual’s face – always results in the same pattern of activity
and gives the same output.

Wisard caused excitement because, against expectations, it was able
to recognise individual human faces after only 20 seconds’ training. Since
then, Wisard has been put to work in factories, picking out defective components
on production lines, and in banks, where it identifies bank notes at high
speed.

Aleksander is now working with Stonham, Thomas Clarke and Manissa Wilson
at Imperial College and Brunel University on a neural net system called
Magnus that could be used to control a mobile robot. Magnus differs from
Wisard in that it simulates the neurons and their interconnections in software
on an ordinary workstation computer rather than being constructed from thousands
of chips linked together. Aleksander claims that when Magnus is completed
it will demonstrate characteristics, such as learning, language and knowledge
representation, which are normally associated with consciousness.

Magnus operates using what Aleksander calls ‘iconic representations’.
These are patterns of activity in the network which represent objects in
the real world. The iconic representations are the equivalent of mental
images, and will be used to translate instructions given in everyday language
into actions by the robot. Aleksander says that in the first experiments,
Magnus is operating in a restricted world with objects such as a pyramid,
sphere and cube, which it views via a camera. The researchers have already
trained the system to link iconic representations to the patterns of activity,
or states, associated with camera images of the real world objects and
their language symbols. They have also trained it to associate a term such
as ‘left’ with the camera view moving to the left and to create iconic representations
for relationships such as ‘on top of’ or ‘to the right of’.

Mental images for robots

The aim of the experiments is to show how a robot driven by a neural
network could operate by associating language, images and actions. For example,
the sentence ‘Pick up the pyramid and put it on the cube’ will be represented
by a sequence of iconic patterns of activity in the network, and the robot’s
actions could be associated with these. Conversely, when the robot views
a scene, the objects in view and their relationships would trigger iconic
representations which could be used to ‘output’ language that describes
the scene. Iconic representations are a machine equivalent of mental images,
and therefore a step towards machine consciousness, says Aleksander.

He says it will avoid the problems that the symbolic school of AI, promoted
by Minsky, have run into when interpreting instructions for robots to manipulate
objects. The archetype of such systems is SHRDLU (a nonsense word made by
the order of keys on a linotype machine, like QWERTY on a typewriter), a
computer program developed in 1968 by Terry Winograd, another student of
Minsky’s. With SHRDLU, the operator could communicate with the computer
about a simple world comprising a box, a number of blocks and a robot arm.
But before answering questions or performing actions in the blocks world,
SHRDLU had to analyse the typed-in instructions. And interpreting everyday
language requires a knowledge of semantics, a large number of rules of syntax
and a large body of general knowledge. For example, SHRDLU could not interpret
the question ‘How many blocks go on top of one another to make a steeple?’,
which requires an understanding that the phrase ‘go on top of one another’
must not be taken literally.

According to Aleksander, Magnus will avoid these problems by dealing
with language in a way more akin to humans – associating words and phrases
with mental imagery. ‘Language is a representation of the real world. It
squeezes the real world into the things you can say and eventually write,’
he says. ‘AI people have taken these linguistic strings and tried to present
them in a computer in an unambiguous way, and have run into trouble. Very
small changes in a sentence represent things that are completely different
in the real world. They are unable to deal with this.’

What Aleksander has done is different. He is not trying to represent
the linguistic strings but to represent the effects in the real world that
gave rise to those linguistic strings. ‘That makes language understanding
easier as you only need to ensure that the linguistic strings trigger the
right representations,’ he says.

‘Consciousness is very heavily dependent on having a learning system
that can represent the world more or less as it is. AI has always tried
to do the representation at the symbolic level – so it has lost mental imagery
which seems to me to be the crux of consciousness.’

The Magnus project is ambitious. Aleksander has a reputation as a scientific
showman, and is confident his approach will pay off. In discussing consciousness,
Aleksander refers to the definition in Chambers 20th Century English Dictionary:
‘The waking state of the mind; the knowledge the mind has of anything.’
The ‘waking state of the mind’ is a synthesis of many things, says Aleksander
and he postulates a number of attributes – learning, language, planning,
attention and inner perception – that are prerequisite for the existence
of consciousness. It is now possible, he claims, to demonstrate that a general
neural system can possess each of these five attributes.

His first attribute, learning, is fundamental to neural networks. Instead
of being programmed like conventional computers, neural networks learn from
examples. The language and planning attributes derive from the ability
of a neural network to create internal states – that is, particular patterns
of activity – in response to sequences of perceptual input and to generate
appropriate actions based on these states. The creation of internal states
in response to perceptual input is the equivalent of language, suggests
Aleksander. Planning requires the internal state to be split, with part
representing the target state, while the other part learns a sequence of
states that leads to the target state.

Attention, says Aleksander, is the ‘ability to select perceptual input’
or to focus on aspects of the inner state. For example, a human might think
of a cat and then focus on its ears, tail, fur and so on. A neural network
will respond, or ‘attend’ to new input, but if the input is static or a
number of inputs are present, some form of stimulus is required to keep
it attentive. This can be achieved by applying a ‘burst of noise’ – a sudden,
arbitrary but temporary change to the network’s variables. The ‘burst of
noise’ stimulus is so important for attention in neural networks that Aleksander
thinks it would be worth looking for the mechanism in biological systems.

Inner perception Aleksander describes as ‘a sensed inner state’. In
a neural network the ‘inner state’ is the state, or pattern of activity
learnt, in response to an input from the outside world. These patterns of
activity can be stored in a computer memory and retrieved when required.

Aleksander acknowledges that his set of attributes is minimal. But were
an organism to possess all the properties they imply, he says, ‘the organism
would be ready for exposure to critics who have to define what else it should
have in order for them to believe that consciousness had actually been isolated’.

Inner perception is the point where the views of Aleksander and Minsky
begin to overlap. For Minsky, memory of inner states is the key attribute
of consciousness. ‘When somebody says they are conscious, what they are
saying is ‘I remember a little bit about the state of my mind a few moments
ago’,’ says Minsky. Most of the things that are attributed to consciousness
are to do with this short-term memory, he says. ‘If you can’t think about
or reason about what your mind was doing a minute ago then you say ‘That
was unconscious’.’

Minsky’s definition of consciousness is typically simple and clear:
‘Consciousness is being aware of what is happening in the world and in one’s
mind.’ Short-term memory is easy to achieve in a conventional serial computer,
says Minsky. For example, the symbol-manipulating language LISP, the most
popular AI programming language in the US, has a ‘trace’ function. A computer
can be programmed to keep a record of all its internal states and then to
trace back through these. For a human to do the same would require the ability
to go back through brain states to find the point where there was a particular
response to certain stimuli.

‘The human brain has only very limited records of what it has been doing
recently. A machine could be vastly more conscious than a person because
we didn’t evolve for that,’ says Minsky. ‘Now some people will jump up and
be angry and say you mean machines are smarter. No, I’m saying it would
be easy to be extremely conscious – but that doesn’t mean you would know
what to do with it.’

Minsky admits to ‘mocking consciousness’, but this is because he abhors
the mystification that surrounds many of the terms in the AI debate. Consciousness
is ‘one of those words we have for things we don’t understand’, he says.
Because we do not understand these things they are often confused with vital
forces, he says. To use words like ‘consciousness’, ‘mind’ and ‘intelligence’
in an ill-defined way can be a strategy for avoiding thinking about difficult
phenomena. ‘Consciousness is something we (humans) only use a little bit.
We don’t have much of it and we boast about it too much.’

Mischievous humour

Behind the provocation there is a wisp of mischievous humour and an
attempt to force people to think about difficult subjects. Minsky sees himself
as something of a Socrates, challenging received ideas, vague terminology
and ‘bad thinking’.

There is little doubt that the views of Aleksander and Minsky on consciousness
will stir controversy. Consciousness is currently the focus of considerable
philosophical and scientific discussion. Two of the fiercest critics of
AI, philosopher John Searle and mathematician and physicist Roger Penrose,
have drawn a firm line between machines and consciousness.

According to Searle, mental phenomena are caused by neurophysiological
processes in the brain. The brain is not a digital computer but a ‘specific
biological organ’. Consciousness, he argues, ‘is a natural biological phenomenon’.
For Penrose, consciousness ‘is such an important phenomenon that I simply
cannot believe that it is something just ‘accidentally’ conjured up by complicated
³¦´Ç³¾±è³Ü³Ù²¹³Ù¾±´Ç²Ô’.

Aleksander will no doubt be challenged that the Magnus project does
not substantiate his claims regarding machines and consciousness. He will,
for instance, have to demonstrate that the project meets ‘the law of requisite
variety’. This was devised in the 1960s by W. Ross Ashby, a researcher in
cybernetics, a precursor to AI. To paraphrase, the law says that an AI system
must contain variety of information equivalent to the variety of information
in the problem to which it is applied. If it does not it cannot provide
a solution. As language interpretation is key to Aleksander’s views on consciousness
and machines, the language element of the project must demonstrate requisite
variety. This would have to include the ability to apply common-sense understanding
– for example, interpreting how many objects ‘go on top of one another’
in a nonliteral way.

Common sense is a challenge for Minksy, too. A machine could have a
high level of consciousness but not use it in any meaningful way, he says.
Consciousness is not the issue in the quest for intelligent machines: the
need is to endow them with common sense. Even the most sophisticated of
present-day robots cannot perform a mundane task like cleaning a house.
To do so a robot would require a vast amount of everyday knowledge ranging
from the force of gravity to the brittleness of glass – so that it would
know that a glass that is dropped will break.

AI systems are currently restricted by the lack of this kind of general
knowledge, and Minsky believes that ‘every country should have a programme
to make a common-sense computer’. As a model he cites the Cyc project initiated
in the mid-1980s by Douglas Lenat at the Microelectronics and Computer Technology
Corporation in Austin, Texas. Cyc will be a mammoth accumulation of common-sense
knowledge that an intelligent machine might require to perform everyday
tasks and interpret everyday situations. Lenat aims to have 10 million interconnected
facts keyed into the computer database by 1995. Cyc will one day be able
to offer its knowledge to other specialised computer programs, Lenat believes.

Minsky has also modified his views on neural networks since his devastating
critique in 1969. More sophisticated neural network architectures than the
perceptron model plus more advanced software methods have overcome the
limitations that he and Papert outlined. Also neurological research has
greatly increased our knowledge about the brain. The discovery of many different
specialised centres in the brain led Minsky to formulate a theory that the
mind is a ‘society’ of interlinked and cooperating parts, a theory he explores
in his recent novel, The Turing Option (Viking), written with Harry Harrison.
He speculates that a thinking machine of the future ‘might look like the
brain with hundreds of neural nets’.

Aleksander and Minsky are no strangers to controversy. Their views are
often provocative and Minsky at least has a reputation for thought-provoking
speculation while leaving others to work out the details. But both men are
continuing to push the frontiers of artificial intelligence, and they have
made it clear that this frontier now extends to consciousness. The debate
and the research have a long way to go.

Clive Davidson is a freelance journalist specialising in computing.