|
|
Computing like the Brain:
Advances in Neurocomputation
Igor Aleksander
Professor of Neural Systems Engineering
Imperial College, London
The myth of the 'Electronic Brain'
 'The Electronic Brain' is a cliché that has become attached to computers
since their inception. But nothing could be further from the truth: there is
very little in common between the way that computers and brains work. The fact
is that for some tasks such as calculating, sorting, filing and even playing
chess, computers can do better than brains. But in other tasks, such as the
recognition and understanding of the world and its inhabitants, computers lag a
long way behind. Were we to land a robot on a distant planet to have the
sensibilities of an astronaut, programming it to survive would be an awesome if
not impossible task. So what is it about the mind of a live astronaut that is so
much better than a computer?
For some years now, we have been attempting to understand the computational
principles of the living brain. Two reasons compel us do this. First we wish to
have a better understanding about how the brain achieves its competence, and,
second we want to find ways of transferring this competence to computers and
robots, over and above that which can be achieved with current computational
methods. This science is called neurocomputation.
What is Artificial Intelligence?
Is this what people call Artificial Intelligence? Well, maybe. I would call
it 'new' artificial intelligence where GOFAI (Good Old Fashioned Artificial
Intelligence) has been with us for over 50 years. Let's start with GOFAI to see
the contrast between the old and the new. This began in the 1950s when computer
engineers recognised that the computing machine that had recently made its
appearance (due to John Von Neumann's design in 1947) could do more than crunch
numbers. It was a genius of a US engineer called Claude Shannon who, in 1950,
wrote the first program to play chess (he had also invented 'information theory'
and was the first to apply logic to the design of switching systems). He devised
an ingenious way of rushing efficiently through the possible moves that both the
machine and its opponent could take in order to chose the next best move. This
searching through vast numbers of possibilities using the high and continually
increasing speed of computers became the style that made the computer seem
intelligent. This method was at work when, in 1997, an IBM computer called Deep
Blue beat the world chess master, Garry Kasparov. But the rules for survival on
Mars might much be less tractable and clear than the rules needed for playing a
game of chess.
Why is the brain different?
So what are the differences between the way the computer computes and the
brain produces 'thought'? Primarily, the brain is made of 100 billion little
cells (neurons) each of which is like the memory of which the computer has just
one. It is true that each cell is less powerful, than the Random Access Memory
of my laptop, so why are so many little memories so effective? The key answer is
that the electrical activity at the outputs of these cells is responsible for
everything of which we are capable of thinking . They are the components that
give us a knowledge of what the world out there is like. Were I to look at a
white wall when a little fly comes to rest on it, that fly causes some of my
neurons to change their activity (they are said to 'fire' when they produce
electrical impulses at about 100 times per second). These neurons not only
encode the colour and shape of the fly but also where, out there in the world,
the fly is. This is what makes us 'conscious' of what the world is like.
Not only are neurons active when we perceive something with our senses, but
also, because they communicate among themselves they can recall their previous
perceptual sensations. This is what we call our imagination or our experience.
These internal patterns of activity can be triggered by incoming impulses
generated by the senses. For example if we read a book or hear a story, we are
soon engulfed by our neurons making us experience the content of the tale. The
internal patterns also respond to our intentions and desires. If I am hungry or
thirsty they will allow me to work out not only how to plan to get food but will
also create my feelings that anticipate the pleasures of eating. At other times
they could spell out warning signs which we would describe as fears.
Simulating conscious brains on computers
The focus of our work is to see whether this firing behaviour of neurons
could be simulated on a computer to create mechanisms that include a simple form
of consciousness. It is all possible in principle, what stops us from simulating
a brain with a human-like mind is the sheer scale of things in the brain. The
100 billion cells form something like 100 modules each with an individual task
to perform. They are interconnected to perfection by evolution. Interestingly,
evolution is something one can simulate on a computer, but not on the vast scale
that gave rise to the human brain. So far we have been able to study systems of
about 50 modules with a total of 2-3 million neurons which provide some visual
awareness, visual imagination and planning. What do we do with it?
The breakthrough with these simulations is that we can decode on the computer
screen what the machine is 'thinking'. We can see if the machine is thinking
normally or defectively. So far this has not been possible with the brain. Yes,
medical brain scans tell us where in the brain the activity is going on, but not
what the subject of the activity is. But because we know the way that the
simulation is connected, this revealing decoding becomes possible. Through this
we have been able to show that the interconnections between modules which
sustain both perception and imagination are very subtle. Get the balance between
these interconnections wrong and the machine stops perceiving or imagining. This
has led us to investigate what goes wrong in some mental diseases such a
Parkinson's where patients may loose some visual awareness and, on occasions,
hallucinate. The practical side? We have worked on providing visual awareness
for robots which enables them to build up accessible experience of their
surroundings (rather than be told by a programmer what they are) and then
respond to simple verbal statements such as 'go to the red cone'.
Can a conscious machine be dangerous?
Where is all this going? Some may feel threatened by robots becoming
conscious in some way. Will the machine become evil and want to get rid of
people? For me, this is science fiction: there is no reason why this greater
ability should lead to a self-generated evil streak. Evil is a product of
competition within and between animal species for territory, food and mates.
Machines don't need these things. Machines that are aware will be safer and more
helpful to humans than machines that rely entirely on the prowess of a human
computer programmer and, goodness knows, how many mistakes a program can hold!
Were a robot to drive your car, would it be better for it to be conscious or
unconscious? So, if the work discussed in this paper continues to be successful,
in twenty years' time it will make obsolete computers that are totally devoid of
brain-like machinery. Also it will give us a new neurocomputational
understanding of the biological brain to complement our current neurological
knowledge.
Professor Igor Aleksander may be contacted at i.aleksander@ic.ac.uk
|
