午夜福利1000集合

DAVID Noever has a dream. One night, his computer wakes him up with a loud,
insistent beep. His own voice, pre-recorded in a moment of optimism, shouts
excitedly: 鈥淒avid, this is the one we鈥檝e been looking for. This is it! We鈥檝e
found life on Mars!鈥� He stumbles out of bed, pulling on his clothes, and goes to
the source of the voice. There on the computer screen will be Noever鈥檚 鈥淜odak
moment鈥濃�攁 digitised picture of a slice of Martian rock. And there, plain
for all to see, is the unmistakable form of a fossilised cell.

This isn鈥檛 as far-fetched as it may sound. Noever is the leader of a
controversial new project that will create an intelligent software
system鈥攁 silicon lab assistant, if you like. Running on thousands of
computers, networked across the Internet, Noever鈥檚 carefully tutored program
will spend its time combing through billions of images鈥攑hotos of
cross-sections through meteorites鈥攕earching for the signature of life.

Noever鈥檚 venture will have another goal too鈥攖o produce a 鈥淏ook of
Life鈥�. Eventually, he hopes, this will form the ultimate reference guide for
biologists. Every living creature known to humankind will be there, sorted and
arranged in a completely new way. The project hopes to answer a fundamental
question that has bothered earthbound biologists for decades. Are there specific
signatures鈥攕ymmetry, shape or a set of combined features鈥攖hat
provide the exclusive key to identifying and categorising living organisms? In
other words, is it possible to define the shape of life?

It鈥檚 an ambitious project, and to get it under way, Noever has had to think
small鈥攎icroscopically small. He has been teaching his software to decide
whether a tiny speck inside a meteorite is a fossilised Martian or simply a
nanometre-sized blob of minerals.

This is no mean feat鈥攅ven human experts can鈥檛 seem to agree what鈥檚
what. In 1996, NASA announced the discovery of what seemed to be microscopic
fossils of primitive, bacterium-like organisms inside an ancient Martian rock
found in Antarctica
(New Scientist, 17 August 1996, p 4). The argument
over whether these tiny structures actually are fossils is still raging.
Meanwhile, researchers are again heading out to Antarctica to scour the ice for
more meteorites. If past expeditions are anything to go by, they could have as
many as 1000 new meteorites in a few years鈥� time.

This volume of material will stretch NASA鈥檚 analytic resources to the limit.
Any record of Martian life is likely to be well below the micrometre
scale鈥攕omewhere around a thousand fossils might fit across the width of a
human hair鈥攕o when the rocks are sliced up, about several trillion
candidate photographs will need examination.

Noever鈥檚 solution to this vast problem is a special type of program called a
neural network that learns from experience
(New Scientist, 20 January 1996, p 22).
To teach his neural network what to look for, Noever shows it
images of microbes. The program then extracts the essential features from each
picture鈥攖hings like colour, size or structure, for example鈥攁苍诲
stores them away. Then he might show the program a picture of a small pebble,
and tell it this isn鈥檛 what to look for. Again, the program extracts the
characteristics of its shape, and stores them under 鈥渄on鈥檛 get excited about
迟丑颈蝉鈥�.

Identity parade

The computer simplifies each image for analysis, in much the same way that
computer scanners reduce text characters to the simplest skeleton in order to
identify them. So, to Noever鈥檚 program, a rod-shaped cell becomes an elongated
ring of dark pixels; the perimeter of a spherical cell is marked out by a black
ring of pixels on a white background, and so on
(see Diagram). To start with,
the program breaks each image down into small squares just eight pixels across.
Then the neural network compares the position and state鈥攚hite or
black鈥攐f each of the 64 pixels in the image with the rest, so that it can
work out what shape it is dealing with. Once it has done this for each segment
of the picture and put them all together, it compares the image鈥檚 scale and
features with the images it has learned about so far. If there is a match with a
positive image, it classifies the new picture as potentially interesting.

The great thing about a neural net is that it can constantly adjust the
importance it attaches to each feature of the image. As the number of images fed
into the program grows, so does the computer鈥檚 ability to make up its own mind
about whether a picture belongs among the hopefuls or on the scrap heap.

But Noever has added another refinement. When he left NASA two years ago, he
co-founded Cyberchemics, a technology company based in Huntsville, Alabama, that
develops special software designed to help drugs designers create new
pharmaceutical products. Drawing on this experience, he has included a genetic
algorithm in his program. This helps the net evolve rapidly, introducing large
changes into its architecture in response to what it learns.

Noever has named his neural network 鈥渢he D鈥橝rcy machine鈥� after D鈥橝rcy
Thompson, a biologist who worked at the University of St Andrews in Scotland in
the early 1900s. Thompson was one of the world鈥檚 first biomathematicians, and
his mathematics provides the foundation for Noever鈥檚 program
(see 鈥淭he morph man鈥�). 鈥淲e鈥檙e building on D鈥橝rcy鈥檚 legacy, looking at how important shape
is in identifying life,鈥� Noever says.

Borrowed time

The effort of analysing billions of images will require fantastic computing
power. So Noever hopes to borrow your home or office computer. He plans to break
the images down into 50-kilobyte packages, and distribute them across the
Internet to computers that are taking part in the project. Each processor then
performs the operations, analysing the pixels within its package as a background
to your computer鈥檚 normal activity鈥斆� la SETI&home
(New Scientist, 25 July 1998, p 46)鈥攁苍诲
extracting the characteristic
features. These are sent back to the neural network for comparison with the
familiar, learned shapes that signify life.

Unfortunately, there are some drawbacks. Not least of which is the fact that
Noever can鈥檛 be sure that he knows what extraterrestrial life will look like.
Astrobiologists aren鈥檛 looking for fossils of little green men. If they do find
anything, they believe it will be a single-celled organism, a primitive life
form, maybe a hint of chemistry gone live. But the only real reference point is
the nature of such things on Earth.

Earthbound microbes come in a variety of shapes and sizes. There are rods,
spheres, filaments, spirals and clusters鈥攌nown as cocci鈥攖hat look
like grapes. So to begin with, Noever is training his program to look for these
simple shapes.

What if this technique actually inhibits the computer鈥檚 ability to recognise
Martian life? By training it to look for known shapes, could Noever be skewing
the system away from what living things might look like on Mars?

Not so, says Simon Conway Morris, a palaeobiologist at Cambridge University.
He thinks that looking at the shape of life on Earth is as good a place to start
as any. Our planet, he says, has had four billion years of history, a vast
diversity of environments and major fluctuations in atmospheric composition. As
such, most of the possible forms of extraterrestrial life should be found
somewhere on Earth.

All life is here

He believes that there are certain shapes and functions that will always
arise in evolution, a process known as 鈥渃onvergence鈥�
(New Scientist, 13 February 1999, p 28).
Chemistry, he thinks, will almost inevitably lead to the
genetics of the primitive cell. 鈥淚 find it difficult to think there might be an
alternative to DNA,鈥� he says. And if amino acids form, then you鈥檙e going to get
proteins, and protein folding, 鈥渁nd if you鈥檝e got protein folding it鈥檚 difficult
to imagine it without a cell.鈥� In almost any environment, says Conway Morris,
chemistry will create life looking pretty much as we would recognise it. At the
microbial scale, we may find there is nothing new under the Sun.

Even if the atlas of life on Earth makes a less than ideal training set,
Noever鈥檚 machine could fall into even bigger traps. Consider the impact of the
wrong choice of negative images in training the program. Many man-made
structures imitate objects in nature; show the program pictures of these
鈥渘ot-life鈥� structures and some serious confusion could follow. Imagine the
embarrassment if the computer ignored a fossilised Martian beetle because it had
been taught that the shape of a Volkswagen Beetle is nothing to be interested
in. Or if the program saw a fossilised fly鈥檚 eye and threw it out, believing it
to be an architectural geodesic dome.

There are so many possibilities that 鈥渘ot life鈥� becomes almost impossible to
define. 鈥淚t鈥檚 a bit of a black art,鈥� Noever says. 鈥淵ou can鈥檛 just tell it that
the car or the chair is a shape we don鈥檛 normally associate with living
things鈥攜ou need to be as representative as you can.鈥�

Such value-weighted judgements鈥攅asy for human eyes鈥攁re among the
biggest challenges facing artificial intelligence researchers. 鈥淭he question is,
are there ways to automate what humans see?鈥� Noever asks. 鈥淲hen I see a horse in
the clouds, how do I know when it is a horse, and when it is a cloud?鈥� It may
sound like a ridiculous, Zen-like inquiry, but that鈥檚 only because our minds
make unconscious judgements that answer the question before it gets asked. If
you鈥檙e looking at the sky, you know from experience that you are more likely to
see clouds, not horses. We can define context, texture, colour and scale without
giving it a thought. Whether a computer can really do the same is one of the
questions that makes artificial intelligence so controversial. 鈥淵ou could
probably write down a hundred criteria to separate those images,鈥� Noever says,
鈥渂ut you could say it鈥檚 actually an infinite list.鈥� If a program shakes it down
to a certain number of criteria, it is possible that it loses the real defining
argument. So, the critics say, artificial intelligence is always going to be a
miserable reflection of the capabilities of real intelligence.

In a way, Noever agrees. It鈥檚 impossible, he says, to write a straightforward
program that will tell a computer how to tie shoelaces, for example. There are
too many variables, too many possible snags. But it is possible to teach the
computer鈥攋ust as we learn鈥攖hrough repeated examples. Hence the need
for a neural network, trained like a difficult pet. 鈥淵ou allow it to fail, tell
it where to go, and let it iterate that with a heavy punishment and reward
system,鈥� Noever says.

So far, his main training strategy is to use multiple viewing angles on every
object. Find as many ways of showing the fly鈥檚 eye and the geodesic dome as
possible, he says, and the computer will tell them apart. Such a strategy will
also guard against dumb mistakes: the program can鈥檛 be allowed to classify an
鈥渆nd-on鈥� view of a rod, for example, as a sphere. 鈥淎ll our energy goes into
selecting candidate images: we are giving it images which show all of the scale
features we can think of,鈥� he says. It certainly seems to work鈥攖he neural
network does now get things right most of the time. 鈥淭he program is accurate. We
benchmark it at about 80 to 90 per cent,鈥� Noever says.

To improve this figure further, another approach could be to identify the
properties of the material under inspection, suggests Noever. It might be
possible to use factors like context and reflectivity, for example, to measure
the differences between the way a fly鈥檚 eye and a geodesic dome reflect
light.

Nothing, however, would be as useful as finding fossilised signs of
movement鈥攃atching cells in the act of division, for example. 鈥淒uring cell
division you have a period of regulated shape change, and those shapes have been
measured very accurately,鈥� says Wolfgang Alt of Bonn University. Put those
shapes into the database and Noever could stand a much better chance of
capturing his 鈥淜odak moment鈥�.

Alt specialises in examining and mapping cell movement, and has watched some
extraordinary examples of cell motion. He has found they can take on unexpected,
mathematical shapes as they move. Sometimes a square-shaped microbe will undergo
periodic pulsations, he says. 鈥淭hey roll up, and then take on a quadrilateral
shape.鈥� Skin cells, for instance, can become triangular or square, sometimes
even pentagonal鈥攏ot the blobby shape we normally associate with primitive
life forms.

Amazing shapes

Add to that the fact that as they move, rod-shaped cells can flex, round ones
can become almost semicircular, and it becomes clear that life could actually be
caught in an amazing array of forms. Alt thinks that Noever鈥檚 program would do
well to look for these kinds of signatures too. 鈥淚t would be good to set up a
project which recognises these typical shapes from movement,鈥� he says.

This isn鈥檛 as easy as it sounds. For starters, forces like surface tension
might act on Martian cells in a different way from Earthbound ones. And Noever
thinks that regular shapes would probably be thrown out straight away by his
computer because the program has been trained to reject geometric
objects鈥攖hey often signify the presence of crystalline minerals in the
rock. It鈥檚 contentious, he admits: some viruses are able to crystallise
themselves and would also be overlooked. But there has to be some baseline for
rejection. 鈥淲e鈥檒l be as comprehensive as we can but we鈥檙e willing to live with
10 to 20 per cent false positive and false negative,鈥� he says.

Ways of seeing

Despite all the hurdles, Kanti Mardia, who specialises in computerised shape
analysis at the University of Leeds, believes the project could work. 鈥淚t鈥檚
certainly feasible to identify structures automatically, although it鈥檚 very
computer-intensive,鈥� he says. And if you know what you鈥檙e looking for, it鈥檚 not
always necessary to go down to the level of individual pixels. Looking at
slightly larger features within the picture can save a lot of time and effort.
But problems could come, he says, if the images are not clear enough. 鈥淚f you
have very low contrast between the object and its background, that can make life
very difficult.鈥�

Norman MacLeod of London鈥檚 Natural History Museum agrees that the image is
critical. 鈥淧art of the object might be obscured, it might be in a different
orientation, or the lighting might play havoc with the boundary of the object,鈥�
he says. 鈥淐reating an automated system鈥攐ne that scans for things that even
experts disagree about鈥攊s somewhat optimistic.鈥�

Noever is aware of the difficulties. Initially, he plans to use his D鈥橝rcy
machine as a 鈥渇irst-level filter鈥� to dispose of the rocks that hold absolutely
nothing of interest. After that, images can be passed to the experts for debate.
He will be happy, he says, if just a few of the project鈥檚 challenges are
met.

Although Noever will begin by looking only at microbial life forms, there is
no reason why his neural net can鈥檛 eventually begin to learn the shapes of
higher life forms, to create the 鈥淏ook of Life鈥濃�攁 taxonomy categorising
the characteristic shape of every living thing on our planet. Whatever happens,
the project is bound to stretch the boundaries of artificial intelligence. And
there will be gains in image processing, not to mention the practice of
distributed computing.

Eventually, in a decade or so, Noever says he would like to see the program
used aboard one of NASA鈥檚 space probes. 鈥淚t鈥檚 very attractive to imagine
integrating it into remote sensing鈥攁 diagnostic tool in interplanetary
别虫辫濒辞谤补迟颈辞苍.鈥�

鈥淚 think he鈥檚 right,鈥� says David Lehman, project manager for Deep Space 1,
one of NASA鈥檚 first 鈥渟mart鈥� probes
(New Scientist, 24 October 1998, p 38).
An intelligent recognition system such as Noever鈥檚 would certainly fit with
NASA鈥檚 objectives for future missions, he says. Processing data in space is
becoming more and more important as the tasks carried out by the craft get more
complicated. Imagine, for instance, that you want your probe to trundle its way
across the Martian surface looking for fossils鈥攋ust think how much easier
things might be if it were able to process data on-board and make its own
decisions.

鈥淯sually when you command a spacecraft you have to tell it exactly what to do
and when to do it,鈥� says Lehman. It can take 20 minutes for a signal simply to
make the journey from Mars to Earth and back again. It would be much better,
says Lehman, if the craft could react automatically to the environment around
it.

But Chris McKay, an astrobiologist at NASA Ames Research Center near San
Francisco, believes it鈥檚 too soon to think about using intelligent probes to
hunt for fossils. At the moment, it makes more sense just to look for liquid
water and organic molecules. 鈥淭hese are the key correlates of life,鈥� he
says.

Helen Matsos, an Ames biophysicist and collaborator on Noever鈥檚 project,
argues that shape is as good a place to start as any. 鈥淵ou have to have a point
of departure,鈥� she says. Whether the shape of life on Mars is similar to life鈥檚
forms on Earth will, of course, affect the machine鈥檚 usefulness, she admits.
鈥淏ut you have to start from somewhere, and our view of life is the only thing we
can go with at the moment.鈥�

Besides, says Matsos, NASA currently employs just a dozen or so people to
look for fossils in Martian meteorites鈥攁苍诲 their only reference point is
Earthly life. Matsos and Noever are convinced that their D鈥橝rcy machine could
help speed up NASA鈥檚 search.

At this very moment, somewhere inside a rock tucked away in a NASA lab, a
strange fossilised Martian bug may be awaiting discovery. Similar rocks might be
picked up by the next Antarctic expedition. Once the analysis starts in earnest,
the world will watch with bated breath. Most enticing of all, the first
indication that there was once life on Mars could come from your PC.

D鈥橝rcy Thompson鈥檚 seminal work On Growth and Formused mathematical
functions to describe the shapes of living things, and then altered the maths to
show that species changed in shape over time according to strict mathematical
principles.

He discovered that superimposing a rectangular grid on a picture of a fish,
for example, gave him a way to relate the fish鈥檚 features to a set of
coordinates. Altering the grid with a simple linear change in the parameters so
that it took the shape of a parallelogram allowed him to 鈥渕orph鈥� the fish鈥檚
features: from one species of deep-sea hatchet fish to another, for example. On
the other hand, a quadratic change that makes the grid curved can round out a
parrotfish into something exactly like an angelfish. The same effect occurs with
skulls鈥攓uadratic mapping on a human skull can make it look like a
chimpanzee鈥檚 or a baboon鈥檚, depending on the mathematical equation used to morph
the grid
(see Diagram).