about the coming computer revolution: AI and evolutionary algorithms
brought together in the development of autonomous computers.
There is a recent interview as well, in the Nouvel Obs. Translation to come.
* * *
source: Le Nouvel Observateur
interviewer: Sophie Fay
translation: doxa-louise
‘Computers are capable of darwinian evolution’
Pierre Collet has developed in his laboratory ‘evolutionary motors’
that make possible for computers to create industrial objects of
higher performance value than those made by engineers. Interview.
The Strasbourg Complex Systems Digital Campus at Strasbourg University,
directed by Pierre Collet, is one of the French Centers of Excellence for
research on Artificial Intelligence. Interview with a researcher on a fascinating
subject.
You teach Artificial Intelligence (AI). What is this about?
There are many forms of Artificial Intelligence (AI). In general, one
defines AI with the use of the Turing Test: when a man interacts with a computer and
cannot tell with certainty whether he is dealing with a human or a machine, the
program is considered intelligent.
This is the ‘chatbots’ principle, dialogue applications. But it is also possible
to make Artificial Intelligence that is non-human. This is what we are up to at
Strasbourg: artificial evolution. That mechanism which makes for nature finding
solutions to adapt to its environment and survive, we use that on computers
who in this way become creative.
So this is evolution, in the Darwinian sense?
Precisely. One confronts the machine with a problem, which finds solutions, cross-
breeds them, discards the less promising, selects the best. With spectacular results.
In 2006, NASA created a micro-satellite the size of a basketball. It needed
an emission mechanism that would consume little energy, to service a given
radio telescope.
No one could solve it, the antennae designed by engineers were all ten meters long.
Unusable. NASA asked computer scientists to have a look and apply their programs.
They came up with a mini-antenna, and a physical concept that no one could have
imagined and which no one has reproduced since. To find another antenna for another satellite, one has to resubmit the problem to machines along with the new parameters,
and make new plans. It’s all possible, but no one is quite sure how.
Thus, in 2206, NASA sent the first bit of non-human technology into space.
Does this mean computers now possess a creative form of intelligence?
For decades people thought that computers were super calculators, doing what they
were asked to do, capable of resolving equations more quickly but that was all.
Yet beginning in the aughts, computers have become creative. They put forward
things humans never suggested to them. Sometimes we are dealing with completely
novel solutions, which we cannot even control. To get an additional solution to a
similar problem with different parameters, one has to ask again. Now computers can
image things on their own.
Computers have become creative and competitive with humans.
But such computers do not pass the Turing Test (wherein a human cannot tell the
difference between an answer given by a computer or one of his peers), because
their form of reasoning is non-human.
Has Strasbourg come up with an analogue to the NASA antenna?
Yes, we have for example found a new zeolite thanks to Artificial Intelligence:
it is a porous crystalline structure. About one hundred variants are currently known
in the world. Each zeolite has applications with respect to filtration. They can be
used for cat litter, for example.
How have other forms of AI evolved?
Artificial intelligence that is historically embedded, using what is called deep l
earning and neural networks, gives computers a reasonable understanding of their
environment: if you show a picture of where we are presently to a computer, it will
respond: you are in a bar. Eventually, it will claim more precisely that we are in
Alsace, in Strasbourg, because it recognizes this pan for kuglof and certain elements
of decoration etc...
But it is not yet autonomous. If it detects a problem, it will not know to organize
itself to solve it, unless it has been programmed to do so. For that to happen, we would
have to link deep learning and artificial evolution. This is what we are working on in my
laboratory: linking up comprehension of the environment and evolutionary motors.
So what happens when you succeed in creating such a link?
Then we will have autonomous robots: once they have identified a problem and
created a model, they will switch to AI mode which will find a novel solution.
What are evolutionary motors called?
We are simulating Darwinian evolution with a computer. It’s a PC. We have in our
lab a supercomputer, with an enormous capacity for computation. and this enormous
capacity will generate millions of possible solutions which will evolve over generation
upon generations as with Darwinian generations. This involves mutations, cross-breeding.
Artificial evolution is a complex system, which is faster than random searches. It is a
highly productive creative process.
How do you know the machine has found something?
We incorporate an evaluation function, which evaluates the quality of each new
individual created. As evolution proceeds, the best individual of his generation -
as a function of its value in resolving the given problem - is selected. After many
iterations of cross-breeding and selection, we end up with an individual that does
a good job of resolving the problem.
Are there already industrial applications for this?
Yes, a great deal of work has been done on the design of airplane wings. There
is no other way to work on this than through trial and error. Engineers design forms,
test them, cross-breed them, all intuitively. The computer, for its part, starts with a
population of airplane wings and makes them evolve until it finds the best design.
Evolution proceeds by trial and error, but not in a random way, but by taking objects
with worked rather well. The yield on this is the creation of new forms which no one
had imagined.
How long did it take to find your zeolite, for example?
We use machines in parallel. In this case, it would have taken a PC 11 years of
continuous calculation, we used many machines in parallel and it took a mere 34
hours. This is an important research node. Since 2005, all computers are in parallel
mode. When you buy a PC there are many cores and there are two sensors. The
graphics card can have up to 3 000 cores. Your telephone has two cores in the processor.
These computers are linked and one can make thousands of cores work in parallel. To
do that one uses algorithms inspired by nature because in an inherent and intrinsic way,
nature works in parallel.
Not everyone does the same thing at the same time. Everything gets done in an
asynchronous fashion. This is the essence of complex systems, which allows us to
direct interactions between a great number of autonomous entities, to obtain a desired
result.
Aristotle had foreseen the power of this interaction, when he said that the whole is
more than the sum of its parts.
The 100 billion neurons in our brains, so many autonomous entities, interact to produce
a decision in one thousanth of a second. That is cognition.
Certain researchers have heralded that there is exponential progress in AI?
I would be prepared to say there are thresholds. Starting in the aughts, machines
became creative, repetitively so. There is a book by John Koza which shows that, starting
in 2003, with respect to many quite different problems, computers were regularly generating solutions competitive with those proposed by humans. These
solutions can be better than those hit upon by engineers.
In the interval 2010-2012, another threshold was crossed, that of deep learning:
computers were able to recognize cats in an image, than everything else in that image.
The next stage, the one we are working on, is to unify the two to yield autonomous
computers.
Is this exponential? All evolution is exponential!
You are a signator to the principles of Asilomar, which is a kind of ethics document for researchers in Artificial Intelligence, co-produced last February.
Please explain.
Thins are accelerating quickly. In the near future, computers will become autonomous.
Once this happens, we will be able to task them and in an autonomous fashion, they will
find what needs to be done to accomplish the task. They will have a goal, and they will
accomplish it.
This can become dangerous in a number of ways. Even if given a ‘sweet’ goal, the
computer might well imagine actions that could hurt people. The real problem is that
we might build armed robots.
Currently in China, there are in transport hubs robots armed with tasers
(The AnBot robot). That raises a number of questions. It can aim for the wrong target. There are still many errors occurring in recognition. Computers make mistakes.
(The AnBot robot). That raises a number of questions. It can aim for the wrong target. There are still many errors occurring in recognition. Computers make mistakes.
At the moment, there are not many autonomous things that make decisions and can
interact with humans, so we are not aware of the dangers. But we must prepare ourselves, put up barriers, agree on a code of behavior. Decide, for example, that
we will not arm them. We have so many more useful things to aim for: bettering
education, medication, care...
What will be the first autonomous decision in your estimation?
In fact, it is already here. You bank account for example is managed by a computer.
It will tell the bank whether we should loan you money or not. Banks go through
your payments. If you buy gas, they know whether you own a small or large car.
Everything you do is gone through to decide if you are ‘bankable’.
Already computers are very autonomous, but there is often an element of human
control. And we have not yet integrated these to robots. When drones take off
to kill the Taliban, today it is with planes direct by humans. Tomorrow we will
be able to assign the task to a computer. do we wish to move in this direction?
No, this is what the principles of Asilomar specify.
No comments:
Post a Comment