Do brains compute? : Rodolphe Sepulchre at TEDxLiege

Translator: Virginie Labeye
Reviewer: Denise RQ How would you feel in a world where half of the doctors, lawyers,
and counselors would be machines? How would you feel in a world
where half of the caretakers of your elderly people
or young children would be robots? How would you feel in a world
where ‘Her’ would be a computer? Those questions are not mine,
they surround us, they fascinate us, and sometimes they scare us because those are questions
about ourselves, those are questions
about how different we are and how different we feel
from the machines that we build. Those are questions about a race, a race between artificial intelligence
and human intelligence. As a control scientist, I’m studying the control mechanisms
that enable intelligence both in machines and in nature. And today, I would like to share with you a control mechanism that is different
in the brain and in the computer. This difference is so simple that it’s almost embarrassing
to talk about it. And yet, I think it tells us a great deal about how differently we compute
from the machines that we build. But let’s start from the beginning. The race between the computer
and the brain started in 1950 in Cambridge when Alan Turing, one of the greatest scientists
of the last century, invented a game: the imitation game. A modern version
of the game is as follows: you receive an invitation on Facebook, but you don’t know whether your new friend
is a machine or a human, so you engage in a conversation with her, trying to figure out
whether she has a brain. And if you do figure out, you win,
if you don’t, you loose. Turing predicted that by the year 2000, machines would win
the imitation game most of the time. Turing’s paper set a fundamental link
between intelligence and computation, and the imitation game launched one of the richest dialogues
of contemporary science. For the last 60 years, the brain
has been a model of the computer, and the computer has been
a model of the brain. And here we are, 60 years later. Our pocket machines solve
our daily calculations, correct our spelling mistakes, complete
our sentences, translate our texts, solve our quizzes,
teach us how to play chess. Machines solve
those complicated computations by decomposing them into a sequence
of very simple operations. So they need memory
to store the intermediate steps, and they need processors
to execute the operations. Computer technology has developed steadily
to a point where those pocket machines can store huge amounts of memory
and execute operations at a speed that literally,
approaches the speed of light. This makes computers incredibly fast at solving computations
that would be formidable for our brains. So the victory of computer technology
is really humbling for the brain, to the point where you wonder
whether there’s still anything our brains can do better than computers
when it comes to computing. Well, I think there is at least one thing our brains still do
much better than computers. We forget what we memorize,
we disregard what we see, and often we don’t listen
to what we hear. That might have been a problem at school. Perhaps you consider this
as a further weakness of your brain compared to the computer because computers do not forget
what they memorize, machines do not disregard what they see, but forgetting and disregarding are the basis of discerning. Our brains keep discerning. Machines are quite bad at discerning, and this makes our brain fast,
in fact much faster than computers at zooming in and zooming out
on the relevant information. A few weeks ago, I was hiking with my wife
in a part of Turkey called Lycia, in the Mediterranean Sea. One afternoon, we decided to go down
500 meters below our hiking path to explore a beautiful little beach. It was already 2 p.m.,
so we were a bit in a hurry, and as we were running
down this very rocky terrain and looking for these small yellow
and red signs that would lead us to the beach, I kept thinking how challenging
it would be for a machine to do what was great fun for us because we are quite good
at detecting those small signs, just because they matter to us. But then, relaxing on the beach
and recovering from the descent, I remember all these awareness experiments
designed by neuroscientists, I remember missing a gorilla
in the team of basket players just because my mind
was focusing on the ball passes. This is amazing.
I think it will take a lot of time before computers can miss a gorilla
among a team of basket players. This ability of detecting
very small things or ignoring very big things
depending on our state of mind, this ability of focalizing
our computations in a very specific window
in time and space, then modulating this window very quickly, I would call ‘plastic localization’
or ‘multiresolution.’ Brains are superfast at executing,
but they are rather slow at localizing, because they are rigid,
they are monoresolution. Brains instead are rather
slow at executing, but they are incredibly fast at localizing
and at modulating the resolution. They are plastic,
they are multiresolution. But why is that? Why can’t we
make machines plastically localized? To me, localization has not much to do
with how fast you compute or how much you can store, it has to do with a control mechanism, which brings me
to the second part of my talk. Two years before Turing’s paper, Norbert Wiener,
a respected control engineer at the MIT, published a book
where he mostly emphasized the fundamental role
of feedback to intelligence. A feedback processor is a processor
that feeds the output back at the input. This processor does not process
the input signal you intend to process. Instead, it processes a comparison between that intended signal and the one
that is actually seen at the output. You are all very familiar with feedback because we all know
how important it is to learning. How could we learn without errors? How could we form errors without comparing
intended actions to actual actions? But there are two types of feedback:
negative feedback and positive feedback. Negative feedback is
how your thermostat works. The input to your heating system
is not the desired temperature, it is a comparison between the desired temperature
and the actual temperature. The processor tries to make
the input and output alike, so it is a regulating device. We are surrounded by negative feedback. Negative feedback is what regulates
the temperature in this room, the temperature of your body, the speed of your car,
the speed of your heart. Positive feedback is the opposite
of negative feedback. Here, you add the output to the input. So if, there is a change of the input,
your further amplify the change. As a result of this instability, you get an all-or-none device,
you get a switching device. Positive feedback is everywhere around us. Positive feedback switches off and on
the light in your living room, but also the genes in your cells,
and the transistors in your computer. So plenty of negative feedback,
plenty of positive feedback. Now, let’s imagine a device
that would have two feedback loops, one for positive,
the other one for negative, a mix of switching and regulation. If you ever raise a child or a pet, you might think that this is not
so bad as a policy. Positive feedback is a switch
between reward and punishment, something stimulating, which you’d like to balance
with a regulation of emotions, which is negative feedback. So it seems that intelligence
promotes the mixed feedback policy. But an engineer would probably tell you
that this device is not optimal because every feedback loop has a cost. So you might convince an engineer
that you need one feedback loop, either for regulating or for switching, but you will have a hard time
convincing an engineer that you need a second feedback loop to undo what the first feedback loop
is doing. And this is one of the many reasons why you will not find
that device in technology, it has not survived
the technological evolution. And that was my control education:
plenty of positive feedback, plenty of negative feedback,
never a mix. And that picture changed when I started studying
feedback mechanisms in nature, first as a hobby
more than as a profession. I wasn’t surprised to find lots of positive feedback
and negative feedback in nature, but I was increasingly surprised to see that every single device in nature
seems to have two-feedback loops, hardwired at twice the cost
of a single-feedback loop. So either natural evolution
obeys different rules than technological evolution or engineers have missed something here. So here comes that different mechanism. Plenty of negative feedback,
plenty of positive feedback everywhere, always separated in technology,
never separated in nature. This difference is so simple that it is
almost embarrassing to talk about it, but today my research is entirely
about exploring that difference because I believe that the mixed-feedback principle
is the essence of plastic localization, of multiresolution, this property
our brains do not share with machines. The mixed-feedback processor
is plastic, this is a long story, but here is a cartoon: on the right, you have
a strong positive feedback, so the output is all or none,
a switching device. On the left, you have a lazy regulator,
input is the same as the output, a regulating device. But in the center,
you have this balanced processor, and the balanced processor can be turned
into a switching or regulating device by a mere modulation of the balance between the positive
and negative feedback. This processor is highly localized. It activates in a very narrow range. But this localization is not just
in range, it is also in time, because if you make the positive feedback
a bit faster than the negative feedback, your device will first switch,
then regulate. This is called an ‘excitable behavior,’ and this behavior is the behavior
of every single one of your neurons. And you can also localize in space because you can make
the positive feedback short ranged and the negative feedback longer ranged. And by localizing in time and space,
you will build a wave with a certain wavelength
in space and time. And once you build one wave,
you can start modulating the frequency, the spatial and temporal frequency
of that wave by relying on the collective activity
of many such balanced processors. And nature has found
an incredible amount of different ways to modulate the resolution. It always seems that there is this very simple feedback principle
at work: never a positive feedback
without the negative counterpart. At the core of the computer,
we have transistors. And transistors are switches.
This is a positive feedback technology. At the core of the brain, we have
neurons that emit action potentials, which are produced by the combination
of a positive and negative feedback, a mixed-feedback technology. And your tens of billions neurons
collaborate in many different ways to change the wavelength
of those localized signals. I wish to think that machines will not win
the next stage of the imitation game until they will have incorporated
a mixed-feedback principle at their core. So whether you enjoy
scientific predictions or whether you are scared
by scientific predictions, you should acknowledge
that imitation game and artificial intelligence
have changed our world indeed, but that the machine
hasn’t eradicated the human yet. Instead, perhaps
artificial intelligence has improved our understanding of ourselves because it has improved our understanding
of how different we are from machines. And this is perhaps something
of interest to all of us, because we all like to play
the imitation game, we all like to assess how much
‘Her’ is different from a machine. So next time you play the imitation game,
you may want to remember that your best advantage is perhaps
not fast computing or quiz solving, your best advantage is perhaps
your ability to balance your positive and negative feedback
at many different scales. Your best advantage
is perhaps to discern, at least for a while. Thank you very much. (Applause)

One Reply to “Do brains compute? : Rodolphe Sepulchre at TEDxLiege”

  1. I might be speaking ahead of myself (I have only watched 3:08 of this video), but anyway, in order to win that game (if I was talking to a person or a computer) I would ask questions, not based on logic, fact, or reasoning… but feeling. for example "what goes through your mind (or what do you think of) when you see a sunset" I don't think think a computer could give an answer that makes me think "This is a human responding"

Leave a Reply

Your email address will not be published. Required fields are marked *