MITS RENEWED FOCUS ON A.I., AND IBMS KITTY BRAINS
(Posted 2009-12-11 16:58:15 by Ray Lopez)
There's been a lot of news lately in the world of artificial intelligence.
First up, was the news from IBM Research. Last month they reported that
they had constructed a cortical simulator built upon the IBM Blue Gene
supercomputing architecture. According to IBM, this cortical simulation is
able to run at near-real-time, and possesses the power to simulate the
activity in a cat cortex. Exciting news, to be sure.
However, a few weeks later Henry Markam, an IBM Blue Brain project
director, publicly issued a nasty letter calling the whole cat-brain
achievement a "PR stunt". The choicest quote from Markham's letter is this
one:
_This is light years away from a cat brain, not even close to an ants brain
in complexity. It is highly unethical of (Dharmendra) Modha to mislead the
public in making people believe they have actually simulated a cat's brain.
Absolutely shocking._
Wow. For their part, IBM has stood by their original announcement. But
it's important to note that this little family quarrel amongst IBM
researchers is indicative of a much deeper level of discord that goes way
beyond IBM. Markham was originally trained as a neuroscientist, while the
team leader who is taking much credit for this work, Dharmendra Modha, is a
computer scientist. Neuroscientists and computer scientists are going to
have different approaches and different expectations for working on the
problem of brain simulations. Computer scientists like Modha are going to
look at their estimates of the raw computing power of a mammalian brain,
and use those estimates to decide how much horsepower they need to simulate
that raw computing power _in silico_. Once they're able to hit those
horsepower numbers, they can declare a sort of victory. Meanwhile,
neuroscientists such as Markham are used to studying, in minute detail, the
complex underpinnings of relatively simple functions and behaviors. To
them, it is impossible to assert that a cat cortex has been "simulated",
because such a statement implies a very deep level of understanding of what
it is the cortex does and how it works, something we currently don't know.
As a result, both Modha and Markham are right: Modha can declare victory
from his perspective because of the fact that, by his own standards, he has
managed to "simulate" a cat cortex. But Markham is also correct, in that
we haven't even begun to simulate the real underpinnings of neural activity
that give rise to even the simplest forms of behavior.
I've met Dr. Modha and have chatted with him a few times regarding his
work. He's a nice guy, and very sharp. What was clear to me, though, is
that he definitely sees neuroscience as nothing more than a support tool
for his work. We neuroscientists should simply provide data to the
computer scientists, and not worry about how those data are being used. I
think that most neuroscientists would prefer to see this work proceed in a
much different fashion. Rather than work on simulating a mammalian brain,
we should have begun by completely simulating the nervous systems of
invertebrates, including planaria, cockroaches, and flies. Understanding
the way these simpler organisms behave would go a long way toward helping
to understand how more complex animals behave.
Maybe we should go down even further: Why not start by simulating the
entire activity of a single neuron? We know much now about how neurons
work, how they control and regulate their electrical charges, activation
pathways, neurotransmitter production, gate control mechanisms, and much
more. Neurons are enormously complex cells, and there's reason to believe
that at least some core behaviors (e.g. Pavlovian conditioning) may be
found at the neural or molecular layer. We may not be able to fully
understand how our brains work until we understand exactly how the
complexity of a neuron contributes to those inner workings.
Meanwhile, in another story, it was recently reported [
http://www.theinquirer.net/inquirer/news/1565254/mit-plans-ai-redux ] that
MIT is launching what they're calling the Mind Machine Project (MMP). MMP
is to be a $5 million effort aimed at "fixing" the mistakes of some 40
years of AI work. A collection of experts will be working together to
create new, intelligent machines that they hope will live up to the dream
and hype of AI that existed back when it was first born. Back then (the
1960s), there was little doubt in anyone's mind that there would indeed be
a HAL 9000 computer by the year 2001. I fear that the MIT workers will be
committing the same sins as their brethren at IBM. They'll likely focus on
top-down approaches to various problems such as "cognition".
What they should be doing is working first on building a framework where
simple machines can be augmented by more and more complex "layers" or
components. This would be much more like the evolution of brains and
behaviors on earth. In addition they should start with components of
simple behaviors, such as sensation and perception, then move on to simple
behaviors such as avoidance, sensitization, and habituation. Any machine
that can do behaviors at the level of a worm or cockroach is ready to be
used to support more and more complex behaviors.
I hope that the folks at MIT realize this go forth in such a manner. The
folks at IBM are much to enamored of their splendid hardware to change
course now. But there are a lot of smart folks over there, and I hope that
some of them realize that they need to take a different approach with brain
and behavior simulation, and listen to a few neuroscientists. Whatever
happens, all of this promises to be very exciting to watch, and I continue
to wish that I could be part of it.
--------
There are no comments on this post.
To submit a comment on this post, email
[email protected] or visit us on the web
[
http://ratthing.com ].
