I've just read the book Alan Turing: The Enigma, by Andrew Hodges, an outstanding and
profound-- if thick-- biography of Alan Turing. Turing's work touched on some deep
philosophical questions about the relationship between brains and computers. I
naturally had my own opinions, but I wanted to talk to somebody with more knowledge
of brains who was also computer savvy--someone with a foot in both worlds. So I paid
a visit to Michael Greenspon, who develops software models of neural systems with
Walter Freeman at UC Berkeley. We got together for lunch and had a very interesting
conversation. Here's a sample:
[Audio embellishment: clinking of nice glassware as Dave and Michael dine in the sun]
DKJ: I heard something recently that struck me as profound: computers don't
manipulate reality, they manipulaterepresentations of reality. The profound part is
that seems to be what brains do, too. Alan Turing, for much of his life, wanted to build
a brain. He firmly believed that consciousness was caused only by the operation of the
brain, and that the brain's operation could eventually be described at any level of
detail.
[Michael looks patiently skeptical, but Dave plunges ahead, oblivious, waving his fork
excitedly.] Further, he had previously proven that in principle, a "universal
machine," of which the computer is a finite approximation, could simulate any other
logical machine, and thus any logical process whatsoever. So if you could describe the
function of the brain as a logical process, you should be able to program a computer to
"be" a brain. The description part, of course, is the killer. But I can't help thinking
that we'll get there eventually. What do you think?
MCG: Whoa, Dave [almost choking on his exotic Thai salad], I think you've hit the
intellectual cul- de-sac of traditional artificial intelligence. The reason it's so hard to
describe the operation of the brain as a logical process is simple: it isn't a logical
process at all. That's a cerebral approach to a fundamentally biological and physical
problem. I'm sure someday we'll be able to logically explain the operation of the brain
in terms of physics, but that explanation won't include a computational mechanics
based on formal logical operations.
DKJ: But then how do you approach the problem of trying to understand and model
brains in your lab, if you can't describe them as logical processes?
MCG: Our approach is that of computational neuroscience; we're doing dynamic
modeling at the level of cell populations, using massively parallel machines with a
Macintosh front end.
When I say representationalist AI is a cerebral approach, it helps to realize that the
cerebral cortex is just a few millimeters thin. It's a tissue essential for generating the
separatist intellectual conception of ourselves as humans, but it's really a translucent
veneer over the bulk of what our brains do day in, day out, which comes from our
animal ancestors. Before we ever learn formal or even naturallanguages, our brains
are already highly developed as processors of spatial, tactile, and kinesthetic
information, to name my favorites. This is one reason why the Macintosh has been so
successful as a tool--because it's the first readily available machine to offer at least
at the outer layer a spatially based interface.
DKJ: And the reason that's so great is that our brains process spatial information
effortlessly, without our even trying.
MCG: Right, a spatial interface allows us to apply more of our innate biological
intelligence in communicating with the machine. But both structurally and
functionally, the digital computer as a metaphor for the brain is almost completely
inaccurate at every level of analysis.
I think if you look further into the nature of thought and perception, and also look
more carefully through microscopes and macroscopes at what real brain tissue is
doing, you'll see a physical system that exhibits chaotic dynamics in time, has fractal
extent in space, and is inextricably linked to the natural world. Computers are
powerful tools for simulating and visualizing these properties, but they don't
themselveshave these properties yet.
DKJ: Especially the links to the natural world.
MCG: Exactly. If you want to apply computational metaphors to the brain, perhaps the
brain is like a fractal architecture computer that can compute infinitely recursive
functions in finite time.
DKJ: Oooh, I like the sound of that. Fractals, computers, infinity, and recursion all at
once.
MCG: I like it too, but that's really just a structural metaphor. I'm interested in what
we can learn about how real brains might work, so that we can apply these principles
to next-generation user interfaces and to new non-von Neumann computing
architectures.
In an engineering sense, we're after machine perception. That is, we want future
machines to interact in the human sensory world, rather than forcing humans to
interact in the virtual world of the machine.
DKJ: Yeah, to use or program a computer today you still have to interact on the
machine's terms. I think one good approach to changing that is to try to build
computational structures that are like the brain, so that our machines will be a little
more like us. There are 10 billion neurons in the brain, more or less, right?
MCG: More. And perhaps 1015 synapses, which you could say is where a lot of the
computation is going on.
DKJ: OK, more than 10 billion neurons in the brain, and they're wired together
inunbelievably complex ways. The point is this: I'll bet that we can simulate a single
neuron fairly closely with a computer, and over time we can get our simulation closer
and closer to the real thing,arbitrarily close. Further, I'll bet that someday it will be
possible to get 10 billion little computers together and talking to each other. I know
this is a little speculative, but my business card says "Limit Pusher," and I feel
compelled to live up to it.
MCG: Rave on.
DKJ: So we set this thing up--10billion little processing nodes--and we turn it on
and start feeding it information. What will happen? What will it do? I can't help
thinking that whatever it is, it will be something very much like life. And just as
mysterious.
MCG: Well, I don't think it's purely an issue of scale. At Berkeley, we're building a
new ring architecture parallel machine based on superscalar processors that can
accommodate multimodal sensors and effectors. It's called CNS-1 and is spec'd at
upwards of 100 billion operations per second.
DKJ: 100 BIPS!
MCG: Right. Or 0.1 TRIPS, which is perhaps a better indication of
how far we have to go. We expect CNS - 1 will be able to simulate many of the
emergent dynamical properties of cell populations observed in real brains--to run
what I call the lava lamp model of the mind. But even this much power won't bring us
"arbitrarily close" to the wetware. I don't think you'll want to say it's alive or that it
works the way a biological brain works.
DKJ: Maybe not, but I think that a network of 10 billion processors couldact
something like a brain, could seem like a brain, even though it's not one by any
stretch of the imagination. That idea fascinates me: that a computer, or a bunch of
computers, can behave like something else. This gets back to Turing's thesis that a
computer can simulate anything, if you can describe the thing in enough detail. That
begs the question, though, of whether the simulation isfundamentally the same as the
reality it simulates.
MCG: Is it live or is it Memorex?
DKJ: Precisely. It's like comparing painting on the computer to painting using canvas,
brushes, and oils. At one level of description they're identical activities: applying
color to a surface in intricate and skillful ways to produce a little piece of space that
other humans can look at and experience emotion toward. But the tools differ hugely
and, perhaps more important, the experience of using them is completely different. So
I guess what I'm saying is that at the right level of description I believe (well, I want
to believe) that it's possible to "build a brain."
MCG: Or to grow a brain. I think you're barking up the wrong dendritic tree.
It'sexperience that's essential. Brains are dynamic systems that actively reach out into
the sensory world for experience; perception is a creative process, not a passive one.
To talk about building a machine with the capabilities of the human brain you have to
include the same kinds of connections to the world that humans have. In the real tissue,
it goes right down to the level of quantum phenomena and beyond--what I call "real
virtuality."
What I've been trying to get across is that real brains operate by virtue of being
physically continuous systems; there's an interplay between the nanoscopic and
macroscopic, the intrinsic and extrinsic, such that structure and function are not
separable. The notion that there exists in brains a "level of description" at which
cognition is implemented as logical operations is a convenient fallacy, what John
Searle calls "closet dualism." It means, for example, that if you want to start
capturing the creative, human aspects of language--not just the literal, but the slang,
humorous, ungrammatical, and allusory--you have to model the dynamics of the
underlying physical processes.
DKJ: Hmm, this point about not being able to separate cognition from sensory
experience is important. It's interesting to compare the development of computers
with the development of life. Computer sensors and effectors--the parts of
computers that by necessity touch the world--always seem to lag way behind the other
parts, the computing parts, in their development. And the gap seems to be widening. So
computers are currently wrapped in sensory cellophane, while the connection of
biological systems to the world is very strong and high-bandwidth.
MCG: Exactly. It's likely that in biological systems, sensors and effectors developed
first and, as part of an evolutionary feedback loop, drove the development of the
nervous system. Though now you could say the demands of more sophisticated user
interfaces are driving the development of CPUs. The perceptual side is limited to 2-D
mouse tracking and 1-D clicks and keystrokes. But speech and pen gestures are about
to expand that. Eventually computer-human interface will be polymodal, including
intonation, spatial gestures, eye position, facial expression, and cortical activity
patterns-- what I call the "think-along interface."
DKJ: It fascinates me that programmers can so easily get sucked into the machine--I
knowI've been there-- despite the very limited modes of interaction with it.
MCG: Yes, in programming, I often feel I'm being sucked into a one-dimensional world
of historical arbitrariness. I think this comes from the fact that while the complexity
of our software systems has increased exponentially, our development tools haven't
kept pace. The current tools fail to providethe real-time, interactive turnaround
that's crucial to maintaining the creative flow. They force us to think too much about
the machine's problems, instead of the human problems we're presumably trying to
solve.
DKJ: Amen. And it's true for nonprogrammers, too. So how would you like to see the
tools improve?
MCG: Well, besides speed--where speed means real-time, no perceptible delay;
anything less is slow-- future tools will have semantic knowledge of the process of
software engineering and eventually of the application you're building. The code
browsers are a good step forward; at least they can automatically determine structure
from syntax. The next step is to automate the build process, the incremental linking of
components, and the maintenance of an audit trail and nonlinear undo space for source
code. Here we start to blur into a dynamic-language sort of model.
DKJ: That's exactly the kind of administrivia that computers are supposed to be good
at. But right now, for most of us, the burden is still on the human.
MCG: It sure is. Where we want to head is to shift the focus of the iterative process
from the syntax level--compilation, debugging--which is what the machine is
concerned with, to the level of design and validation, which is hopefully where the
programmer is trying to solve the semantic problems of the application.
DKJ: Way back in the 1940s Turing talked about the fact that ". . . as soon as any
technique becomes at all stereotyped it becomes possible to devise a system of
instruction tables which will enable the electronic computer to do it for itself." In
other words, as soon as we can describe how we do a job, we can program the machine
to do it for us. This is happening, but slowly. As an amusing footnote, he went on to say
"It may happen however that the masters [programmers] will refuse to do this. They
may be unwilling to let their jobs be stolen from them in this way. In that case they
would surround the whole of their work with mystery and make excuses, couched in
well chosen gibberish . . ." He was a pretty prescient guy.
[Setting his napkin on the table] Well, I guess we should try to wrap it up here; our
readers' MacApp builds are probably finished by now, and we'll be losing them soon.
Let's try to wring a message out of our ramblings, something developers can take home
with them. How about this: Strive to bring computers ever more firmly into the world
of people, rather than trying to cram people ever more firmly into the world of
computers. The differences can be subtle, but the distinction is very important.
MCG: Well, I think we can and will go much further toward humanizing the experience
of using computers. But I don't think we have to couch what we do in gibberish to keep
our jobs, because programming is fundamentally a creative discipline. Like other
creative disciplines, when you've done it long enough and intently enough, you tend to
see its way reflected in everything you perceive. You could say programming is a way
of seeing. That leads us to computers as tools for extending human visualization.
[Flipping up his shades] The point is that it's human vision--not the
technology--that's crucial. When we create tools and toys and lifestyles that separate
and insulate us from nature, we further the consumption and destruction we see all
around us. But I think we can see past the empty goal of creating trillion dollar
markets for our products. As humans, we've always had the infinite power to change
our minds. It's time we tap that power by creating tools that connect us--to each
other, to the earth--and enable us to meet the real life-or-death challenges we face on
this planet. As programmers and technologists we're in a key position to determine the
future by the choices we make every day. I hope each of us can make every keystroke
and every mouse click a step toward a sustainable society.
RECOMMENDED READING
DAVE JOHNSON once borrowed a friend's video camera so that he could spy on his
dogs when they were alone. He carefully--and gleefully--set up the camera near the
front door, turned it on, and went out for dinner and a movie. The dogs mostly just
slept, with an occasional barking fit, apparently just for doggie grins. It was really
very dull viewing except for one incident about halfway through: the smallest dog,
affectionately known as Dinky, sat down right in front of the camera, stared balefully
into the lens for a moment, then put her head back and howled for a full five minutes,
something Dave has never seen before or since. *
MICHAEL GREENSPON is a doctoral student in the department of Electrical
Engineering and Computer Science at UC Berkeley. When he's not cramming for quals,
he can often be overheard trying to explain the cost benefits of telecommuting to Apple
managers. (We're still not sure when he sleeps.) If the sun's out, you're sure to find
him soaking up some of it; since the release of the Macintosh PowerBook and
ToolServer, he's hardly been seen indoors except for an occasional rave. In fact, he and
Dave Johnson were recently spotted rigging a LAN in the outfield at Golden Gate Park.
He does, however, respond to his e-mail: he can be reached via AppleLink as INTEGRAL
or on the Internet as mcg@icsi.berkeley.edu.*
Dave welcomes feedback on his musings. He can be reached at JOHNSON.DK on
AppleLink, dkj@apple.com on the Internet, or 75300,715 on CompuServe.*