Sunday, April 01, 2007

The Neural Plausibility of Connectionism

For those of you not interested in cognitive science, here's a picture
of my pug wearing her winter coat. For everyone else, onto the cogsci...

I was recently looking at Martindale's book Cognitive Psychology: A
Neural Network Approach
. Most ideas of how the mind work are
broadly categorized into "information processing" theories and
"associative" theories. As described in the introduction to this book,
both are based on metaphors: the former that the mind works like a
computer, and the latter that the mind works like a brain.

The idea that traditional cognitive science looks at the mind as a
computer has always bothered me. It's not quite right. Or at least,
it's not quite right anymore-- maybe it was right at some point. My
first reaction to this is "no, we don't look at the mind as a
computer, we look at it as a computer program." There's a big

Computers have registers, a central processor, a memory, input
devices, etc. Computer programs, on the other hand, are much more
broad. If fact, they're almost too broad to be any use as a
metaphor. But there is a nugget of importance there: traditional
cognitivists view the mind as basically a computational object, and
that thinking consists of manipulating representations.

The problem with this notion, though, is that it includes the
associative theories too (neural networks, connectionism, etc.)
Indeed, when people make connectionist models of the mind, they are
often implemented as computer programs!

However, the traditional cognitivists think, roughly, as the brain as
the hardware and the mind as the software. And, as we know, the
workings of the software need not have any similarity to the working
of the hardware. To take a simple example, something can work in
parallel in software but serially in hardware. The idea of a "virtual
machine" is very valuable here. There is a physical machine, but the
virtual machine, which is created through the processing of the
physical machine, has its own rules of operation that we can endeavour
to understand.

I really like the idea that connectionists see the mind as working
like a brain. At first glance, though, doesn't it sound kind
of... obvious? Some like to say that the mind is the brain. It
seems, as many say, that the connectionist approach is "neurally
plausible." I think this is why I'm finding so many undergraduates
fascinated with connectionism. It appears, at first glance, more
scientific, more likely to be right, if we are actually taking about
how the brain works at the same time as talking about how the mind

But hold up, that's not what is being said! To say that the mind works
like the brain is not to say that the mind is the
brain. A metaphor might help here. Imagine someone were to tell you
they had a new theory of traffic congestion, and that it works like
schools of fish. You might think "okay, that sounds like it could be
promising." Then another person told you that they had a better
theory: traffic congestion works like a car. The school of fish
theory sounds much more plausible, doesn't it? I mean, if traffic
congestion systems worked like a car, there should be analogs to the
AC, the engine, the seat, the oil, the starter, the windshield wipers,
etc. in traffic patterns. Nuts, right?

To a traditional cognitivist this is how we see the inherent
legitimacy of the connectionist approach. The brain is not a good
metaphor for the mind simply because the mind is implemented on
the brain, any more than cars are a good metaphor for traffic jams, or
cell physiology is a good metaphor for the workings of the nervous
system, or saying that Microsoft Word works like a computer, or that
DNA works like an atom. In all these cases, even though the
higher-level system is implemented on the lower-level one, it's clear
that they work on very different principles. Just because the nervous
system is, in some sense, nothing but cells does not mean that the
system as a whole works like a cell.

The credit we give connectionists are due to the ability of their
models to explain and predict what we know of the mind-- that that's
nothing to sneeze at. They've been great for lots of things. But the
fact that it's brain-like should not give it some special appeal.

It's irritating when connectionist fans don't appreciate this. Every
one of them bends over backwards to say that the nodes, or units, in a
connectionist network are not models of actual neurons. But if they
are not models of neurons, then what, exactly, is the inherent
benefit of having a neural metaphor?
That is, if the
connectionist network is a virtual machine running on the physical
neural hardware, just like a production system could be running on
neural hardware, then why should we give it any more credit for
"neural plausibility?" Until connectionists take a stand and say that
they are actually modelling the neurons in the brain, I don't want to
hear anything more about connectionism's "neural plausibility."

1 comment:

Dustin said...

What a wonderful juxtaposition of A cute dog and mouthy title.