Joe Blaylock
April 27, 2007
B553, Homework 5
Creative Commons Attribution, Non-Commercial, No-Derivitives, Share-Alike

I think that simulation, empathy, and "the self", are all intimately tied up
in one another.  It should be understood that I don't believe in a self, per
se.  I think that the collection of indicators that we take to be evidence of
some intrinsic "I" are either illusory phenomena, or else byproducts of other
mental processes (what these processes might be I'll get to in a minute.)  The
notion that we are separate unique beings, fundamentally different from the
world of phenomena that we observe, seems clearly untrue: as physically
instantiated beings in the world, we are part of the world, as are our thought
processes.

This may sound vaguely nihilistic, but it's not at all.  In fact, it gives me
great hope for the future success of the programme of Cognitive Research.  

I think that simulation is the key.  I think that brains are very fancy
simulation engines, designed to draw simulation parameters from sense data.
We're built to do science, and babies collect data about the world.  They
initially simulate to understand the outside world, but eventually, the
causality gets applied to one's own feelings and actions (but perhaps not
without some social pressure) and this leads to the development of the idea of
the self as distinct from other observable things.  Eventually (if we're
lucky), the process iterates once again, and the self (running in emulation on
top of a simulation of the environment) becomes a modelling ground for
understanding others (empathy).  It's an essentially recursive kind of
process, although I think that the substrate that engine is built out of is
modelling.

This is the obvious place to dredge up mirror neurons and start talking about
all kinds of spooky stuff.  But instead, I'm going to try to ground the
importance of simulation.  Before I do that, however, I'd like to note that I
think all of the big problems of cognitive science are really aspects of the
same thing.  When I say that I think simulation is key, that's not to say that
compositionality or any of the rest of it (except maybe symbols; I don't
really believe in symbols) are any less important.  I can, however, imagine
features like object distinction being part of some quasi-statistical
machinery that just sort of comes with the brain.  It doesn't seem easy, but
it seems less hard than the gestalt problem that simulation represents.

Oh, but what do I mean by simulation?  Extracting from the environment
distinct objects, observing their behaviors and the effects of those behaviors
on other objects, and inferring generalizations about those "relationships".
Note that the environment includes one's own body, so some of these rules will
be object correspondences such as realizing that the eyes you see in the
mirror are your eyes, and the eyes you see on faces are also eyes, even though
they're not yours.

Learning changes the fitness lanscape explored by evolution (riffing on
Baldwin).  Depending on how you look at it, simulation is either the central
tool of learning (real learning, as opposed to just memorizing), or else it's
the whole game.  This implies that the ability to simulate is crucial for
learning creatures; it's fundamental.  The advancement of cognitively
interesting lifeforms in some sense is just the gradual improvement (by
evolutionary selection coupled to learning-granted advantage) in the mental
machinery of simulation.  At the most primitive level, a worm that burrows
away from sound pulses in the soil may be considered a kind of very coarse
simulator - it predicts that soundwaves lead to pain.  The fact that there is
no observer to interpret the prediction as such matters little; the point is
that every learned association is a kind of coarse simulation.  I contend that
the machinery of personality is roccoco embellishment enabled by the
refinement of those abilities.

I think that Baldwinian evolution is a potentially valuable tool in exploring
fitness landscapes for complicated alife instances (which, let's face it, are
the ones we're going to have to get to eventually, if we want to do anything
interesting).  This is a blessing and a curse; I think that it's a real
phenomenon that we can exploit to ensure that, given enough time and model
complexity, we can get cognitively interesting behavior (ie, us) out of a wide
range of different neuromorphic architectures (though I'm coming around to the
CTRNN camp, personally).  But it's so very expensive, and so very slow.
Obviously not as much so as waiting for real evolution, but an intrinsically
expensive undertaking.

As someone personally committed to the idea that the best way to understand
something is to build it, however, I think that it's an undertaking well worth
the cost.