Re: Is sentience an emergent brain behavior?

JGCASEY wrote:

Michael Olea wrote:

[ ... ]

It is these sorts of mindless assertions, like Curt's "I've
solved AI" assertions that from time to time "piss me off".

Ouch!

The reality is you are better educated Michael. What can we do?

Everybody's better educated in some set of topics or another. I don't doubt
Dan knows hecka more about spiders than I ever will. What ever happened to
his "wobbly spiders" and his GA gait learner? Why doesn't he extrapolate
from things he knows, rather than pontificate on things he doesn't?

My formal training is in geology. This is at heart a historical science -
you arrive at the "scene of a crime" and deduce from fragmentary evidence a
sequence of events. There is no way to do that systematicaly without
adopting a strict discipline of "multiple working hypotheses". So when I
switched from geology to computer science I built my reputation, at first,
not so much on design ability as on troubleshooting.

I'm not talking about debugging some app - hey, the spray can tool in
pixelpaint makes a random mess when I set the brush width to zero. I mean
you get a phone call at 2am, you're on a plane by 4am, walk into a staff
meeting two thousand miles later, crammed with panicky and often hostile
managers, each with a "theory" of what's wrong (this or that "hanging",
"dying", "freezing"), none able to describe what they've observed rather
than what they think, and within an hour you are in a chilly computer room
with maybe 50 computers servecing 200 workstations on 5 intersecting
ethernets and remote links to 3 processing hubs on 3 continents, each of
those 50 computers running some 30 or 40 processes concurently, and each of
those 200 workstations running some 10 or 20 processes concurrently, and
every now and then some process, say "image fetch", "stalls" for about 8
minutes. That sort of thing.

Success in that game demands a strict discipline of multiple working
hypotheses. You have to be very clear on the difference between what you
think and what you know, what you beleive and what you have measured. And
you need a good working knowledge of the sorts of things that can go wrong,
ordered by likelihood - a hypothesis space that grows with experience. And
you need a systematic plan for gathering maximaly informative evidence to
home in on the problem (is it plugged in, fuses all good... no, I never got
on a plane just to replace a fuse, but close enough!). And you need a
decent "bedside manner" to sort through mysticism laden operator verbal
reports to separate observed symptoms from presumed explanations without
offending the client (hard to believe I was once considered good at that).
You ask an operator what the problem is and you usually get a diagnosis,
not a description. ("Just the facts, maam" - Joe Friday.)

Geology, especially field geology, was excellent preparation for
troubleshooting. And troubleshooting was a strong lesson in practical
epistemology. But not my first love, which would be mathematics. Group
theory. Metric spaces. Measure spaces. Topological spaces. Abstractions
that distill recurrent patterns into essentials. Dan always wants to dive
off into the details and mechanisms of implementations - I've got nothing
against that, I am a programmer, I do conceive, design, build, debug, and
maintain software. But I'm less interested in solving A problem than in
solving a CLASS of problems. What do all solutions have in common? What are
the fundamental limits imposed by environmental information gradients
independently of any and all mechanisms? So, for example, I do not conceive
of recognition devices (like threat detectors) in terms of neural nets,
decision trees, self organizing maps, vector quantizers, support vector
machines, nearest neighbor tesselators, AdaBoost, Bayesian networks,
Fourier descriptors, rational splines, linear classifiers, polynomial
classifiers, Gaussian classifiers, radial basis functions, kernel methods,
wavelet basis functors, or whatever. I think like this: all classifiers,
regardless of the underlying mechanisms, carve up some input space into
"equivalence classes". A digit recognizer operates on bitmaps, and it sorts
bitmaps into categories - all these bitmaps are 2's, all these are 3's, and
so on. So any specific digit recognizer boils down to a particular
subdivision of bitmap space into a finite set (10 in this case) of
categories. It induces a "partition" on its input space. The specific
partition it arrives at (whether this is fixed after some finite training,
or continues to evolve) is some partition out of a set of partitions the
device is capable of inducing. That set of possible partitions, the
partition space available to the device, has attributes - for example, a
"capacity" that can be characterized, roughly, by something called its "VC
dimension". This in turn has consequences for the ability of the machine to
generalize from examples. There is a sort of "impedance matching" law at
work between the complexity of classification tasks and the complexity of
hypothesis spaces. This approach leads to genuinely fundamental issues
applicable to any and all classifiers, mechanical or biological.

But does it predict the shape of sandunes?

I did, in those troubleshooting days, meet now and then with Jim
Crutchfield. Never one-on-one, always in a group - elevator ride, shushi at
Yoshi's. Clipped snippets: regular expressions, stochastic state
transitions, about your article in Scientific American, fuzzy grammers,
Lyaponov functions... Jim would not know my name or face, but K would, and
Jim would recognize K. The one thing I never forgot was the quote "a
department of nonlinear mathematics is like a department of nonelephant
zoology", which I repeated with great effect to the mathematical friends of
Rufus (ergodic theory) Bowen's courageous and charming widow, Carol (at
Chez Panisse, while waiting for calzones, the food, not the garment).

You can even converse in the language used in the experimental
analysis of behavior with Glen.

Glen has specific expertise that I value. He brings to this group a level of
precision and rigor, in his area of expertise, that I find refreshing.
There is, no doubt, a fruitful dialog to be had between statistical
learning theory and the experimental analysis of behavior. Glen is quite
capable of dispassionate assesments ("sometimes the most basic questions
have been insufficiently explored") if you are willing to make the effort
to discuss the actual EAB rather than some strawman simulacrum. Glen has
his agenda, I have mine (does the experimenter shape the pigeon, or does
the pigeon shape the experimenter?), but Catania's book on "learning" might
be the best purchase I have made in a couple of years.

Then there was your bassakwards purely mythological account
the history of computer vision and its relationship with
neuroscience:

Speaking of being better educated what does bassakwards mean
exactly? I couldn't find it in the dictionary or an online
dictionary but I found it was being used in many sources.

It's a joke - a self-referential play on "ass backwards".

-- Michael

.

Relevant Pages

  • Re: Is sentience an emergent brain behavior?
    ... switched from geology to computer science I built my reputation, at first, ... That sort of thing. ... Fourier descriptors, rational splines, linear classifiers, polynomial ... It induces a "partition" on its input space. ...
    (comp.ai.philosophy)
  • Re: "Sorting" assignment
    ... And many others prefer to call partition exchange because "quicksort" ... bin B depending on whether it is greater than, ... If the array is already sorted, this means that you end up ... attempt to sort them. ...
    (comp.programming)
  • Re: Multiple Key Sort in Fortran
    ... The sort might optionally be repeated based on subsets of the input ... GLIBC's QSORT, which is actually QSORT down to a threshold, where it switches to INSERTION sorting, which is faster for smaller ranges. ... insertion sort is used for the remainder, because it is faster for small sort ranges. ... Ignore both small partitions: ...
    (comp.lang.fortran)
  • Re: Partition and Dual Boot (was Windows XP)
    ... Yep - there have actually been quite a few Threads about that sort ... So you may not be able to use an Upgrade WinXP Install CD then. ... Well, actually a few years ago, it was *more* common to get an Install ... Unless you fork out Money for something like Partition Magic, ...
    (uk.people.silversurfers)
Loading