Re: Goal of AI: Perfect or Bounded Rationality

JGCASEY wrote:

Michael Olea wrote:
Curt Welch wrote:

Michael Olea <oleaj@xxxxxxxxxxxxx> wrote:
Curt Welch wrote:

...and even a turning test will be hard...

I think you mean "Turing test".

:)

Yeah, every time I try to write Turing my fingers type Turning. I
normally catch it and correct it, but not this time. (well, actually, I
have no clue how many times I've made the mistake and not caught it...)

Have you seen the FedEx commercial?

We don't get "french benefits", we get "fringe" benefits.
It's not the leaning tower of pizza, it's the leaning tower of Pisa...

Yeah, I can relate. I've always had problems with language and I've
had people correct me on things just as dumb my whole life.

When I took an introductory psychology class in college I had seen the
word "placebo" in print, and knew what it meant, but I had never heard it
spoken. So when I answered a question I pronounced it like "place bow".
The whole class cracked up.

Nice to see others have had a similar experiences. The same thing
happened to me first year secondary school when I was reading
aloud to the class and pronounced the 'b' in subtle.

Although my compositions always had a few spelling mistakes
the stories were always well received while some good spellers
complained they couldn't think of anything to write about.

Well, I've embarrassed myself far worse than mispronounce words in public.
Spouting gibberish at a high level meeting comes to mind. I sat in, as a
consultant, at a meeting between representatives of one of my clients, and
one of my client's clients about a project called "forms subtraction". The
basic idea was to compress an archive of images of documents with data
entered on standard forms by first "subtracting" from the images the form -
the boilerplate, the lines and other form-invariant imagery - leaving just
the variable data, the entered data, the informative data, before running
whatever compression algorithm on the image. Then, when the image needed to
be diaplayed or printed, it was to be "reconstituted" by decompressing the
variable-data only image onto a backround image containing the form. So you
store one image per form, and data only images per transaction (which on
average, so the thinking went, would be more compressible since the form
pixels were now all white).

The whole scheme was a bit misguided (there are much better ways, for
example 2-D LZW type ways, of exploiting the redundancy of recurring forms
in achieving high compression ratios than introducing the host of suBtle
problems involved in "subtracting" out and then "adding" back in form
pixels), but that is not where I went astray. I had studied Michael
Barnsley's book "Fractals Everywhere", and was in the midst of reading a
later Barnsley book on fractal image compression. I blithely quoted one of
Barnsley's claims: "fractal image compression is achieving compression
ratios of a million to one!" This utter nonsense created some excitement,
but after the meeting a friend of mine who was also in attendance pulled me
aside and pointed out that if this were true then all one-megabyte images
could be compressed to a single byte. And since there are only 256 distinct
bytes if such compression were possible this would mean there are only 256
distinct one-megabyte images. DOH! Technically, since fractal image
compression is a lossy rather than lossless compression scheme it would
really only mean that the compressor would partition megabyte images into
256 equivalance classes - those that compressed to the same byte. But it
amounts to the same thing - the decompressor could only generate 256
distinct one-megabyte images.

The reason for all this is that Barnsley had gone from brilliant expositor
of mathematics, in "Fractals Everywhere", to a shameless patent-happy
huckster when he was promoting fractal compression (a perfectly reasonable
technology when separated from the hype). There is in fact a lot of good
mathematics in Barnsley's book on fractal compression (a book I later
traded in for a couple of tacos, and I don't remember it's exact title).
But it also contains much misleading hype. Basically fractal compression
amounts to finding compact formulas for generating an image. Rather than
storing the image you store the formulas. So, for example, if you have an
image of a circle, rather than storing the bits, at whatever resolution,
that make up a rasterized circle you can store the coordinates of its
center and the length of its radius. Now, when you display the image you
can display it at whatever resolution you want. The original image may have
been 256x256 pixels, but given the *inference* that this is the image of a
circle, whose formula takes only a few bits to store, you can display it as
a 1024x1024 pixel image, or as a 2 billion by 2 billion pixel image, or
whatever you like. This is how the so called "million to one" compression
ratio was justified - blow up the image to whatever scale you want and
claim whatever "compression ratio" you want. This is, of course, not
compression, but interpolation.

So, after my public idiotic hype-driven comment, and the subsequent discrete
friendly correction, I went home and pulled out Barnsley's book and figured
out what he had done to lead me astray. And I did get a chance to redeem
myself later when I was able to demolish an argument against lossy
compression of legaly binding images of documents. The idea was that you
could not discard information, by using a lossy compression scheme, when
archiving legaly binding images. But what they were doing was: 1) making
grayscale images of a document (a loss of information), 2) BINARIZING these
images (a HUGE loss of image information), and then 3) insisting that these
much information reduced binary images be compressed losslessly, preserving
every last bit of the already very lossy artifact. My point was that a
lossy compression of the grayscale image, which was as compact as the
lossless compression of the binary image, lost much less information than
did binarizing the image in the first place. That argument and a demo won
the day.

I tell you all this, of course, in anticipation of ripping apart, time
permitting, your latest remarks about "The animal, not the environment".
(If I get around to it, I hope you understand it will be the comments I am
ripping, not the commentator).

-- Michael

.

Relevant Pages

  • Re: Compression by descent
    ... full data square. ... fractals is selected, I can show that a file containing random data, ... industrial_...@xxxxxxxxxxx, to answer your question, the compression ...
    (comp.compression)
  • Re: 4mp doubled = 16mp!?
    ... quality JPEG settings did indeed make more detailed images ... my 2 Kodaks have no ability whatsoever to alter compression, ... large prints from as large an image in mega pixels as you can ...
    (rec.photo.digital)
  • Re: newbie question about changing resolution/size
    ... pixels wide by 2592 pixels high, and the resolution is 72 pixels/in. ... For proper display however that would require the ... physical size of the print requested, the amount of data in the image and ... Compression is a method of reducing file size for a given image. ...
    (comp.graphics.apps.gimp)
  • Re: Are CODECs used by Blu-ray and HD-DVD owned by Microsoft?
    ... general and DCT based compression in particular. ... Yes - all video codecs are lossy when using Blu-ray, ... And Blu-Ray is not full AVC like HD-DVD, it's only a cut back version. ... all day long always ending back with the original file you started ...
    (alt.tv.tech.hdtv)
  • Re: microsoft photo editor: size reduction = ?
    ... The menu says my photo is some 19 cm wide. ... something about pixels, too, but did not look at it. ... Then there is "color depth", a range of numbers that describe the shade of color for each pixel. ... Lastly, there is compression, to make the files smaller. ...
    (alt.photography)