[Info-vax] Pattern matching (what do my VMS colleagues think?)

[G Everhart]

 From:
Glenn C. Everhart
156 Clark Farm Rd.
Smyrna, Delaware 19977
everhart at gce.com
18 Jan. 2010

My work on a model of intelligence in the very late 1960s was able to explain a
good many features of how we think, based on a model whose primitive operation
was detecting the most similar match of data patterns. (Of course there were
inputs and storage and the like also, but the fundamental cycle it assumed was
that it found the most similar parts of patterns stored to patterns being sensed
or considered from previous cycles. The basic logic was at least 3 levels, which
I though of as on or off pixels, in areas either of interest or not, so that a
part of data could be either off, on, or uninteresting.

Much later, in the late 1970s I think, I was convinced to write it up and apply
for a patent, which was granted on the image matching device. This design used a
material that allowed a field to control the direction of a current to be
detected and used a strategy of matching 0 and 1 pixels separately, but with the
field moving to in effect move one pattern over the other, and summing the
currents. By varying the direction and divergence of the field incrementally,
one could follow a steepest-descent scheme to find the best match. The most
current would mean the best match...

I proposed this using magnetic fields and material with a large anisotropic
transverse magnetoresistance (very cold bismuth might do), but there are other
ways to do it. The material in liquid crystals is large oblong molecules that
line up with electric fields. If such a molecule could be used as a light pipe
for some wavelength of light, it could be steered by an electric field instead,
which is easier to manipulate than a magnetic one. The field can be whatever
works, and the current would in that case be an amount of light (or other em
radiation, getting a frequency that works). By using a moving field, this
superposition can be achieved without moving parts. By allowing the field
to diverge, some variation in size of images can be handled.

What the gadget does not do well is rotation, which in principle might be doable
with magnetic fields as well as electric in one device, but recall that people
do not match patterns well that are rotated either, so I thought that this may
not be essential.

Anyway, getting back to the imaging device's use, it seems likely that the basic
capability of humans from which human intuition and understanding come is an
ability to match similar patterns.

Traditionally such matching has been hard for computers, with things like the
Hough transformation taking O(n^3) cycles for patterns of size n on a side. A
steepest descent following scheme is approximately linear, and if done with an
analog system as I had proposed could be pretty simple. (The comparison would be
between a smaller image area of interest with a larger area of something stored.
It would be useful to be able to include a second area on the edges of the image
of interest and be able to detect when currents to that area would substantially
improve the match to some area stored, so that could feed adjustments to the
area of interest...)

Pattern matching need not be for images or sets of sensed data only. Nor need it
only be for cross matching of data (as occurs for some people who experience
numbers or sounds as colors or shapes, incidentally dramatically improving their
ability to remember patterns in them). If any patterns can be produced, where
there is some continuity between the adjacent ones (e.g. time series of
different but correlated values) then patterns might be similar to some
previously seen ones. A fast pattern matcher would allow such similarities to be
noted.

One could perhaps consider measures of behavior, and perhaps be able to detect
signs of ill intent, of creative ability, of benevolent intent, or whatever by
patterns which are made up of pieces which are hard to distinguish individually.
One might be able to take financial measures which are correlated, and display
time series of those and notice patterns which were similar to, say, past cases
where frauds were done, or where substantial changes in some market might be in
the offing.

A great deal is done now by attempts to match patterns of these sorts, and a
gadget of this sort may be too slow, or the continuity of the parts of patterns
too limited to allow gradient search to be used, where purely digital techniques
might manage by sheer brute processing power. It is possible though that such a
gadget could make useful kinds of searching cheaper to do. The patent is long
expired now but is number 4,475,238 for those interested in background.

(I suppose that now one could use shift registers in numbers that were years ago
impractical, to move images around in a detector, but the analog scheme's
intrinsic tolerance of small changes is still hard to get so directly.)

Thoughts?

Glenn Everhart