MIT talk

[From: Bruce Nevin (Fri 921002 13:14:03)]

I'm back from my month's field trip among the Indians, jet-lagged, and
with much to catch up on. The following of possible interest was
forwarded to me from the MIT AI list.

···

-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-
             WHY DO WE SEE THREE-DIMENSIONAL OBJECTS?

                           Thomas Marill

                          October 1, 1992
                               4pm
                         8th floor playroom

                             Abstract

When we look at certain line-drawings, we see three-dimensional objects.
The question is why; why not just see two-dimensional images?

In this talk we explore the theory that we see objects rather than
images because the objects are, in a certain mathematical sense, less
complex than the images.

However, there are infinitely many objects that project to any drawing.
As a result, a second question arises: Given that we are going to see an
object when we look at a drawing, which one will it be?

The theory under discussion holds that the object selected by the vision
system will be the least complex of the available alternatives.
Experimental data supporting the theory will be reported.

This work is based on the pioneering ideas of Solomonoff and Kolmogorov,
and on the more recent ``minimum description length'' concept of
Rissanen.

    ****Revolving Seminar****Revolving Seminar****Revolving Seminar****

The schedule for the rest of the semester:
Oct. 8: Bart Selman
Oct. 14: David Liddle
Oct. 22: open
Oct. 29: Eric Sven Ristad
Nov. 5, 12, 19: open
Dec. 3: Brian Subirana
Dec. 10: Andrew W. Moore

If you are interested in giving a talk send email to mdlm@ai.mit.edu.
The Revolving Seminar has a small budget for reimbursing the travel
expenses of senior researchers. If you are interested in a particular
speaker, please let us know. We are particularly interested in
inviting people who espouse views that are not widely represented
within the lab.

-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-

[Martin Taylor 921009 17:45]
(Bruce Nevin 921002 13:14:03)

The theory under discussion holds that the object selected by the vision
system will be the least complex of the available alternatives.
Experimental data supporting the theory will be reported.

This work is based on the pioneering ideas of Solomonoff and Kolmogorov,
and on the more recent ``minimum description length'' concept of
Rissanen.

Interesting. In the mid-70's I was very concerned with exactly such concepts,
and it is a good part of the reason why I harp so much on statistics as an
essential part of perception. I also used the concept of "minimum
description length" as an argument for why Occam's razor was not only a
handy guide, but a nearly infallible rule for favouring one theory over
another, as well as for why we perceive differently using active exploration
as opposed to being presented passively with "sensory stuff." But I didn't
know about Kolmogorov at that time, nor that it was a novel idea. It just
seemed obvious, not really publishable. It still does, and I'm rather
surprised that people can make seminars out of it.

There's also an article on this in the (most?) recent Science, which I don't
have with me so I can't check whether it is the same authorship (Thomas Marill).

Martin