on projection and hallucination

[Hans Blom, 950829b]

ON INTERNAL MODELS, MODEL-BASED CONTROL AND INFORMATION THEORY

A quote from E.H. Gombrich's book "Art and Illusion" [1956?].
Gombrich (born 1890) was a British art historian and aestheti-
cian, who has significantly influenced recent understanding of
how we perceive the visual arts. This quote sheds some light on
the sometimes overwhelming power of internal models, on
adaptive control, and on the relation between adaptive control
and information theory. The additions in square brackets are
mine.

   "[During the second world war] I was employed for six years
by the British Broadcasting Corporation in their "Monitoring
Service", or listening post, where we kept constant watch on
radio transmissions from friend and foe. It was in this context
that the importance of guided projection [Gombrich's term] in
our understanding of symbolic material was brought home to me.
Some of the transmissions which interested us most were often
barely audible [i.e. there is significant observation noise],
and it became quite an art, or even a sport, to interpret the
few whiffs of speech sound that were all we really had on the
wax cylinders on which these broadcasts had been recorded. It
was then we learned to what an extent our knowledge [internal
model] and expectations [predictions by the internal model]
influence our hearing. You had to know what might be said [you
had to have an accurate internal model] in order to hear what
was said [in order to be able to disregard the observation
noise]. More exactly, you selected from your knowledge of
possibilities certain word combinations and tried projecting
them into the noises heard. The problem then was a twofold one
-- to think of possibilities and to retain one's critical
faculty [to combine the predictions of the internal model with
the actual perceptions]. Anyone whose imagination ran away with
him, who could hear any words -- as Leonardo could in the sound
of bells -- could not play that game [relying on the internal
model only does not work]. You had to keep your projection
flexible [admit uncertainty], to remain willing to try out
fresh alternatives [prevent premature convergence], and to
admit the possibility of defeat. For this was the most striking
experience of all: once your expectation was firmly set and
your conviction settled, you ceased to be aware of your own
activity, the noises appeared to fall into place and to be
transformed into the expected words [convergence is reached,
but to incorrect model parameter values]. So strong was this
effect of suggestion that we made it a practice never to tell a
colleague our own interpretation if we wanted him to test it.
Expectation created illusion.

   "While I was struggling with these practical tasks, I did
not know that these problems of transmission and reception of
communication -- terms such as "message" and "noise" -- were
destined to become a most important, not to say fashionable,
field of study under the name of "Information Theory". The
technical and mathematical aspects of this science will always
remain a closed book for me, but my experience enabled me to
appreciate at least one of its basic concepts, the function of
the message to select from an "ensemble of possible states"
[or, in a non-discrete model, to select a value from a range of
values]. The knowledge of possibilities in the monitor is the
knowledge of the language and the contexts in which it is used.
If there is only one possibility, his receptor apparatus is
likely to jump ahead and anticipate the result of what William
James called the slightest "auditory hint". But it also follows
from this theory that where there is only one such possibility
the hint is in itself redundant and there is, in fact, no
special message [the observation corroborates the model's
prediction]. The word we must expect in a given context will
not add to our "information". We receive no message in the
strict sense of the word when a friend enters a room and says
"good morning" [if he always does so]. The word has no function
to select from an ensemble of possible states, though
situations are conceivable in which it would have [if he said
it differently, or at a different time].

   "The most interesting consequence of this way of looking at
communication is the general conclusion that the greater the
probability of a symbol's occurrence in a given situation, the
smaller will be its information content [the more accurate the
model's prediction, the less we need the observation]. Where we
can anticipate we need not listen. It is in this context that
projection will do for perception.

   "The difficulty in distinguishing between the two in seeing
as well as in hearing was well brought out in a fiendish
experiment. The subjects were seated in the dark in front of a
screen and were told their sensitivity to light was to be
tested. At the request of the experimenter, the assistant
projected a very faint light onto the screen and slowly
increased its intensity, each person being asked to record
exactly when he perceived it. But once in while when the
experimenter made the request no light was, in fact, shown. It
was found that the subjects still saw it appearing. Their firm
expectation of the sequence of events had actually led to a
hallucination."

Questions:

1. How can we -- if we can do so at all -- discriminate whether
we project our internal model upon the world rather than adapt
our internal model to the world? In which conditions are we
most likely to project rather than to process?

2. How can we -- if we can do so at all -- discover whether we
hallucinate? In which conditions are we most likely to halluci-
nate rather than to perceive?

3. What could it mean, in our daily lives, to KNOW that we,
sometimes or frequently, project and/or hallucinate?

Greetings,

Hans