[From: Bruce Nevin (Wed 920424 08:22:46)]
I suspect that there is more than one sort of thing being called
adaptation. I have proposed before that one of these is a change in the
value of a reference signal, basing it on short-term memory more than on
long-term memory. This would be real adaptation. Another sort of thing
might be reduction in gain. Later repetitions of a word in discourse
are produced with less careful articulation than the first occurrance,
Martin tells us (a fact that, to my delight, fits very well with the
Harrisian notion of reductions). With a reduction in gain there might
be a statistical tendency but not the individual consistency that is the
hallmark of control.
There is always the possibility of boredom, listening to 70 repetitions
of a synthesized while (whilst) clenching ones tongue between one's
teeth.
The idea of adapting the reference signal has an interesting twist for
language learning. We do accomodate small shifts of dialect without
noticing, preserving contrast of words; we also resist large dialectal
differences at a different level of control, the contrast between the
kind of person who speaks noticeably that way and the kind of person
with whom I am most familiar. If it were possible to vary the relevant
parameters for a difference in pronunciation (say English vs. Spanish
VOT) in ongoing dialog with the student, the student might "adapt" to
the new reference values in graduated steps, never encountering a
differential sufficient for control on the self-image level. One can
imagine a SF world, where the student was immersed in an environment
with androids whose reference signals were able to be tuned gradually in
this way. I'm not sure if it would be possible for a human instructor
to make the gradual adjustments in a consistent way--sensitive always to
the student's current and changing settings of those signals, and while
carrying on natural and engaging conversation.
(Martin Taylor 920623 21:15) --
Replying to me (920423 08:32:46)
One possibly fruitful way to approach this is by attributing
differentiating perceptions to the individuals in memory and
imagination.
...Suppose we have ECSs (elementary control systems) controlling for ways
to differentiate between like perceptions by identifying other
perceptions associated with them as attributes. Is there any reason
this is implausible? In such a way we might develop control of category
perceptions.One way to track four dots that soon becomes intractible is using their
history ("the one that started in the northwest corner"). An ECS alert
for differentia might entertain various hypotheses ("the bounciest one")
and develop a history of them (". . . is tired now"). A story, not in
words but in remembered perceptions. ...
I find this approach highly plausible and natural within the HPCT
structure. Does it answer your own question? I think it answers mine,
which was supposed to mirror yours.
Yes, I hope so, and thank you for taking up the question and
rearticualting it so well.
Now, what might a difference-detector ECS accept as input? Two signals
from an ECS, one associated with one set of signals from other ECSs,
another associated with an intersecting but different set. A
discontinuity in the signal or signals that stay the same, reflecting a
discontinuity in sensory input or a shift of attention. (Attention
involves at least the real or imagined focus of sensory organs on a
target and the adjustment of gain on selected ECSs, as for example
tuning out extraneous noise for conversation in a crowd.) The
discontinuity marks a shift from one individual to another, unless the
difference-detector itself reports continuity (no differences that make
any difference).
a m p a b c
\ | / \ | /
b -- X -- q m -- X -- n
/ | \ / | \
c n r p q r
Individual Individual Individual Individual
A = X plus B = X plus A = X plus B = X plus
{abcmn} {pqrmn} {mabcpqr} {nabcpqr}
Here are two hypothetical cases of "detecting an X" plus associated
perceptions {abcmnpqr} that may be present. X is a category perception.
My proposal regarding category perceptions is that they develop from the
association of e.g. {abcmn}, {pqrmn}, {mabcpqr}, {nabcpqr} as distinct
individuals by a difference-detector. The difference-detector then
constitutes the category perception out of these perceptions of
attributes. Any association-set like {abcmn} or {pqrmn} (or in the
second hypothetical case, any association-set like {mabcpqr} or
{nabcpqr}) indicates that an X is present. Were one to define an X, one
would say anything that has attributes {mn} (in the first case), with
other attributes {abc} or {pqr} optional. (In the second case, the
defining attributes would be {abcpqr} and the variability would be
between m and n.)
According to this proposal, there are difference-detectors, and there
are associations of perceptual signals (attributes) in associative
memory, but there are no category detectors per se. Does this still
sound feasible and sensible?
Bruce
bn@bbn.com