[Martin Taylor 931025 15:05] Federal election day
I will probably not see responses to the following until next Friday or Monday.
I post this without having seen any postings since last Friday, so please
forgive failures to take note of concepts that may have been developed since
then.
The following is a semi-formal presentation of a different conception of
the upper part of the perceptual control hierarchy, intended to resolve
some of the issues that have plagued the category-contrast debate by
going around them.
···
=============
Category and contrast
A restructuring of the perceptual control hierarchy
The following rethinking of the relationship between the logical and the
analogue parts of the perceptual control hierarchy is intended as a straw
man. It is based on restating the concept of "contrast" between "categories"
as a process rather than as a relationship or a perception. I neither
believe nor reject it, though it has some attractive properties. All
comments welcome.
Several participants in the discussion on contrast and the discussion on
category that followed it have suggested that category is something that
happens at every perceptual level, rather than representing in itself a level
of perception. Bill P. has proposed that categorization is something done in
the perceptual input functions of logic level ECSs. Since "level-skipping"
of perceptual signals is not prohibited in the "standard" hierarchy, Bill's
proposal allows for categories to be based on any perceptual level.
Certainly, when we label categories, the things that are labelled seem to
come from all levels of the hierarchy: "bright," "blue," "quick," "running,"
"democracy." So the suggestion that category is something related to every
level is a priori plausible.
The standard hierarchy treats "category" as a level above the set of
analogue levels, or perhaps is itself the highest analogue level. Either
way, it interfaces between the highest analogue level and the lowest logic
level.
I would like to propose an alternate view. In this view, category is indeed
an interface between analogue perceptions and logic perceptions, but is not a
level in the hierarchy. Instead, it relates an analogue hierarchy with a
second, digital, hierarchy. Category is defined not by a place in the
hierarchy, but by a process. That process is "contrast," and the contrast
process can occur between any two or more perceptions at any single level of
the hierarchy. Perceptual signals related by a contrastive interaction,
described below, become suitable sensory inputs to a logical perceptual
control level.
In a previous posting (reference not available here at home), I proposed
that two contrasting PIFs at the category level might be connected in a flip-
flop relation. I now propose that the flip-flop connection defines the
"contrast" process and thereby the existence of a category.
In a flip-flop relation, we consider two ECSs, A and B, the perceptual signal
of each being provided as sensory input to the PIF of the other (later, the
same relationship is shown to applicable to groups of more than two ECSs).
The output of PIF-A is amplified and provided as an input to PIF-B, and vice-
versa. When the output of either is high, it tends to drive the output of
the other low. Such a relation corresponds to the normal interpretation of
the existence of a contrast between perceptions A and B. There is no
contrast between the category of a visible chair and a smelled cigarette, nor
between a visible chair and a visible antique (both being the same object).
Contrast is normally said to exist where the perception of category A implies
the non-perception of category B and vice-versa.
Here, again, is a diagram of the flip-flop relation. I think it worth
repeating, because it has some interesting behaviour. Each PIF is supposed
to be a component of an entire ECS, not shown in the diagram.
PA | | PB
^--->--\ /--<---^
> \ / |
PIF-A \ / PIF-B
/| amp \ / amp |\
/ | \---<--\--/ / | \
other \---->-/ other
inputs inputs
The "amp" in each cross-link could easily be shown as part of the PIF, but is
separated in the diagram because it is important on its own. The sign of the
gain of "amp" determines whether the occurrence of PIF-A aids or suppresses
perception of B and vice-verse, and its magnitude determines by how much this
occurs. The same structure could serve "assocation" (when "amp" makes PA and
PB rise together) as well as "contrast" (when "amp" makes PA tend to suppress
PB and vice-versa).
Both association and contrast are seen as processes rather than perceptions.
The process of contrast is the suppression of perceptual signal PB as a
consequence of inputs to PIF-A that make PA go high; the process of
association is the enhancement of PB when PA goes high.
If the gain of "amp" is set to zero, the two PIFs operate independently.
As they are neural elements, the magnitude of the perceptual signals PA and
PB cannot go negative. PA and PB are positive magnitudes that depend on how
the perceptual input function responds to the signals marked "other inputs."
The linkage for association or contrast is trivially extended to groups of
more than two PIFs. Since none of the outputs PA, PB,... can go negative, it
is quite possible for all of the members of the group to have zero output,
but if the gain of "amp" is inhibitory and high, only one of the group can
have a high output at any moment. The members of the group all show mutual
contrast. If the gain of "amp" is excitatory and high, none or all of the
group members will have high (probably saturated) output. Effective
association occurs when the gain of "amp" is positive and low (<1). In this
case, the perception of one member of the associative group would enhance the
probability of perceiving another member of the group.
The revisionist element I am proposing in this posting is that the PIFs in a
contrastive relationship, apart from their cross-linked inputs, can be of any
kind in the whole hierarchy, rather than being of one specific level. But
the next higher level may be logical, no matter what their level. The PIF
for the logical level is likely to have as sensory input the perceptual
signals from all the ECSs that are in the contrastive group (A and B in the
example, but such groups can be larger), whereas each could still
independently provide its perceptual signal as input to higher analogue
levels.
What this means is that the logical structure (including program, system, and
principles) should be conceived as lying beside the main analogue hierarchy
rather than on top of it. Program fragments such as "If intensity of red
light greater than green light and event is driving toward rather than away
from intersection ..." make sense if one sees category construction as
occurring at each level of the analogue hierarchy equally. They are harder
to deal with if there is a separate "category level" that interfaces between
the lowest logical level and the highest analogue level.
The outputs of logical level ECSs provide, eventually, reference levels
for category ECSs ("I want to see a green light, rather than a red one").
The outputs of category ECSs must provide references for analogue, not
logical, ECSs, since the categories are the consequence of the process of
contrast operating within the analogue hierarchy. The contrast PROCESS
provides the link between the analogue and the logical worlds at all levels
of the analogue hierarchy, but the perceptions that are categorized are
within the analogue hierarchy. They may nevertheless take on nearly binary
values, intermediate values between zero and saturation being highly
improbable if the gain of "amp" is high.
If the gain of "amp" in the contrast process is high one of the PIFs in a
contrastive group is likely to be giving a saturated output while the others
are giving perceptual signals near zero. Only if the "other inputs" would by
themselves give small perceptual signals that were quite similar across two
or more of the group, and at the same time the gain of "amp" were low, would
there be appreciable perceptual signal in the ECS of more than one of the
group at any single moment. Ordinarily, the perceptual signals within a
contrastive group should be expected to be nearly binary, "on" or "off," only
one member of the group being "on."
The association process is identical to the contrast process, except for the
sign of the "amp" connection in the figure. Association provides a linkage
among perceptions that could be of wildly different kinds in the analogue
hierarchy (such as between the sound of a word and the sight of the
corresponding object). In an associative group, the gain of "amp" cannot
afford to be very high, or else the entire group would be permanently
switched on to saturation as soon as any one of them received adequate "other
input." Association requires low positive feedback, whereas contrast works
well with high negative gain across the PIFs. Association does not link the
analogue hierarchy with the logical. Contrast, which produces category,
does.
=================
As an extension to this concept, I would propose that outputs from the
logical level affect the gain of the cross-link amplifiers "amp," as well as
the reference levels in the contrastive group of ECSs. When a contrast is
required, the gain of "amp" is increased, forcing a choice to be made between
the contrasting entities, whereas when the logical level does not require a
contrast, it provides signals that reduce the gain of "amp," leading to mild
or non-existent inhibition between PIF-A and PIF-B. When the gain of "amp"
is low, each PIF responds mainly to its "other inputs", and a continuous
change in the values of those inputs generates a continuous change in PA and
PB. When the gain of "amp" is high, continuous change in the "other inputs"
can lead to abrupt, discontinuous change in PA and PB simultaneously, only
one being high at any moment. The border at which this change happens
necessarily shows hysteresis, to a degree that depends on the gain of "amp."
As I said, I find this structure attractive, partly because of its
simplicity, but more because it leads directly to a functional architecture I
inferred to exist in support of the reading process (Chapter 10 and 11 in The
Psychology of Reading, Academic Press 1983), in which a logical hierarchy
parallels a massively parallel pattern-matching hierarchy in mutual support.
I then called the two hierarchies "tracks," LEFT and RIGHT. The LEFT
(logical) track operations seem largely to be carried out in the left
hemisphere, though the RIGHT track operations are performed in both
hemispheres more or less equally.
To recap, I propose that the concept of "category" is defined by
"contrast," a process of mutual inhibition between certain perceptions.
Category perception can occur at any level of the hierarchy, and provides an
interface to a logical hierarchy that exists not on top of, but at the side
of the analogue perceptual control hierarchy. The same process that
generates contrast can also generate association, in which the existence of
one perception enhances another.
I would be very interested to pursue the implications of these concepts, and
anticipate that if we do so, we will wind up with something rather different
from what I proposed above. We may wind up determining that it makes no
sense at all, and that the "standard" category level corresponds more closely
with practical necessities and introspections about perception. But at first
sight treating "contrast" as a process that generates categories at all
perceptual levels rather than as a perception related to a specific
"category" level does seem to handle many of the problems raised in the
contrast-category discussions of the last few weeks.
Martin