The recursive connection of higher-order maps to the lower-order
maps which feed them can generate "synthetic" sensory input via
"recursive synthesis" (RP p. 88), resulting in effects such as
the illusory contours of the Kanizsa triangle (RP pp. 79
[Fig. 4.4]; BABF p. 39 [Fig. 4-2]). "The possibility of
reentering signals in a recursive fashion to a lower-order mapped
input after they have been processed in several high-order maps
is an enormously powerful way of creating **new** functions.
Since this process of recursive synthesis occurs in various
successive time chunks, it allows for asymptotic approaches to a
great variety of different outputs in complex nervous systems"
(RP p. 69; see also BABF p. 89). Similarly, "crosstalk" among
sensory modalities and submodalities can generate other sensory
"illusions" via "cross-modal construction" (RP p. 88), causing,
for example, a sensation of color or stereo depth to undergo
apparent motion, or illusory borders to appear surrounding
streams randomly-placed dots in motion (RP pp. 79-89).
Edelman emphasizes that the complexity of interconnection due to
reentry in higher brains has no analog among other natural
phenomena or among human artifacts: "A jungle or food web, like
the brain, has many levels and routes for the passage of signals,
but has nothing corresponding to reentrant anatomy. Indeed, if
asked, 'What characteristic uniquely differentiates higher brains
from all other known objects or systems', we would say 'reentrant
organization.' Note that while complex wide-area computer
networks are beginning to share some properties with reentrant
systems, such networks rely fundamentally on codes and, unlike
brain networks, they are instructional, not selectional" (UoC
p. 85; see also UoC p. 49, p. 151 [Fig. 10.1]; ND pp. 147-148
[evolution of reentry in vertebrates and their precursors]).
Reentry is not the same as feedback in the traditional
engineering sense. Feedback is generally carried by a single
unidirectional connection path (versus rentry's multiple,
reciprocal paths), and has the fixed function of supplying a
computed control or error-correction signal (whereas reentrant
connections have no prespecified function or information
content). "Like feedback, however, reentry can be local (within
a map) or global (among maps and whole regions)" (UoC p. 85; RP
pp. 285-286).
The reentrant signalling between a primary map of a particular
sensory modality or sub-modality and other primary maps or
secondary maps (which are simultaneously linked, directly or
indirectly, to primary maps of all other sensory modalities)
enables the brain to spontaneously mirror the spatial and
temporal correlations among features and events in the external
world. This capability, according to Edelman, forms one of the
bases of perceptual categorization. The simplest example of this
is the "classification couple": "a minimal unit consisting of two
functionally different maps made up of neuronal groups and
connected by reentry. Each map **independently** receives
signals from other brain maps or from the world... Within a
certain time period, reentrant signalling strongly connects
certain active combinations of neuronal groups in one map to
different combinations in the other map. This occurs through the
strengthening and weakening of synapses within groups in each map
and also at their connection with reentrant fibers. In this way,
functions and activities in one map are connected and correlated
with those in another map. This occurs even though each map is
receiving independent signals from the world: One set of inputs
could be, for example, from vision, and the other from touch...
This connectivity is not limited to a pair of maps or to any one
moment of time. The interactions of **multiple** maps can be
coordinated in the same fashion" (BABF p. 87).
When maps associated with more than two sensory modalities are
thus correlated by reentry, the result is a "classification
n-tuple": "pair-wise or higher-order interactions of reentrant
maps... form so-called classification couples or n-tuples, which
combine the results of independent disjunctive samplings to yield
categorization" (RP p. 243). This correlative activity obviates
the need for a "projection screen", a "Cartesian theater" (see
Daniel C. Dennett's _Consciousness Explained_; Little, Brown;
Boston; 1991; p. 39), or a supervisor in the brain: "The result
of the selection of appropriate neuronal groups is a covariance,
**not** an image or a sketch... The existence of at least
thirteen separate visual centers in cortical and subcortical
regions, for example, does **not** require a central 'scratch
pad', as is implied by certain... computational theories" (ND
p. 233). "According to this interpretation, the categorical
response of an organism depends to a very large degree upon the
particular physical features of its receptor sheets and their
positions on body parts as developed during evolution. This is
not meant in the banal sense that if cones were not present,
color would not be sensed. Rather, it is intended to imply that
the physical structure of rods and cones or of the cochlear
apparatus, for example, are already adapted by evolution to
abstract major adaptive properties, and that the function of
higher-order neural networks is not to 'compute' but to correlate
these abstracted properties by further reentrant selection among
selected populations of particular neuronal groups, as driven by
motion. Given the possibility of neuronal group selection, with
its rich combinatorial possibilities, the nature of the
transducers and of the relation to the motor ensemble puts by far
the single greatest constraint upon the capacity of such a system
to categorize and generalize" (ND p. 234).
The architecture of reentrant interconnection among cortical maps
thus provides a solution to the "binding problem": "When we
see... a scene, we are not aware of colors, movements, and forms
separately and independently, but bind the color with the shape
and the movement into recognizable objects... How can a set of
diverse and functionally segregated maps cohere without a
higher-order controller? ... A set of models and computer
simulations [UoC pp. 114-120; the model can discriminate and move
a camera to track a red cross in a visual field containing
independently moving red cross, red square, green cross, and thus
binds shape and color]... has shown that binding can occur as a
result of reentry across brain maps that establishes short-term
temporal correlations and synchrony among the activities of
widely spaced neuronal groups in different maps... Thus, reentry
correlates a large number of dynamic circuits in space and time.
The selection of those circuits that are temporally correlated
under constraints of value leads to a coherent output. This
binding principle, made possible by reentry, is repeated across
many levels of brain organization and plays a central role in
mechanisms leading to consciousness" (UoC pp. 106-107).
The automatic correlations of disjoint properties performed by
classification n-tuples solve the problems that would otherwise
result from having to fix the boundaries of an imposed, top-down
categorization scheme: "While the world is not amorphous and the
**properties** of objects are describable in terms of chemistry
and physics... it is clear that, at the macroscopic level,
objects do not come in predefined categories, are variable in
time, occur as novelties, and are responded to in terms of
relative adaptive value to the organism rather than of veridical
description. This lends a relativistic and disjunctive flavor to
the categorization of objects by animals: things are partitioned
according to those factors that are significant for and available
to the perceiving animal. Feature shaping occurs according to
the particular sets of saliences, cues, and contexts presented at
some time and in some sequence. There is a fundamental ambiguity
in classifying objects..." (ND pp. 259-260; BABF p. 233). By
means of reentrant maps forming classification n-tuples, "The
unlabeled world (which follows the laws of physics but in which
biologically adaptive patterns occur that are not described by
physics) is disjunctively sampled by various parallel
sensorimotor channels" (RP p. 243). "Generalization can occur on
the basis of any combination of local features or feature
correlations resulting from disjunctive sampling of signals from
objects not encountered before. The reentry will link the
responses to those combinations to previous patterns of
responses" (ND p. 264 [Fig. 9.5]; see also ND p. 62 [Fig 3.4]; RP
pp. 48 [Fig 3.3], 49-50).
Since the occurrence of a classification n-tuple modifies the
likelihood of the subsequent occurrence of a similar (but
probably not identical) classification n-tuple because of the
plasticity of the synaptic connections among the neuronal groups
comprising it, and since such an n-tuple may subsequently be
triggered by sensory input sufficiently similar to that which
caused the original, possibly in only a subset of the sensory
modalities (or a subset of the input in a given modality) that
originally activated it, these n-tuples also form one of the
bases of memory (BABF p. 103 [Fig. 10-1]). Ongoing perceptual
categorization is therefore always to some extent a
re-categorization, depending not only on current sensory input,
but on an organism's past history of such categorizations. In
other words, the present is always to some extent a "remembered
present" (hence the book title of RP; see UoC p. 78).
Recategorical memory is an inexact and probabilistic re-enactment
of prior similar events, not the retrieval of an exact replica of
a prior event, as occurs in human-made information storage and
retrieval systems (BABF p. 104).
Edelman applies principles of neuronal group selection and
degeneracy to the motor (output) side of the nervous system as
well as the sensory (input) side (see ND Chap. 8). In the
interaction between the nervous system and the musculoskeletal
systems of the body, "the linkage between brain regions and
muscles **cannot** be absolutely strict. This suggests that any
fixed model of control loops for motor behavior is not likely to
be generally correct, at least not for more complex nonautomatic
movements. The units of action are not muscles or joints or
simple feedback loops but functional complexes or synergies
(patterns of movements) in the phenotype that are more akin to
gestures, postures, and their transitions..." (ND p. 220). "By a
'gesture', I mean the degenerate set of all those coordinated
motions that can produce a particular pattern that is adaptive in
a phenotype" (ND p. 227). "[T]he brain deals mainly with
**patterns** of movement. A gesture is one of the degenerate set
of all those coordinated movements that produce a particular
pattern. A closely related notion is that of a synergy -- a
class of related gestures" (RP pp. 120-121). The postulation of
unitary gestures and synergies follows from the consideration of
"[m]ovements of an organism... as morphological objects -- they
develop, react, and evolve in definite epigenetic patterns" (ND
p. 223). During infancy and childhood, "through successive
stages, the locomotor system reorganizes epigenetically, setting
new problems for the C[entral] N[ervous] S[ystem] to which it
must adapt. This process involves costs in energy that must be
minimized" (ND p. 223).
The details of these "synergies", or patterns of movement, cannot
be known in advance any more than the categories of the inputs
that will be received by sensory systems: "[T]he functional
nonunivocality of connections between the CNS and the periphery
are closely related to the notion of degeneracy... [The] sources
of this many-to-one mapping [are] (1) anatomy -- the number of
degrees of freedom in complex kinematic chains, the multiplicity
of action of muscles, the variation of action with the
disposition of limb segments, and the impossibility of the
existence of fixed antagonists; (2) the mechanical complexity of
multisegmented kinematic chains, the activity of which leads to
variant reactive forces following Newton's laws; and (3)
physiological variance in central effector impulses. The overall
conclusion is that motor consequences of central impulses must be
decided peripherally as well as centrally. Even in repetitive
motion of limbs, for example, it cannot be that the brain
**computes** the postural algorithm to accommodate for body
recoil by Newton's third law. Selection must occur in such a
fashion as to bring to lower values the number of degrees of
freedom in the periphery..." (ND pp. 232-233). "Such gestures
must be considered as **patterns to be recognized** by somatic
selection in the nervous system" (ND p. 220). Synergies are
treated by the category-forming interactions in the brain "the
same way sensory systems treat signals from the environment:
categorical and associative relationships can be **built up**
centrally (mainly in the cortex..." (ND p. 230).
To complete the picture of perceptual categorization, Edelman
describes sensory classification n-tuples (comprising multiple
reentrant local maps) as well as classification n-tuples
dedicated to the categorization of gestures, synergies or
patterns of movement (comprising multiple motor maps) as being
linked together in larger-scale structures which he calls "global
mappings" (RP pp. 54-56, 121; BABF p. 89), whose activity is
distributed widely throughout the thalamocortical system, which
correlate motor as well as sensory activity, and which interface
with nonmapped, specialized regions of the brain outside the
cortex (the cortical appendages, organs of succession related to
sequencing of movements and sequencing of categorizations leading
to memory): "a global mapping [is] a dynamic system consisting of
**multiple** reentrant local maps correlating sensory input
**and** motor activity and interacting with nonmapped regions to
form a representation of objects and events" (ND p. 227, see ND
p. 228 [Fig. 8.5]). "Motor activity as reflected in global
mappings leads to the production of synergies, which are
essential elements in categorizing certain kinds of sensory input
as well as in categorizing movement itself. The result of the
combined activities of synergies and tactile and visuomotor
coordination is the selection of neuronal groups in such a way as
to correlate and link the synergies and their transitions...
Because, in forming synergies, sensory sheets are moved in real
time, their linkage is a source of feature **correlation** and
not just of feature **detection**. In this correlation, the
cerebral cortex relates the neuronal groups receiving
proprioceptive input from musculotendinous receptors to groups
receiving visual and tactile inputs, principally in the parietal
cortex" (RP p. 121). This completes the loop connecting the
ongoing sampling of sensory receptors to the movements of the
body that carries them: "Such a global mapping ensures the
creation of a dynamic loop that continually matches an animal's
gestures and posture to the independent sampling of several kinds
of sensory signals" (BABF p. 89). "The construction and shaping
of a global mapping are... driven by ongoing motor activity,
building upon the result of past activities... Such a mapping is
dynamic, and it depends... upon unceasing exploratory movements
and constant motor rehearsals in interactions with the
environment" (ND p. 230).
To reiterate, according to Edelman, a "given classification
couple, or even a classification n-tuple, is in general
insufficient [to yield perceptual categorization]. Instead, a
global mapping... is required... The concept of a global
mapping takes account of the fact that perception depends upon
and leads to action. In this view, categorization does not occur
solely in a sensory area that then executes a program to activate
motor output. Instead, the results of continual motor activity
are considered to be an essential part of perceptual
categorization. This implies that global mappings carrying out
such categorization contain both sensory and motor elements" (RP
p. 54). The motor component of perceptual categorization and of
the global mappings it is based on is, claims Edelman, an
essential ingredient in the ability of an organism exhibiting
primary consciousness to partition the world into objects and
scenes: "[An] essential aspect of categorization is the
**definition of an object** (and, above all, its continuity
properties) by the responding organism... [T]he categorization
of surfaces or textures... is necessary but not sufficient to
define objects, inasmuch as the continuity properties of a
contour are not inherently definable topologically by a static
receptor sheet. There are several ways in which an 'object' may
be categorized via motion or depth cues, however... It is in
connection with object definition that the relation of various
forms of motor activity to sensory input comes to the fore. The
evolutionary development of eye movements, head and neck
movements, and coordinated evolutionary patterns of correlation
of these movements with limb extension are key adaptive features.
They not only prepare animals for appropriate responses but also
serve to relate feature extraction by sensory sheets to the
global feature correlation that accompanies the definition of an
object. This is a key point in relating movement to
categorization" (ND p. 232).
Edelman hypothesizes that the cortex gets specialized assistance
in binding sensory classification n-tuples to gestures and
sequences of gestures to form global mappings and the perceptual
categorizations that depend on them, from morphologically
distinct areas of the brain which he calls "cortical appendages".
"Although the specific ways in which [the] different cortical
appendages interact with the cortex are of central importance,
the appendages all seem to share a fundamental mode of
organization (especially the cerebellum and basal ganglia): Long,
parallel paths involving multiple synapses leave the cerebral
cortex and reach successive synaptic stations within these
cortical appendages and eventually, whether they pass through the
thalamus or not, they go back to the cortex... This serial
polysynaptic architecture differs radically from that of the
thalamocortical system: The connections are generally
unidirectional, rather than reciprocal, and form long loops, and
there are relatively few horizontal interactions among different
circuits... In short, these systems seem admirably suited to the
execution of a variety of complicated motor and cognitive
routines, most of which are as functionally insulated as possible
from each other, a feature that guarantees speed and precision in
their execution" (UoC pp. 43 [Fig. 4.4], 45-46).
One cortical appendage associated with motor activity is the
cerebellum, which operates over the relatively short periods of
time required for the generation of a single gesture or synergy.
The cerebellum "smoothly coordinate[s] the sensorimotor
components of a synergy in short periods of real time in a
feed-forward fashion" (RP p. 121). "For such a task, it must
clearly have a very large sensory input, and both anatomical and
physiological evidence suggests that this is so" (RP p. 122).
"The cerebellum acts in motor adaptations of vestibuloocular
reflexes, saccades, and functional stretch reflexes, and it
regulates reflex gain control. Concerned with the signaling of
errors in movement control, it does not itself generate motor
programs, because, without input, it acts only in 300 msec
intervals or less. But it is essential for timing and
synchronization of smooth movements and for linkages of responses
to novel sensory stimuli... It is **not** connected to the
limbic system in any strong fashion..." (RP p. 138). "[T]he
provision of... temporal coordination is the main role of the
cerebellum, which **together** with the motor cortex (and the
spinal cord) contributes to global mappings that allow
registration, linkage, and smooth succession of movements. Such
a succession is a critical part of the categorization of motion,
particularly in novel tasks and situations. It is also an
essential part of the workings of the classification n-tuples
dedicated to the categorizations of gestures themselves" (RP
p. 121).
The "basal ganglia" (a group of structures, including the
striatum, globus pallidus, and substantia nigra, treated
collectively (RP p. 135 [Fig. 7.3]) constitute the other cortical
appendage involved with movement. "These are a large and complex
set of structures located deep at the center of the brain. They
connect to the cerebral cortex in a series of parallel circuits
involved in eye and body movements, as well as to the frontal
portions of the cortex, the function of which is related to
behavioral planning and emotions" (BABF p. 106). While Edelman
acknowledges that his hypothesis about the role of these organs
is to some degree speculative (RP p. 133), he believes this
cortical appendage to be concerned with the sequencing of motor
programs and the choice and execution of motor plans, over time
periods longer than those associated with the cerebellum (300
msec to several seconds [RP p. 139]), and under the influence of
motivational states due to input from the limbic system and brain
stem (RP p. 137). Motor programs are "sets of muscle commands
put together before the beginning of a movement sequence. They
permit that sequence to be carried out without peripheral
feedback..." (RP p. 133). Motor plans are larger behavioral
units entailing sequences of motor programs: "[Motor programs]
are linked by the motor system into complexes that can be
affected by planning, thought, and intention" (RP p. 133). "[The
basal ganglia] not only help regulate movement in a motor program
by coupling sensory and motor responses but also help direct what
is to be done according to a motor plan. Notice that, unlike the
cerebellum, which smooths and coordinates more immediate
gestures, this appendage works over longer time scales and helps
correlate whole sequences of gestures in a plan... The basal
ganglia are also intimately connected to the hedonic centers of
the brain, and... they very likely play a role in attention (BABF
p. 106). Following Edelman's logic so far, one assumes that
motor plans are also categorized in the cortex by means of
classification n-tuples, and that these highest-level motor
complexes also participate en bloc either in global mappings
corresponding to perceptual categories or in the classifications
of global mappings corresponding to conceptual categories (see
below).
Selection at the level of perceptual categories (global mappings)
is constrained by values: "[C]ategorization always occurs in
reference to internal criteria of value and... this reference
defines its appropriateness. Such value criteria do not
determine specific categorizations but they constrain the domains
in which they occur... [T]he bases for value systems in the
animals of a given species are already set by evolutionary
selection. They are exhibited in those regions of the brain
concerned with regulating bodily functions: heartbeat, breathing,
sexual responses, feeding responses, endocrine functions,
autonomic responses. Categorization manifests itself in behavior
that appropriately fulfills the evolutionarily selected
requirements of such life-supporting physiological systems" (BABF
pp. 90-91). "[T]he driving forces of animal behavior
are... evolutionarily selected value patterns that help the brain
and the body maintain conditions necessary to continue life.
These systems are called homeostats. It is the coupling of
motion and sensory sampling resulting in behavior that changes
the levels of homeostats. Aside from those occasional
species-specific behavior patterns that have been selected for
directly by evolution, however, most categorization leading to
behavior that changes homeostatic levels occurs by a **somatic**
selection of neuronal groups in each animal. Categorization is
not the same as value, but occurs **on** value. It is an
epigenetic developmental event, and no amount of value-based
circuitry leads to its occurrence without experiential selection
of neuronal groups. But it is also true that without prior
value, somatic selectional systems will not converge into
definite behaviors" (BABF p. 94).
Up to this point, the global mappings linking classification
n-tuples dealing with sensory input and motor output have all
been concerned with immediate interactions with events and
objects in the world outside the animal's skin. As the next step
to consciousness, Edelman postulates that the cortex is capable
of forming classification n-tuples which categorize its own
global mappings (without going into great detail about the
machinery by which this takes place, other than to lean heavily
on the characteristic cortical morphology of reentrant maps).
Edelman calls these categorizations of categorizations
"concepts", employing his own non-linguistic definition of that
word (RP Chap. 8; BABF pp. 108-110). "[I]n forming concepts, the
brain constructs maps of its **own** activities, not just of
external stimuli, as in perception... [T]he brain areas
responsible for concept formation... categorize, discriminate,
and recombine the various brain activities occurring **in
different kinds of global mappings**... [They] categorize parts
of past global mappings according to modality, the presence of
absence of movement, and the presence of absence of relationships
between perceptual categorizations... Structures able to perform
these activities are likely to be found in the frontal, temporal,
and parietal cortices of the brain. They must represent **a
mapping of types of maps** (BABF p. 109). "To succumb to the
temptation [to consider concepts as properties of language] I
just warned against, one might suspect that **if** an animal with
concepts could speak, it would reveal its concepts as ontological
categories -- things, motions, and classes... In their most
elaborate form, concepts may serve as bases for image schemata
('object', 'motion', 'barrier', 'container', etc.) summarizing a
variety of general physical situations" (RP p. 141). Concepts
are a form of long-term memory: "The categorizations and
generalizations reflected in concepts are much richer, more
abstract, and therefore less immediate than those seen in
perceptual categorization. They may involve the memory of much
earlier or more distant events... and less than immediate
reference to various objects. Concepts may involve judgments
made in absence -- judgments made ahout classifications of
stimuli long past" (RP. p. 142).
The cortex does not form concepts directly from the "raw" global
mappings being activated in the immediate present, Edelman
hypothesizes, but instead relies on a third specialized cortical
appendage to gather input from currently-activated global
mappings in the cortex, edit and sequence them (based partly on
value signals from the limbic system), and replay the edited
stream, as a sort of home movie of the recent past, back to the
cortex. This cortical appendage is the organ of succession known
as the hippocampus: "[T]he hippocampus... receives inputs from
practically **all** regions of the cerebral cortex, through a
smaller region known as the entorhinal cortex. These inputs run
through the hippocampus in a sequence of three successive
synapses. Having passed through these structures, the signals
loop back through the entorhinal cortex and are relayed back by
reentrant fibers to the cortical areas that originally connected
with it. Cells inside the hippocampal loop all receive
simultaneous connections indirectly from the midbrain and hedonic
areas, areas subserving value" (BABF p. 106; see also RP
pp. 127-133). "[W]hile classification couples and global
mappings undergo neuronal group selection by synaptic change,
this in itself is not enough to assure the relationship between
short-term and long-term memory. Unless organs of succession
such as the hippocampus intervene and order the results, it
appears that severe memory defects will ensue" (BABF p. 107).
"The time units of succession are roughly (1) those of cortical
excitation patterns occurring in a period of 10 msec, the basis
of ongoing perceptual flux; (2) **sequences** of such events from
100 msec to 1 sec, corresponding to the integration of perceptual
categories; and (3) **chains** of such sequences occupying longer
times and related to learning and fixation of secondary synaptic
strengths in the cortex. It is important to emphasize that the
ordering of time chunks (i.e., of signals represented in global
mappings) is not a simple reverberating circuit. Overlap,
divergence of inputs, and mixing of signals prevent such a simple
circuit property from emerging" (RP p. 132). "While necessary
for long-term memory, [the hippocampus] is not likely to subserve
it; rather, changes in cortical synapses probably serve this
role" (RP p. 139).
This archive was generated by hypermail 2b29 : Thu Jul 27 2000 - 14:13:51 MDT