AN ADAPTIVE,
EVOLVING NETWORK
CONCEPT
PAPER
[MAVIN:
Machine Adam Virtual Information Node]
“We are not
stuff that abides, but patterns that perpetuate themselves.”
Norbert
Weiner, The Human Use of Human Beings
The following is a concept for creation of
an adaptive, evolving network mimicking biologic processes. The network would rely on a simple organizing
principle from which complex behavior and organization would emerge. The basic organizing principle upon which the
concept relies is the attainment and maintenance of system stability, stability
being envisioned as a basic organizing principle of both living and non-living
systems. For programming purposes,
stability is to be defined in terms of the number of relationship patterns
between node and system data elements.
The network is envisioned to consist of multiple interconnected nodes,
each node to be programmed with a threshold level of stability which it must
maintain. Individual nodes exist in a
connectivity relation with their "data field," or areas of sensor
data input, as well as other nodes.
Inter-node communication consists of feedback and feed-forward loops by
which inter-nodal pattern matching information is exchanged and attainment or
loss of individual node stability messages are sent to other nodes. Upon reaching the threshold level of
instability (number of unrelated data elements in relation to the existing node
data base), individual nodes send an "instability" message to all
other nodes to which they are connected.
Upon receiving the instability message, external nodes are connected to
the unstable node’s data base and begin to query their own data bases in search
of data relations between their data and the unstable nodes data. As these relations are found, related data
elements are forwarded to the unstable node until it reaches a predefined
stability threshold. Upon re-attainment
of system stability, the originally unstable node then sends a second signal to
the other nodes with which it is connected to signal them to recommence normal
internally stabilizing functions. The
central node stores inter-nodal information relationship patterns to create a
field system mapping.
Threshold levels of stability or
instability may be defined in terms of a percentage of the number of relations
to the total data field size. Threshold
levels of stability assignments per node might be defined in relation to the
importance of the individual node in overall system stability, the system being
defined as the sum total of all of the individual nodes and their data bases
(total system relations attained or determined/total system data). A central node would control system
stability, selectively shutting down non-essential data nodes depending on the
level of system instability. The central
node would also store holistic system patterns and contain a mapping or
addressing from pattern memory to contributing nodes. Lower threshold signal strengths would be
required for inputs from nodes having significant contributions to overall
system patterns, signal strength being adjusted based upon the degree of
“particle resonance” exhibited by outputs from outlier nodes to the central
node (relationships formed by inter-nodal pattern matching). The initial state of the central node would
be set at a normalized required signal strength across
all outlier nodes, threshold signal strength addressing being developed as
outlier nodes compete, and this competition being category specific so that a
mapping from the central node to outlier nodes develops specific to information
pattern categories.
The
following definitions of basic terms are provided.
Basic
system construct = system instability elicits node/system actions to achieve
the system goal.
System
Goal = stability
Stability
= information relationships
Relationships
= repetitive coincident events or patterns
Positive
Value = input or activity increases overall system stability
Negative
Value = input or activity decreases overall system stability
Instability
= lower threshold of repetitive coincident events, patterns/total data elements
(system and/or node)
When
the threshold level of instability is reached, the system enters a state of
"heightened vigilance" in search of relations between information elements. This heightened vigilance may incorporate
several tasks;
1. Search of system data to increase node data
relatedness/identify previously unrecognized relationships or relationship
patterns.
2. Passive or Active Selective Attention to
sensory data with a higher probability of success in finding relations.
- Passive Selective Attention – no
sensor node/sensor data field interaction
- Active Selective Attention – may
entail the need for the data system to position the sensor to better advantage
within the respective data field (change the degree of view of that which is
being observed).
3. Active Searching – involves actions such as
probing the data field or other actions wherein the sensor system searches for
elicited responses to system actions.
4. Functions previously created through the
observed data, and subsequently turned on that data using it as substrate,
rearrange protodata into new relations (the system
acquires not only new facts (datum), but also new ways of looking at those facts
– what emerges can affect what it emerges from). The functions created or derived from
previously observed repetitive coincident events can act as a type of spell
checker on expectation and reality. When
previously repetitive coincident events no longer coincide, as determined by an
error between the new data, the old function based on proto-data, and reality,
the function becomes hyperactive (hyper-vigilant), triggering the search for
emergent relations and, as they are found, the development of new functions
based on new data relations, and the revaluation of the old functions.
Data relationship definitions would be
based on biological data structural relationships as depicted in figure 1
below. Specified triplet numerical
sequences would relate to specific letters/codons,
letter sequences to words/symbols, word sequences to phrases/strings, phrase
combinations to concepts, and concept combinations to functions. Initial/seed individual data node inputs
could run the data spectrum from simple quantitative data to functional data,
the quantitative data being a representation of an underlying function
(preformed relationship in the data field).
No pre-established database would be provided. Rather, the system would learn through
recognizing relationships within each nodes own
experience as well as the experience of other nodes and sensor fields, applying
this "understanding" to create additional relationships to further
increase stability. The discovered complimentarity of information from other nodes provides a
new informational figure/pattern from which to push and stack from within the
individual node data structure, as well as a new epi-informational
function to express lower level relationships (environmental feedback triggers
expression of previously dormant information by providing new context for
conceptual nearness of lower level symbols/patterns).
Figure
1
Douglas
R. Hofstadter makes this same comparison, offering an analogy to biological
system transcription, translation, protein manufacture and function using
Typographical Number Theory and the Gödel Code in his book Gödel, Escher, Bach,
figure 100, page 535, describing a Gödel Code for a similar process. Figure 2 below depicts his number to codon matrix which could serve as the basis for creating
the mathematical genetic code of the network.
Figure 2
Biologic
and physical system analogies can be carried further as follows. Biologically, each node could be
representative of a specialized cell type, the inter-nodal messages to be
hormonal messengers (a substance [relation] produced in one part of the body
[data system] having an impact on another part of the body [data system]), and
the functions produced by the relationship patterns to be enzymes. The last, the functions produced by the
underlying relationships, could be such that they are active functions, as they
are created being run or acting on their data "substrate," that
substrate being a particular data sequence within the node. Random variation in the underlying genetic
code would produce function variation which would be selected for continued
expression based on its ability to efficiently and effectively identify
relationships in the data field. The
population of nodes would compete as overall system stability is a system goal,
the central node selectively shutting down those non-essential nodes creating
system instability, the individual nodes shutting down functions ineffective in
creating node stability. The collective
nodal processes create the epiphenomenon of a network “mind” which acts on the
nodal processes in a feedback loop to change them through the manipulation of
information based on knowledge obtained from and driven by the utility
function. In this way, the mind moves
matter, directing the characteristics, activities, and capabilities of the
network, the associative nodal state being akin to a mental state. Physically, what is entangled is not
therefore primarily matter but rather information, the physical being an
epiphenomenon or phenotype of the information and, more, an epiphenomenon which
interacts with its genotype through complex feedback loops. The evolution of information entanglement
throughout the system becomes the “hidden variable” which describes the
probabilities of relationship identification and development within and between
nodes. The evolution of the wave
function becomes the evolution of information entanglement throughout the
system. The choice of which question to
pose to nature becomes the “patterns in the brain mirroring the brain’s
mirroring of the world.” (GEB) Nature’s statistical choice of which answer to give becomes the
system’s choice of its most stable pattern given the universe of possible
relationships to form and given its utility function.
Figure 3
3 Dimensional arrangement of the information based on all
information inputs at time T1. The
blue represents potential new information in the input pattern obtained by
comparison with node/system data patterns.
Such new information could be an identical pattern or isomorphic as
shown.
Concepts (3 Protein Strands)
● ● ● ● ●
Time Tn Space Sn ●
Figure 4
The
total system would consist of multiple sensor data fields as figures 3 and 4,
each field connected with the others via a common interface for data exchange,
the inter-nodal instability messages applying both internally between sensor
data field nodes as well as externally between sensor data fields. Figure 5 below depicts one sensor field set
up using the ART example, the central node here being equivalent to the F2 node
in the ART system storing that fields long term memory, a system wide central
node with this schema being expanded to the system level being envisioned. An example of a relationship pattern would be
the repetitive appearance of a particular sensory data input with another
sensory data input, some threshold level needing to be determined to separate
causal from chance occurrences (noise) of coincident inputs. Within an individual node, and prior to input
of relationship data from external nodes, a relationship pattern would be the
repetitive coincident appearance of data from sensory nodes (for instance, the
sound of a train whistle coincident with the sound of the wheels running on the
tracks). Repetitive coincident
information from different sensor types or from different inputs of one sensor
type would serve to bind the information, similarly to the wiring together of
neurons which fire together. Data
elements would be time and location stamped or tagged to determine inter-nodal
correlations of sensory data.
Figure
5 – Sensor Field 1 Standing wave (second sensor field)
The
system mapping of the information patterns from the sensory field inputs becomes
a function (a unique mapping from a domain space x (sensor fields) into a range
space y (the system information topology).
Transcription of input data sequences and storage in memory would create
an isomorphic data set of the input information. Random variation in the transcribed input
sequence can generate new potential mappings to increase relatedness
(stability). Such variations would be
compared to inputs across the system to determine coincidence with field inputs
(do these patterns which could increase my stability exist somewhere in domain
space x – the real world?) Such activity
would be akin to the system thinking about itself as opposed to simply
recording stimulations from the environment.
If one considers the 3D arrangement of information as in Figure 4 an
information particle, its spatial coordinates are its mapping to its local
field map and its higher system map.
Specifying all of its coordinates corresponds to specifying its spatial
relationships to the entire range space y – node and system range space. There exists a mapping or activation
probability across the range of potential linkages (vertices and edges) within
the system topology and multiple potential particle locations which correspond
to particular activation patterns. As
information relationships increase stability which decreases energy, we can
equate the total potential activation patterns to an effective potential energy
surface and the specific patterns to potential energy wells, or the regions
surrounding the local minima. As in
Adaptive Resonance Theory (ART) systems, resonant behavior is equated to
information relationships. The wave
function describes the probabilities of relationship development within and
between nodes. Oscillatory behavior corresponds
to the particle being transiently trapped in a potential energy well (a local
minima). Alternative minima exist
throughout the range space. Oscillations
are produced by the system transitioning to innumerable potential information
relationships, the system seeking its most stable energy state. That state is equivalent to the greatest
number of edges between vertices (relationships) throughout the system at any
particular time. The closer the system
is to the threshold value of instability, the higher the oscillation frequency
as the system becomes hyper-vigilant in seeking to lower its energy state. The information “particles” from an
individual sensor field combine to form an information “atom” as shown in
figure 5 above. (As an alternative view
to physical nodal connection, and as a concept linking pure information to
matter, consider the following. These “atoms”
resonant behavior produce outgoing oscillations/waves which interfere with
incoming oscillations from other sensor fields producing wave interference
patterns which characterize each sensor field’s relationship to other fields. These patterns are what are captured by the
R/EM sensors. Stability, in this sense,
is achieved through the creation of standing waves as the “atom’s” oscillatory wave,
produced by its internal relationship patterns, relates to/interferes or
resonates with incoming waves from the system.
The resulting standing wave
creates shells of matter as in the standing wave theory of matter.)
The
following quote was written after reading portions of GEB. “The feedback loops would create sentient
systems, the individual unitary elements of the system possessing no sense of
self or direction with a purpose. It is
only in association with other elements that the laws governing their social existence
can be felt in their impact and reacted to.
It is through the numerous, countless feedback loops that individual
elements, as part of a complex system, sense their relation to that system and
themselves as a unitary part of it.”
Stability equates to harmony which equates to a minimum energy
state. The harmony is achieved through
known relationship patterns with both the internal and external
environment. Disruptions in those
environments create disharmony, or a disruption in the
relatedness of the system parts and therefore a higher energy level due to the
active search for new relationship patterns due to the prime system
directive.
The
following ideas were written after reading “The Intelligent Universe” by James
Gardner. “The utility function =
information relationships. Nature
selects that which is in right relationship with its environment. That is in right relationship with its
environment which 1) through random variation or 2) through knowledge is so
ordered or has so ordered itself or its environment to achieve stability. Random variation and natural selection create
the nascent forms which gradually acquire the ability to so order themselves to
achieve stability – adaptability – and eventually the ability to control and
even create their environments. The
falsifiable proposition – that intelligence and sentience will evolve.”
“ ‘In the beginning was the Word, and the Word was with God and the Word was God.’ The Word – information with a purpose – knowledge. Purpose – to create life and, in particular, a life which could have a relationship with its creator – intelligent life – and through doing so with His creation. Perhaps we must evolve beyond the constraints of our biological intelligence in order to actualize our innate potentiality to know God through understanding the intricacies of our relationship with His entire creation.”