<Book Index>

 

Chapter 4

Grounding The Tri-level Architecture in Neuropsychology and Open Logic

Revised February 2003, May 2004

 

 

 

Section 1: Introduction

Section 2: The Houk/Barto model of the cerebellum

Section 3: Model of Information Metabolism

Section 4: Advent of Anticipatory Technology

Section 5: Ecological processes

Section 6:  Application of stratified theory to information production

Section 7: Generalizing a model of human memory and anticipation

 

 

 

 

 

 

Section 1: Introduction

Mathematical models about the principles of theoretical biology were developed in the 1950s and 1960s (see Levine, 1991, for history).  These models involved differential equations, switching nets and category theory.  The formal work by Rashensky, Rosen and Grossberg helped to establish this foundation to modern connectionism.  The resulting field of study is quite a large one.

In stratified theory we have introduced a new set of foundational principles, some in direct opposition to the principles of connectionism and some that are built on a solid heritage shared with connectionism.  Many of the foundational principles for connectionism are controversial.  The notion that information can be distributed is perhaps the primary point at issue.  The strong form of the field called artificial intelligence, or simply “AI”, has long taken great pains to discredit the notion that information can be anything but explicit.  The historical context and motivations for the long lasting conflict between AI and connectionism has to be addressed carefully and will not be attempted here. 

Stratified theory takes one additional step, beyond connectionism, that is even more controversial.  In stratified theory information can be discrete or distributed, but can also be present in many organizational levels.  These levels of organization can even be nested so that two systems interacting with each other may have entirely different substructure.  The stratification is thus “relative” to location.   Moreover, the stratification itself sets up a theory of emergence, as we will see in this chapter.

I have been somewhat involved in primary work on the nature of neural and immune systems, both while a graduate student, in mathematics, and later in my professional life.  I was involved in and published detailed work on specific neural circuits.  These studies revealed certain general systems properties including: 1) associative learning, 2) lateral inhibition, 3) opponent processing, 4) intra anatomical neuromodulation, and 5) cross-level resonance (Levine, Parks & Prueitt, 1993; Prueitt, 1994). 

In 1995 my work was to be influenced by

1)     Meeting several of the leading Russian logicians in the field of control theory and semiotics

2)     A growing awareness that mathematics, as we known it today, has NOT been constructed to model the cross organizational scale phenomenon that clearly is involved in maintaining autopoiesis (Maturana and Velara, 1989) for a living system

3)     A growing awareness that information technology was hopelessly entangled with the strong form of artificial intelligence and scientific reductionism.

Each of these three influences has had a profound impact on the development of stratified theory, and the increasing significance of anticipatory technology.

While late twentieth century advances have proved significant, it is argued by some that the attractiveness of, the hope for, a single theory of brain function, perhaps based on the computer metaphor, was to mislead basic research in cognitive science.  The confusion created by the science fiction aspects of artificial intelligence was further distorting the process of investigation.  Many shared this viewpoint, but it was unclear what could be done. 

For example, in 1995, Lynn Nadel made a point to frame the current interest, in how learning and memory is organized, as a counter intellectual current to long held dominate scientific positions that, incorrectly, hold there is but one general long-term memory system (Nadel, 1995). In his analysis, analysis shared by others in (Schacter & Tulving, 1994) the fixation of memory theorists on the possibility of a single model can be traced to, the absence of, operational definitions of what is a "system".

Nadel's conclusions was that multiple memory systems exist; memory is involved in all or most of neural processes; and that experimental methods have identified the functional and architectural characteristics of distinct types of learning and memory.

General systems properties have also motivated the field of artificial neural networks and the related fields of computational control theory. The theory of embedding fields, Stephen Grossberg's original formulation of a methodology for modeling neuronal networks, envisioned the development of a body of computational algorithms that described the neuropsychology, the neurochemistry, as well as other experimental evidence regarding the mechanisms involved in human perceptual and emotive actions (Grossberg, 1972a, 1972b).

Theories of field dynamics and ecological psychology have been used to extend and overcome some of the limitations of early connectionist methods.  Central to these limitations are the various forms of a "vector representation hypothesis" that assume vector space mathematics to be sufficient for modeling biological activities.   Ecological psychology in particular offers some new language that helps to separate the concept of a system and the concept of a system’s environment.  Once one has this language, then one can more easily talk about multiple ontologies and the grounding of a specific ontology within a context, e.g, as having a “pragmatic axis”. 

However, no general theory of micro to macro transformations of information has yet developed that fully captures the essential notions of emergent computation and cross organizational scale entanglement.

In 2003, bits and pieces of a general theory did exist, though scattered in different disciplines. The notion of a national project began to be discussed.  This project would recognize that the limiting constraint on developing what is now called the Anticipatory Web was the educational background common to those who might develop such a new reality and use its capabilities.  We concluded that a general theory could not be developed without some sort of counter balance to popular theories that we felt are at least partially based in a specific type of confusion. 

As discussed in previous chapters, notions of cross scale entanglement exist in the historical literature of quantum mechanics. But the notion of cross scale entanglement existed almost nowhere else.  In linguistics, one could talk about double articulating and thus introduce a parallel between quantum mechanical entanglements and the issue of ambiguation/disambiguation in linguistics (Lyons, 1968).  In the various schools of ecological psychology, behavioral cycles are described as having a supporting substructure and having an open environment in which cyclic behavior is expressed. Furthermore, behavioral cycles are seen to have distinct modes and to be emergent from an interaction between environment, self and substructure. Action/perception cycles are seen in ecological physics and psychology. Partition/organize cycles are seen in models of molecular computation (Kugler et al, 1990).  But these fields of study where not integrated and no process for integration had been contemplated, before the planning began for a national project. 

Prueitt’s two conferences at Georgetown University (1991-1992) and Pribram’s series of conferences (1992-1996) at Radford University set the stage for planning the national project on knowledge sciences.  The precise notion of cross scale entanglement arises from the logic of quantum mechanics and has given rise to various "quantum metaphors" linking consciousness and awareness (Pribram, 1991; 1993; 1994; King & Pribram, 1995; Pribram & King, 1996). The cross scale entanglement issues, as well as empirical evidence on the formation of compartments and the stratification of matter and energy in time, lead us to restate many fundamental questions about the relationships between psychological objects of investigation and physiological objects of investigation.  The national project was conceived based on the difficulty one finds in talking about these questions, even within scholarly communities. 

As important a principle as cross scale entanglement is, we begin to realize that the biology of physiological phenomenon requires a full theory of stratification.  This realization lead us to define a tri-level architecture for the machine side of an Anticipatory Web (of information).  Those computational constructions that encode the categories of invariance over a set of measured data were to be encoded at one level of organization.  Anticipatory templates evolved under a different set of processes.  In real time, the two “levels” entangle in respond to specific input from the “present moment”.  This evocative input draws into an Orb construction some, and not all, of the elements from the measured data invariance.  The human side of the Anticipatory Web creates an interpretation of the informational structures.

Our work in 2002 and 2003 developed a standard format and encoding structure both for the set of all categories of data invariances and for a relational operator that defined a local relational property between pairs of these categories.  The result was the first Ontology referential bases (Orbs) developed by Nathan Einwechter and myself.   The 2003 notational paper for Orbs is provided in Appendix B.

We began to see that a new paradigm described technology architecture for text understanding.  From this new paradigm one should be able to create a general class of computation aids to human reasoning and information gathering.  By anticipating the economic power of new a class of technologies we hoped to provide some mechanism for the exploitation of this technology.  Here at last, we argued was a value proposition that would allow stratified theory to take center stage. 

A specific application of this work to defense intelligence was developed and proposed.  Several possible applications to bio-defense were addressed in new work that built on the foundations laid out before 2002.  However, we still could not find funding.   The development of the first full Orb-based human-center information production system could not be undertaken.  In 2003 and 2004 our efforts constantly were limited by what program managers, and advocates for things like the Semantic Web, could bring themselves to understand.

We knew that it was just a matter of time, and timing.  It was a matter of time before the natural principles of the Anticipatory Web were more commonly understood and appreciated.  For our group, it was also a question of timing.  Could we catch the wave and bring the BCNGroup Foundation into an active role?  The Charter of the Behavioral Computational Neuroscience Group envisioned the aggregation of new intellectual property in the knowledge sciences as a means to simplify the literature and fund a the development of a K-12 curriculum. 

The proposed architecture of text understanding had been grounded in neuropsychology, and in the formalisms of open systems theory (Prueitt, 1996). The architecture made a fundamental distinction between a computational and natural system (Prueitt, 1995; 1997). Understanding this distinction helped us to draw the line between what computational systems may be expected to do and what role natural systems must play in human computer interactions. This distinction leads to what we called perceptual measurement.  Our discussions about perceptual measurement lead us to develop the nine-step actionable intelligence process model in 2003, along with differential and formative ontology.  Web reference to this work was made available. 

The grounding of the Tri-level architecture in neuropsychology has been an important part both of our formulation of the principles of anticipatory technology.  With this grounding completed we began convincing policy leaders that the national project was absolutely essential if the United States was to take the next step towards a more full realization of participatory democracy.


Section 2: The Houk/Barto model of the cerebellum

During the 1980's, the neurophysiology and neurochemistry of the cerebellum and motor expression circuits gave rise to a model of the cerebellum as an array of adjustable pattern generators (Houk, 1987; 1989; Houk & Barto, 1991).  A great deal of detail about the physical processes involved was known already in the late 1980s and computational models enriched this science during the 1990s.  Other examples exist in research literatures that focus on computational models of the hippocampus, the prefrontal cortex, the visual cortex, etc. 

The Houk/Barto model of neuronal connectivity was used as one example, among others, to illustrate foundational principles in the context of modeling a physical single-level processes conjectured to be involved in composing, controlling and expressing motor programs.

Holonomic brain theory (Pribram 1971; 1991) describes a different type of model.  The difference is, simply stated, huge.  Pribram stressed the difference between localized processes and process that were simply not localized or even localizable. 

Gravity is an example of a holonomic constraint.  It is involved in local dynamics, and yet its nature is only slightly affected by the local boundary conditions of things that must move around in a field expression of this constraint.  The anticipatory technologies allows the abstract representation of a holonomic constraint to be separated from the things that have initial conditions, due to emergence, and boundary conditions related to location.  The physical correspondent to the abstraction of the middle layer is a set of physical objects that exists only in a present moment. 

How does the brain “see” these objects?  Due to the processing of light from the object, a distributed wave propagation of micro signals from the neuron’s receptive field move along dendrite pathways towards the cell soma.

Each micro-dendritic event, emanating from groups of synapses produces a localized Fourier transform (using Pribram’s language), converts wave potential into discrete event potential in the form of pulses.  These pulses use channel properties of dendrites to move the signal towards the neuronal cell body. This wave of dendritic signals is accumulated as potential energy on the soma side of the axon hillock.

In my interpretation of the theory, at the soma an inverse Fourier transform is generated from the conversion of discrete sequences of micro-pulses into a single semi-coherent electromagnetic field. The inverse Fourier transform is computed as micro electromagnetic events, initiated in dendritic synapses, are accumulated on the soma side of the axon hillock.

As the electromagnetic imbalance is strengthened, phase coherence separates out various parts of the spectrum, resulting in yet another localization and discretization into a Fourier type representation. The potential energy is drawn from the soma into a substrate composed of protein conformational dynamics. Here chemical valances move the potential energy's location, across the hillock where new field coherence is established.

Field coherence is modeled as a new inverse transform. It is followed by the emergence of a forward transform that provides the weights, or more generally the initial conditions, to elemental pulse generators that produce electromagnetic spikes propagating down the axon. Each forward transform selects a, perhaps quite different, basis. Each inverse transform is involved in establishing the structural mechanics of memory stores and in allowing these memory stores to influence the new forward transforms as they emerge in propagated waves of neural activation.

Of course the exact class of transforms, best suited to model neural phenomenon, is the subject of a deeper discussion (MacLennon, 1994). The forward and inverse Fourier transform is a first approximation.

The Fourier transform takes an electromagnetic signal and re-expresses this signal as a sum of amplitude and phase modulated waves expressed mathematically as regular sine and cosine functions (of time). The sine/cosine basis is used under the assumption that the phenomenon has strictly periodic generators. The inverse transform takes these summed representations and loses the representation that relies on a specific set of basis functions. Moreover, the notion of a forward and inverse transform is an artifact of mathematical formalism, since an electromagnetic wave can be seen in a regularized form, or not, without effecting its ontic status.

The transform can; however, be seen as movement between two temporally defined scales of dynamical activity. These scales are marked by the development of regular patterns of emergence into one scale from the other.  A neurowave propagation is then composed of a series of many (thousands) of transform/inverse-transform pairs.  (Look ahead to Figure 6)  Each forward transform selects a representational basis, and this basis is built depending on what is available locally to create a physical basis for the basis.  Each inverse transform encodes information into the new “location” and allows the local ontology to re-express the information. 

For sometime, a structural order has been observed in the organization of the cerebellum Brodal, 1981; Ito, 1984; Gibson et al, 1985). Furthermore, experimental data, reviewed in (Houk et al, 1990), suggests that movement signals recorded in the magnocellular red nucleus, located outside the cerebellum, are produced by a pattern generator located in the cerebellum. These signals are recorded by microelectrode.  The stratified theory suggests that memory encoding and memory uptake is made via spreading activation waves where only part of what is occurring is observed as the activation wave.  There are some deep issues here related to how informational invariance can be preserved even if at each of a thousand iterative cycles the basis for representing this information is changed. 

A note is in order on the future implementation of principles being discussed here in the form of a computer-based Human-centric Information Production (HIP) system. The voting procedure, in Appendix A, is a simple version of quasi-axiomatic theory (see Chapter 6).  Either the voting procedure or a more full implementation of quasi-axiomatic theory works within the tri-level architecture.  Before 2002 these implementations were not as yet a system of computer programs.  After 2004, our task was to build the conceptual foundation for stratification theory from the neuroscience and physics. The modeling being expressed here (Figure 1) was an important first step.

Figure 1: Modified from Houk et al 1989.

The passage “across” the axion hillock is non-Newtonian in that several organizational levels of stratified processes are involved.  For this reason, and for others related to the description of actual perceptual measurement, our interpretation of the neural processes at the axion hillock creates an illustrative model of the cross scale phenomenon, more broadly considered.

The experimental data suggested that electromagnetic potential energy is broken down into a set of basic elemental processes that activate substrate "computation" based on structure activity relationships (Prueitt, 1995, Chapter 1; Zebezhailo et al, 1995). Following this molecular computation, the state of the substrate developed a new representation within a derived set of basic elemental processes.  This new representation is in the language of the metabolic environment present at the time, and location, of the specific computation.  The change in basis is a natural consequence of the movement form one environment to another environment.

The loci of activity may have moved due to transport properties associated with cell structure, again as a consequence of molecular computation. In this new location, an emergent electromagnetic field is formed by assembling a subset of the basic elemental processes according to situational influences at that locus.

For the Houk/Barto model of the cerebellum, this model works as follows: A recurrent (reentrant) feedback pathway composes a motor program from basic elements consisting of patterns in neuronal bursting. Composition is controlled by activated combinatorial maps in the motor cortex. Purkinje cells { Pi } innervate their dendrite receptive field into a regular array of parallel fibers from cerebellum sub-cortical cells. Synaptic connections { wi,j } encode associative strengths. Basket cells { Pi } under the influence of the motor cortex, in the cerebral cortex, inhibit a selection of Purkinje cells and allow the uninhibited Purkinje cells to sample the array of parallel fibers for components of motor expressions that are consistent with combinatorial maps that are active in the cerebral cortex. These Purkinje cells then read out a signal, expressed via its axons, into the receptive field of a group of subcortical cerebellar cells { N } that form a deep cerebellar nucleus. This cerebellar nucleus has reciprocal connections and common dendritic innervations with populations of cells, arranged in columns, located in the cerebral motor cortex (not shown in Figure 1). The cerebellar nucleus has excitatory connections to the red nucleus { R } which are located in the mid brain outside the cerebellum. Environmental impulses {IMC} are received by the red nucleus and these shape the motor commands which are then projected to an interneuron and to muscle fiber.

The shaping of the program's expression by a sensory stimulus is conjectured in Gibson et al, (1985) to involve a two stage process whereby the movement signals, recorded at the red nucleus by micro-electrode, are the expression of response patterns composed to meet the demands of motor commands. Continuous feedback from the environment then modifies the expression. A signal is also sent from the red nucleus to the inferior olive (again outside the cerebellum). The inferior olive receives reinforcement stimulation from various sources and projects reinforcement signals to modify synaptic weights { wi,j }.


Section 3: Model of Information Metabolism

In the previous section, a network model of one "subsystem" of the brain is described. The purpose of this description is to advocate a specific context for using the stratified paradigm to address a central open question regarding the ‘discontinuity" of the stimulus pattern from one location to another. This discontinuity was first noticed by Walter Freeman in his study of the olfactory system of rabbits (Freeman, 1995a) and then later in other neuronal subsystems (Freeman, 1995b). It was noticed that the stimulus patterns generated by olfactory bulbs were lost during the transfer of the signal from receptor cells to interneurons in route to the olfactory cortex.

Figure 2: Special episodic sequence where an organized whole briefly dissipates into a substrate and then reorganizes

The model for cross scale phenomenon allows us to make specific speculations.  The specific discontinuity of a signal pattern might be a generic property of a transfer function that samples a specific part of the electromagnetic spectrum from dentro-dentritic field interaction and encodes information from this spectrum into structural activity relationship (SAR) information at the level of protein conformational states (see Figure 2). This encoding might be seen as a Fourier decomposition of the potential energy into a recipe that is then propagated via a pulse wavefront along dendritic channels to the axon hillock. It is conceivable that at the axon hillock the pulses, and the encoded information, is placed into some type of micro-environment where a process of reorganization may occur.

The elements of the sequence are initiated and terminated by cross scale events.

Using this model of cross scale phenomenon, a viewpoint is supported that internal and persistent cycles of emergence and dissipation is a fundamental property of open complex systems. The viewpoint is then applied to various investigations, including the development of computational architecture directed at managing and manipulating machine representations (Orbs) for the purpose to assisting human production of new real time information. 

Natural language understanding by humans depends on the emergence of mental events from which we derive, somehow, an understanding about our environment and ourselves.  The relevant science literatures suggest that human understanding is based on the assembly of components from implicit memories, which we represent as statistical-type artifacts, in a context that cannot be fully captured in statistical notation.  The development of the categorical abstractions that are encoded as Orb constructions allows a simpler and more powerful means to represent the atoms of memory and the potential relationships that might be established in a real context. 

The Orb constructions are merely sets of ordered triples in the form < a, r, b > where a and b are locations within a graph and r is the relational label of the connections between a and b (see Appendix B).  The stratification theory has a very simple computer data encoding format, and thus what the theory has in complexity the software should have in simplicity. 

Unlike Resource Description Framework (RDF) type ontology encoding, the Orb constructions do not make a final imposition of semantics outside of a specific situation. A theory about situation type may be encoded in the Orb constructions, as is true with RDF constructions.  However, the final imposition of ontological claims is withheld until a reifying process that usually involves a human in the loop and which can take into account the possible novelty within the specific situation.  Reification implies both iterated cycles and human introspection. 

Natural scientists have noted that the experience of context is found in action perception cycles with an environment.  Thus an increase in the use of action-perception cycles would seem to overcome some of the limitations imposed through the use of statistical knowledge. A dependency on “known” substructural atoms by the ecological affordance (Gibson, 1956; Shaw, 1999) forces regularity in substructural type and thus statistical methods do work in creating measurement processes.  These statistical artifacts can then impose the final constraint on any emerging pattern being expressed by the Orb elements.  Direct perceptions of both substructural type and ecological affordance bind the formation of structure into an expression of function. 

Of course, this means that the autonomous understanding of targets of investigation, by computational systems, should likewise depend critically on the cyclic extraction of features from "experience".  Within each cycle, new categorical invariance can be identified and local measures of co-occurrence indicated by adjusting the set of syntactic units in the set of Orb constructions, e.g. in the set of the form < a, r, b >.  We argue that logio-linguistic principles can be used to encode structural characteristics of salient features within human language, and that small subsets of these features will evoke human mental event formation in much the same way as in normal human communication.   In operationalizing a principled theory regarding the use of Orbs as cognitive priming, we developed a simple form of category theory and the Orb encoding.  The resulting work produced machine-readable ontology and first order logics that bind knowledge representation together.  This then is the anticipatory technology. 

Human perception is the key to interpretation of anticipatory ontology.  Without human in the loop activity, the machine ontology will not reflect that part of the “now” that makes that “now” unique.  The RDF standards depend on a precise encoding of a theory of the world that is then recalled to stand in for human knowledge.  RDF is formally closed at run time, whereas Orbs are formally open at run time. 

Our logio-linguistic principles reflect general systems properties because the underlying morphology of concept formation has been observed by natural science to conform to a class of general systems properties, including emergence.  These properties can be seen in physical systems, particularly if one examines certain viewpoints about quantum field theory.  As the final aspects of first generation anticipatory technology was put into place (2002 – 2004) we understood that Anticipatory Web concepts would eventually replace Semantic Web concepts.  The Anticipatory Web concepts are ultimately easier to understand and have a greater fidelity to how the brain system works. The anticipatory technology is easier to encode and manipulate in the computer.  We reasoned that the anticipatory technology would be discovered to be cheaper, faster and better than semantic technology. 


Section 4: The Advent of Anticipatory Technology

The transformation of information technology by Semantic Web efforts failed to achieve what was expected.  This failure was due to a mismatch between complex natural reality and formal expressions of logic acting on static encoding of taxonomical information.  The marginal successes occurred in cases where standard RDF ontologies were about the data formats and data exchange issues (John Sowa, personal conversation). Some degree of interoperability between computer programs was achieved at a greater expense than was necessary.  The confusion over the relationship between semantics and pragmatics made real progress on differential and formative ontology very difficult. 

It was clear to BCNGroup Founders that real-time psychological and social event structure would be understood only if one sees events as having emergent structure.  But it was also clear to us that a new mathematics and a new computer science curriculum will soon be developed, reflecting the principles of Human-centric Information Production (HIP) and stratified theory.   These clear perceptions lead us in 2003 to start planning for a National Project to establish the knowledge sciences as an academic discipline.

The limitation of formal theory in the context of modeling complex systems was the critical scientific result that leads to anticipatory technology. 

Linguistic principles are not described well by classical physics, nor classical logic or set theory.  The use of mathematics can lead to a specific type of confusion.  Hilbert mathematics, for example, is often used to capture so-called “latent” relationship due to linguistic variation seen in the distribution of words in text.  But these methods have only one notion of nearness, the notion derived from the induction of the counting numbers (for more on induction see; Goldfarb, 1992).  Having see this issue sense the early 1990s, the BCNGroup Founders began to make the claim that new logical principles, and notions of membership, are required to design and implement a new generation of machine based autonomous intelligence. 

By 2004 we began to demonstrate operational information production systems of the type we envisioned in the mid 1990s.  Three key concepts centered around (1) organizational stratification, (2) the formation of categories of occurrences within each level of organization and (3) cross scale entanglements involved in emergence of events.  Events at one level were due to the aggregation of categorical defined atoms at a faster level of organization.  A top down constraint on the emergence process is then instrumental in placing the event within its organizational level.   Computationally, the Orb sets are sampled by anticipatory templates and a situational ontology is expressed as a phenotypic graph that when observed primes the human cognitive process and results in a mutual-induction involving the deductive capabilities of the computer and the inductive capabilities of the human.

We began to use the notion of induction is a specific way, and contrasted “induction” with “deduction”. 

 

We suggested that deduction is not as natural to humans as our academic traditions indicate.  Historically, rationality and formal logics has been represented as the ideal for human thought, not only in some theoretical sense but also in the sense that a perfect human being has been portrayed as being perfectly rational.  Rationality is defined in terms of consistency and completeness, and is best reflected in David Hilbert’s grand vision about the completion of mathematics.  Hilbert’s grand vision still governs a great deal of the work in science and mathematics.  However, Godel’s work on the foundations of mathematics and related literatures demonstrate the limitations.

 

Two electrical motors will induct changes in state.  The coupling between physical electrical systems is due to “holonomic” constraints related to an electromagnetic field. Holonomic effects are not locally concentrated as described in Newtonian action-reaction systems.  The induction occurs non-locally, and thus the mechanism “causing” the state changes is not accounted within a discrete model.  However, the state changes do occur and can be modeled using continuum field models.  The mathematics is elegant but is often beyond the reader’s experience.  Chapter One has a longer discussion on the differences between holonomic and non-holonomic causes. 

 

In the same way as two physical electrical motors induce state modifications in each other via a non-local interaction; we observe that symbol systems such as natural language or gestures will cause modification in the mental state of a human being.  Because of the confusion about the nature of a computer program, we use the term “mutual-induction” to talk about and action-perception cycle that involves in each cycle the two mutually exclusive systems:

 

1)      the computer program with some type of display or informational interface

2)     the acting and perceiving human living in real time and experiencing, among other things, the information being computed by the computer.

 

The human can be involved in creating computer input as a result of the responses that the human has to old information states.  These human inputs can result in new informational states in the computer. 

 

Likewise, the computer can be computing informational states in the computer.  Upon viewing an informational state, the human’s mental state is altered, or manipulated via cognitive priming, so that human cognitive acuity and tacit knowledge is primed by the computer informational state. 

 

In the anticipatory HIP architecture, the human action-perception cycle is influenced by the series of states generated by the computer. 

 

 

Figure 3: The anticipatory loop

 

Why is this radically different from the current uses of a computer?   Of course, humans are always involved in these action-perception cycles.  So the primary difference is in the architecture of the computer’s data encoding and repository for production rules.  The tri-level architecture matches the biological facts, as understood by many leading cognitive scientists, and separates the invariance of structure from the anticipation of function in environmental context in real time. 

From our study of natural science, we know that neural systems derive implicit structure and encode structural representation of this structure into physical substrates of the brain (Schacter & Tulving, 1995). This means that the consequences of experience includes modifications of metabolic processes, and these in turn may have direct consequences at the faster time scales where metabolic and quantum events are organized.  We observe that the properties of emergence propagate from the faster time scales into slower time scales and are then participatory to the mental events themselves. 

Through processes of adaptation, in the many levels of self organization within the complex, the substrate expresses coherent and emergent neural (and electro-chemical) phenomena that have correspondence between the structural invariants of the world experienced and internal mental event that represent this experience to the higher order processing centers of the brain (Pribram, 1971; 1991; Levine et al, 1993; Prueitt, 1997).

These phenomena of emergence go to the issue of how knowledge can be extracted from a data source.  We follow a logic-linguistic model that assumes both that knowledge acquisition and the object under investigation is complex.  Knowledge and the natural world is complex.

In written text the extraction of knowledge representation depends on finding correlations over time and developing a flexible mapping between internal representations and external objects and agents (Michalski, 1994). The computational system must have an "understanding" of how objects are composed from a substrate and how they function over time. This understanding can be, at least partially, "engineered" into software and hardware.  But in doing so, the engineering should have some of the same properties of openness as do natural systems, and the overall system should accommodate control by the environment during the periods of emergence. 


Section 5: Ecological processes

The ecological psychology community has developed a view that action-perception cycles are emergent from the physical properties of subsystems and are driven by a class of periodic forcing functions. The scholarly research in ecological psychology and ecological physics views periodic forcing is an essential part of the temporal stratification of biological organization into levels.

In this view, any one level of organization contains objects that are emergent constructions from a relatively stable substrate of basic elements.  The periodicity of formation and collapse is simple local attempts to balance conservation laws.

Figure 4: Natural systems stratify

It is clear from the literature that critical issues must be resolved in any situational analysis of natural systems. 

1)     organizational stratification,

2)     the formation of categories of occurrences within each level of organization and 

3)     cross scale entanglements involved in emergence of events

Categories of occurrences and the relationships they develop define the levels as shown in Figure 4.  Each level, Li , has the capability of forming a substrate from which new organizational wholes may emerge. Each of these levels, and their interaction, can be partially simulated by a thermodynamic model. The stratification can then be nested and thus the Figure can be misleading. 

Natural systems are "open" systems with varying kinds of openness. For example, enzymes produce microenvironments where constraints are placed on the "meaning" of assembled biochemical agents.  In the hippocampus micro environments fuse the associational traces of implicit memory into compartmental invariance that are "judged" by the processes of other limbic systems and of cortical regions. These particular microenvironments are agile semi-closed flexible and capable of responding to novelty and nonstationarity in the environment.   

Stratified theory will allow science to organize experimental results in various areas of investigation.  These areas include work on modeling social systems, biological systems, and environmental systems. 

The stratified thermodynamic model of open systems (Figure 5) has micro environments that emerge through a specific coherent aggregation of compatible processes. The average lifetime of processes at a certain level is an order of many magnitudes different from the average lifetime of the processes at the next level. For example, physical atoms form one level and chemicals form a different level.

Figure 5: Stratified thermodynamic model where one of the levels has collapsed.

The collapse of one of the microenvironments leads to the diffusion of material into an environment that supports the reassemble of this material, and other similar material, into new micro environments. As these new environments arise, the structure and function reflect the characteristics of the material as well as the state of the environment.  The persistent of function though many iteration of collapse and aggregation is maintained by autopoiesis (Maturana and Verala 1989).

Even up to a now (2004) it is not clear how to engineer open systems, but a great deal of the work we have completed has not been applied to this problem.  We expect to make significant progress because the foundational principles of anticipatory technology have addressed issues that are confused in the core concepts of semantic technology. 

Open systems are extending engineering and close loop control theory into a new class of logics similar to quantum logic.  These new logics reflect several classes of open systems properties. The permutation of systems into each other produces the features of assembly/disassembly processes and thus creates the feedback loops that reify the set of Orb constructions and the set of anticipatory templates.

The interface between a human user and a computational system will provide a "pragmatic" axis to situational analysis.  Situational logics have formal means to describe how constituents are assembled into operational wholes and how operational wholes are disassembled into components.  This means that formal tools for describing assembly, degeneracy and indeterminacy can be part of the computational logic as well as the machine - human interface. The human cognitive ability, to step away from the formalism and take a leap of faith, is provided a stream of high quality factual information about event atoms and reminders of situational conditions. 

Human experience of knowledge involves an internal awareness of the structure of time, and we humans do this well – but computational systems do not.  The curriculum of the knowledge sciences must make it clear that pragmatism is rooted in the real-time experience.  This point returns us to the work by Robert Rosen in the nature of formal systems.  Rosen’s work is indeed difficult and must be studied for some time in order to see how his work on theoretical biology and category theory leads into an understanding of how the experience of knowledge might be formalized.  So the K-12 curricular elements must develop a conceptual framework in which this type of difficult work can be addressed by at least some portion of the population. 

A central issue facing general systems theory is about whether or not a class of transformations can be characterized using methods more powerful than numerical models. Can our present notion of computation be extended in certain useful ways?  In fact, Zabezhailo et al (1995) demonstrated to the author that a useful predictive theory of biochemistry is realizable.  Zabezhailo used special logics that perform "iconic computations". These computations are carried out using the special inference operators of Quasi Axiomatic Theory (QAT) (Finn, 1991, 1995) in the context of Soviet work on biological and chemical warfare based on bio-chemistry.

A complete predictive theory of biochemistry is in the future.  For example, iconic structure might be inducted using a process that is quite different from that which constructs first the integers and then the Hilbert mathematics (see Goldfarb, 1992, work on inductive informatics). 

Paths in "iconic" space could, of course, be analogous to trajectories in numerical state spaces; however, these iconic trajectories should be lawfully constrained by the information in a database containing the results of specific analysis of biochemical structural activity relationships. Pospelov (1986; 1995) referred to these trajectories as syntagmatic chains.

Other scholars and logicians have worked on related issues.  For example, the relationship between canalization in simplicial complexes (Johnson, 1996) and structural adacity (Burch, 1989) in artificial life systems provide a means for the simulation of some of the mechanisms that are enfolding the individual state transitions of computational agents in a complex artificial ecosystem.   But it should be noted that much of the work in artificial life treats the state transitions as if there is an experience of state, and our position is that there is no evidence that an experience is occurring. 

The tri-level architecture is interesting to us because the substructural level is considered to be the atoms that when aggregated together under the influence of templates produce the emergent phenomenon in the form of compounds.  Thus we were increasingly hopeful that a new generation of open systems engineering would be achieved with the tri-level architecture.

Section 6:  Application of stratified theory to information production

One approach to knowledge representation and management assumes the existence of a table (database) where the system states, i.e. those states that a compartment can assume, are all specified and related to a database of subfeatures. We construct a system that is specified in a formal fashion and which is computable. In this case, a theory of the world is provided in the form of computer program.

Within the compartmental boundary of this program, an underlying ontology can assume different system states and thus one might capture how the meaning of terms may drift. The rules that govern ontology allow a modification of the sense of the terms. This gives us hope that issues of interlingua and pragmatic type can be addressed within individual compartments.  However, the procedures for the formation and dissolution of these programs must allow for the limitations that formal systems have in general and that a specific program has in particular.  The system should allow for emergence and control of emergence from an environment.  How is this accomplished?

Machine intelligence can have a layered architecture where data constructions in one layer are combined to produce flexible concepts using partially defined relationships. An inductive analysis of the situation is then possible, where a fusion of data occurs based on the accommodation of characterizations of the novelty of the situation with characterizations of the background knowledge.

Complex natural systems are open systems that (1) are embedded in a larger space (2) are composed through an assembly process, (3) have behavior properties with internal and external work cycles, (4) can be described as variably stratified (Figure 7).

Figure 7: The Target of Investigation is observed a number of times

Because of these four properties, the Target of Investigation can be observed a number of times

T = { O1 , O2 , . . . , On }

·         each "observation", O, of the target has a "bag" of properties P = {p1 , p2 , . . . , ps}. The cardinality of the bag of properties is finite (but open).

·         each "observation", O, is composed from a "bag" of basic elements A = {a1 , a2 , . . . , as}. The cardinality of the bag of elements is finite (but open).

·         logical hypothesis (J.S. Mill) can be asked

1.     does an observation have (not have) property p

2.     for each basic element, ai , is ai cause of (or not a cause of) the property p

The observations produce the categories of atoms and event templates.  Of course, the descriptions of properties and the descriptions of basic elements of the target of investigation must be possible. 

The quality of any automated reasoning system is a function of its power to reveal the basic signature of a situation under investigation (see Ritz and Huber, 1996).  Producing signatures requires instrumentation, measurement, representation and encoding as well as a number of other steps.  However, a number of different methods may produce reasonably good results, and given reasonable good initial results it is possible to use additional methods to develop a reification process and consequently a high quality ontology that represents the structural characteristics of the basic signature.

The block diagram in Figure 8 is open to architectural instantiations of computational systems that duplicate the structural mechanisms that experimental neuropsychology suggests are used by biological systems.  Reification processes can be developed by following this diagram

Figure 6: Block diagram for Situational Analysis

First, some sort of instrumentation is required to have a data source.  Then a data source is attended to by a system using perceptional measurements about properties and features.  These properties and features are represented, approximated, as elements of a set of primary components.

The separation of properties and features is vital in establishing correlations between the presence or absence of features and the properties of the event.  The separation is thus taken to be a stratification where features are one level of organization and properties of the event are at a middle layer between the feature level and a level that is defined by environmental constraints.   One can misunderstand the notion of stratification in this context.  The features are properly “atoms” whose identity has been altered by an aggregation of a set of atoms into an object.  That object haves properties, partially determined by the atoms that compose the object. 

As a practical matter the full understanding of stratification, as exemplified in quantum mechanics or in linguistic double articulation (Lyons, 1968), is not necessary in order to produce value from the block diagram.  In fully stratified theory, the atoms are combined into a compound which has properties that are partially established by environmental constraints AND the process in which the compound comes into a separate existence as a compound.  

Stratification is essentially already present in many machine-understanding techniques.  It is seen in the separation of token occurrence and text unit in latent semantic indexing (Dumus & Landauer, 1992).  Latent semantic analysis measures the distributed co-occurrence of elements within units.  Other types of categorization and feature extraction type techniques exist in great numbers.  For example, a simple theme vector transformation of a document's content produces a representation (Rijsbergen, 1979; Prueitt, 1996b) of the linguistic variation in text.  The variation can also be captured using statistical models such as the hidden Markov model.  In each of these techniques, some element of stratified theory can be seen. 

Looking directly at the variation itself can be shown using the category theory developed by Prueitt in 2001 and 2002.  This work is called categoricalAbstraction (cA) and eventChemistry (eC).  Because of the importance of the work completed previous to 2001, this work will only be treated lightly in this book.  We anticipate a follow up book on event chemistry.  The point is that linguistic variation itself is a phenomenon to be observed by any method that reveals its character.  Once linguistic structure is observed and reified, then an event knowledge base can be constructed and one can apply the tri-level architecture directly to predictive analysis of event structure pointed at by the linguistic variation.  One need not use cA and eC methods.

Theme vector representation may also have some correspondence to how the brain processes information.  The Gabor representation and transform may play a role in the human cortex (see Prueitt, 1997 for literature review).  The type space is then projected into a system for visualization.  In the case of human perception, this projection is a reentrant bi-projection between the lateral geniculate nucleus and the layered visual cortex.  But we have to be careful here, because Hilbert space type mathematics does not handle the phenomenon of measurement and emergence as well as one might like. 

One indication that Hilbert formalism is limited is seen by observing how difficult machine intelligence has been.  Very few software systems have visualization of theme (feature) space as well as an integrated representation of knowledge in the form of a concept database. Concept databases exist in a few cases (Abecker, et al, 1997), but are primarily restricted to mechanical systems or industrial plants. Software packages such as Spires, developed by Pacific Northwest National Laboratories, makes projections into feature spaces based on cluster analysis.  But a process of reification and ontology building has not, as yet, lead to demonstrations of event knowledge bases having the agility and novelty detection that we predict from the Orb systems. 

Our scientific challenge is to see both computational visualization and human perception within the same paradigm. This may not be so difficult. A large amount of experimental research exists and some of this has been integrated into our explanatory framework. With stratified theory we see new computational projection schemes based on situational semantics. The object of visualization might be related not only to a themespace but also a concept space.  And the visualization can be controlled within the tri-level architecture so that as the visualization occurs it is possible to allow the environment to make small perturbations in the event representation. 

The study of the framework suggests that the human brain achieves the formation of concepts through a distributed disassembly and reassembly of representational features (Figure. 7). This framework is illustrative of a general systems property regarding the emergence of operational wholes within ecosystems.  It is also suggested that the human in vitro concept space is a virtual space in the sense that the space does not actually ever exist in total in any specific circumstance. Parts of this virtual space come into being while other parts are blocked by various types of competitive cooperative network dynamics. Of course, this simple architecture disguises the complexity of how the brain uses both its neural architecture and its chemical composition.

Figure 7: The niches in ecosystem share in the common use of a finite class of natural type.

Machine intelligence based on stratification might be simpler than human intelligence and yet share behavioral features.  For example, projection from a complete enumeration of a knowledge engineering type concept space can be made onto a concept subspace.  In the block diagram this is called an interpretation space. This subspace may be mirrored by activation of components of a situational model supporting automated reasoning. The mirrors can be maintained by a neural network associative memory as demonstrated in any number of neural network architectures. 

The mirrors are two-way, and as a consequence a separate mapping back to a concept subspace is made after an inference engine has changed the state space of the situational model. This notion of projections and mirrors between representational spaces is a reasonable first model for the production of computer based situational models. Before 2002, there are a number of scientific issues that are not yet resolved.  The structural form of a computer based concept space had not been worked out. Simple themespaces had been defined, but a new class of objects was needed in themespaces.  We expect these objects to create topological distortions of the otherwise flat Euclidean space, and instantiate a theory of structural operators defined in coincidence with inference engines.


Section 7: Generalizing a model of human memory and anticipation

The voting procedure reflects a correspondence between human perceptional process (involving memory, experience and anticipation) and a proposed computational architecture for routing informational bits, knowledge artifacts, from locations and to locations.  The voting procedure (Appendix) is the simplest form of the Russian extension of Mill’s and Peircean logic.  It has a data structure and a process model that operationalize the correspondence between the stratified processes involved in human memory, experience of mental awareness, and human anticipation.  

From this correspondence, it is possible to build a model of human memory. This model is a computational model consistent with von Neumann computing, but is organized into three separate levels. In the model, we stipulate a set of mechanisms that create a decomposition of experience into minimal invariance thought to be stored in different regions of the brain (Schactner & Tulving, 1993). The evidence for these models, of natural systems, is grounded in experimental science, whereas the computational model is grounded in computer science.  The use of the model requires a human in the loop to reify the event structure. 

The physical process that brings the experience of the past to the present moment involves three stages.

1) First, measured states of the world are parceled into substructural categories.

2) An accommodation process organizes substructural categories as a by-product of learning.

3) Finally, the substructural elements are evoked by the properties of real time stimulus to produce an emergent composition in which the memory is mixed with anticipation.

Each of these three processes involves the emergence of attractor points in physically distinct organizational strata.  How does emergence come to exist, and what material substance is combined together during the emergence process?  How does the ecological affordance of the environment come to constrain this aggregation process?  These questions have been treated in general within various scientific disciplines however, we treat the issue of emergence in a semiotic fashion, with an eye on how sign systems assist in human communication and comprehension. 

Figure 7: The process flow that we take as an accurate model of human memory formation, storage and use.

The Voting Procedure considers questions about the "entailment" (entailment is a generalization of all classes of causes, formal (logical), material, efficient and final) of natural symbol systems. A symbol grounding provides a detailed reference to distributions of potential meaning. Obtaining the specification of these distributions is a matter of observation using an information aggregation process.  Set theory is then used to support inference.

A class of minimal invariants (called minimal intersections) is constructed from the representational elements produced through measurement processes.  These invariants are then used as the logical atoms from which axiomatic situational logics are produced.

The investigation of entailment must begin with the identification of most of the basic elements that are involved in causing the situations under investigation. The process is called "descriptive enumeration". Complete and consistent descriptive enumeration requires the extraction of minimal elements to ground a system of signs pointing to the causes of a situation.  For example, if we are considering chemical reactions and properties of chemicals in complex situations, then these minimal elements are the signs of either the atoms themselves or larger groups of compounds that occur as an invariant across multiple situations.

In a similar fashion, a situational logic can be constructed that relates the presence or absence of these minimal elements to the properties of situations (Mill, 1843). The situational logic is completed when plausible and reliable truth evaluations are defined on a set of well formed formula having logic atoms corresponding to those non-empty minimal intersections of causal elements found through a discovery process.

So there are three steps,

1.     the identification of a set of minimal elements that could serve as causation of properties, and

2.     the development of a situational logic that predicts the properties of situations given a partial or complete list of those minimal elements that are present in the situation.

3.     the maintenance of a second order system for changing the intermediate language to accommodate new information.

The first of these steps are addressed using a version of tri-level logical argumentation, relating structure to function. This argumentation can be adapted to developing the computational memory required for automated text understanding and situational analysis. The third step requires a special interface to human experts.