Linguistic biocomputation model

in the wabe... 

Universal Governor (UG)

In the first section, we found a way to build a universal governor (UG), a cybernetic circuit capable of combining strategic (feedforward, program-based) and tactical (feedback, goal-based) governance paradigms. We saw how the brain seems to have made this discovery all by itself, as shown by the evolution of hyper-dimensional polyvariate pyramidal circuits in cerebrum (for spatial modelling) and cerebellum (for temporal modelling). 

The second level (TDE2) operates intrasubjectively ('intra-' means within), ie from the subject's viewpoint, also known as the PCT (perceptual control theory) paradigm*. In the traditional engineering approach, an 'eidetic' model (ie accurate, complete, not a projection or a reductionist view) would most likely be used, constructed with the implicit belief that with accuracy comes the ability to predict events exhaustively, and to not miss anything important. Due to update (sensor polling) latency, eidetic (accurate, 'objective') models 'fall over' in anything larger than 'toy' implementations. 

The rather commonsense solution which nature clearly seems to choose, is to derive control signals not from exhaustive knowledge, but from key perceptual differences. These differences appear as goal-target 'deltas' (drive state differentials**) in viewfinder-level*** input devices. When you aim a gun or a missile, the distance to be minimized is the gap between the crosshairs (located centrally within the 'scope') and the target.

Since conventional computers are procedurally-coded (ie feedforward) machines, the TDE universal governance paradigm (UGP= feedforward/command + feedback/control) includes them as a subset. 

Therefore, UGP has the capacity to fully explain biological intelligence in the broadest possible terms, as a computational component interacting with a cybernetic one. The reader should be aware that this dichotomy, important though it is, is for some odd reason, not explicitly taught at most universities. Each practicing engineer must usually rediscover it for themselves. Most on-line (ie real-time) applications, such as missile guidance systems, must use cybernetic methods (feedback, declarative models/coding). Off-line models can use more conventional mathematical approaches (feedforward, procedural models/coding) because they are not as time and data critical. Feedforward methods usually have lower error levels, while feedback models incur significantly larger 'control' errors. These so-called 'errors' are themselves often used as higher order signal metrics.

*William T. Powers invented PCT as the basis of a novel style of psychotherapy. Uexkull discovered it first, presenting it as a solution to the mind-body (identity) problem. 

**The current author (M. C. Dyer from Flinders University in South Australia) coined the (recursive) term 'Situation Image' to describe embodied (posture-like) synchronous polyvariate (recursive) states. These are identical to Uexkull's 'umwelt'. Dyer's discovery is independent of (but entirely equivalent to) both Uexkull's and Power's solution. The difference between a goal-state SI and a target-state SI is a Drive-State Differential or DSD. DSD's are the way to manage representations within cybernetic frameworks. Note the subtle difference in meaning between Drive-State and Goal-State adjectivals. The former has a procedural connotation while the latter is unambiguously declarative, and teleological.

***The easiest way to find sense in all this jargon is to Imagine that you must write a program for a SAM (Surface-Air-Missile). Clearly, you would start with the image that the target aircraft forms in the TV camera's CCD or infra-red retinal array. You would need to write code that sends corrective direction commands to the missile's guidance fins -no different in theory than a thermostat, except there would be two homeostat loops, one for the scalar u, and one for the scalar v, where (u,v)*<x,y> = [ux+vy)] describes the goal-target coordinate mapping, in which x and y are its unit basis vectors.

The main purpose of figure 6 is to indicate just how very different is the TDE computation paradigm from the two options available in conventional computing.  Both Princeton and Harvard models are random access stored program (RASP) machines*. The Harvard architecture, with separate dedicated memory spaces for instructions and data, is used in small computer-on-a-chip designs. However it is the Princeton architecture (Von Neumann machine) which has instructions AS data - there is literally no difference. At the time, this much needed innovation gave the fledgling computer the maximum possible organisational and operational flexibility, allowing its software to be self-referential, as well as allowing development of new prototypes at an unprecedented rate.  

TDE memory model = WOM / ROM

The TDE, however, is a WOM / ROM machine. This acronym means Write-Only-Memory** / Read-Only-Memory. The acronym 'ROM' is usually associated with a computer's 'firmware' or basic input-output services (BIOS). The core concept is not unlike that of DNA, meaning an executable image that is read into working memory spaces without any modifications being written, ie without changing the original pattern just copying it any number of times in different executable contexts. With ROM, since data doesn't move, there is no need to search - like the registry in a typical operating system, the resource name and its location are tied together in a look-up-table (LUT). This technique, generically called memoization, has one notable downside - though much faster, it is also much less efficient in its use of space, hence the excessive folding of the higher primate's enlarged cerebral cortex.  

The WOM / ROM pair of acronyms paraphrases the logic behind the Producers/ Consumers problem in linguistics. The reason that no reading occurs with WOM is that none needs to be done- the writer already knows what is written. Similarly, the reason that no writing needs to occur with read-only-memory ROM is that none needs to be done- the writing has already been done, and so the location and the memory contents are permanently tied together (semi-synonymous), as with any other associative memory.   The TDE system stores knowledge in hierarchical data structures in exactly the same way that computers store files in hierarchical file systems (mounted directories). Figure 6(c) is an attempt to depict the interaction of hierarchical data structures within the two channels, and at the three TDE levels.

 *Micah Rubin and Riley Boucot 

**Once upon a time, the WOM acronym was an 'in'  joke for the geeking classes, but as they say, many a true thing said in jest.

Hierarchies and 'Lowerarchies' containing Catenae

How does the human brain change with experience? The doctrine of synaptic plasticity is highly flawed, if not outright impossible. There is insufficient time for synapses to undergo systematic change with the kind of sub-second  latency observed in reality. Besides, how would that systematic change be translated into pre- and post-synaptic changes eg in calcium channels, or G-protein transmission? The mechanism of threshold bias adjustment is much more biologically plausible, as well as being supported by the data. These low level changes (which occur because of feedback at the level of the neural circuit) cause higher level changes in data structure architecture. The link between lower and higher levels is covered in a later page. At this higher level, the TDE machine is a make//match machine, rather than a write//read machine. The difference is akin to the separation of subroutine call mechanisms into 'call by reference' and 'call by copy', respectively. 

Each side of the GOLEM diagram contains a data hierarchy. The reproduction hierarchy (a.k.a. the motor channel) is top-down.  The representation hierarchy (a.k.a. the sensor channel) is bottom-up. The reproduction (top-down) side of the GOLEM has the task of constructing meaningful sequences of symbols, while the representation (bottom-up) side of the GOLEM has the task of deconstructing sentences into meaningful symbols (context-dependent recursive subject::predicate pairs). The catenae in the hierarchy (descending links in the tree structure) form syntactic sequences, combining symbols*** at the nth semantic level into sentences at the (n+1)th semantic level****, as per figure 7(a). 

***It is educational to read Al Newell's original paper on the Symbol System Hypothesis (SSH). Although Searle's Chinese Room criticised SSH for its failure to address semantics (and correctly so), this criticism must be a qualified one.  Newell's original paper makes it quite clear that all his symbols are fully semantic. It is all the subsequent third party interpretations that fail to communicate his original intent.

****As an aside, consider the comparison between human languages and arithmetic expressions - although they are allegedly separated by orders of magnitude as far as complexity and coding capacity are concerned, nevertheless they both consist of two types of data - structures and processes, forms and functions, symbols and operators, nouns and verbs, states and changes (transitions).

Hermetic model of semantics

It is essential to understand the many advantages that linguistic memory (LM) systems have over simpler data storage methods, eg 'simple' memory in which a data pattern is read at input, then stored verbatim, or with minimal (eg Lempel-Ziv) attempts at compression.  In terms of basic mechanical metaphors, LM acts more like a typewriter for which you make the letters (Write Once, Read Only), than pen and ink. The Rule:- we can't recognise those representations which are made from memory units (see figure 5(a) ) which we haven't first learned to reproduce. Briefly, we can't recognise any pattern that we haven't first created. This idea has been labelled the 'hermetic' semantics model.

-emic / -etic adjectivals

Languages are built upon semantic-symbolic building blocks called words.  Words are constructed from phonemes. The phoneme is a declaratively coded unit, made from procedurally created phones.  The declarative goal (eg phoneme, a pattern or representation) is held in the right-hand GOLEM channel, in the short-term memory-see fig.7(a)(i) , until the subject is able to reproduce it - eg create the matching sequence of phones in the left-hand GOLEM channel - see fig.7(a)(ii) with sub-liminal (below threshold) error. This mechanism is the source of Ken Pike's -emic / -etic distinction.

Syntagmatic / Paradigmatic combinatorics

Structurally, the combination process is syntagmatic (figure 7(a)(i) ), drawing on a symbol selection process which is paradigmatic (figure 7(a)(ii) ). Empirical evidence for this view is compelling- these two concepts, respectively, form a more precise description of the relative functions in Broca's and Wernicke's areas in the LCH language pipeline. In the syntagmatic process, grammatically constituted units are combined, whereas in the paradigmatic process, context dependent units are selected.

In infancy (see figure 7(b)(i) ) -  the human learns to craft phones into the phoneme patterns heard in adult speech. These speech patterns are stored at embodied, TDE1 level.  In early childhood, this process is largely complete, and language development moves up into the situated, TDE2 level - figure 7(b)(ii). In childhood, the human learns about grammar- not so much how to say them but how to use them. The child is still learning co-articulation skills, as with phones, but at a conceptually higher level.  

Language is an empathy/ experience engine

In adulthood, the human learns about meaning, narrative and the inner life of other subjects - ie each person learns to be skillful at the top-level purpose of language, which is the INTERsubjective transfer of first person experiences. Language is an empathy engine, when used to its full potential, eg in novels. The author tells us what it is like to be in love, to be lonely, to be part of a car chase, to murder someone, to be a cop trying to find a killer, etc. Language puts us in 'the driving seat', if you will.

Sentences are supersymbols (macrosemantics) made from subsymbols called words (microsemantics)

In figure 7(a) we depict a situation in which the subject is using grammatical (syntactic) convention to combine old tokens (ie copies of symbols) into new sentences. But there is nothing unusual about this at all- when the computer programmer 'cuts' code, that is precisely what they are doing, creating or using data variables (symbols with generalised semantics) to create a bunch of little virtual machines, flow-control 'sentences' with procedural semantics. 'Natural' (ie human-readable-writeable) language seems different only because it is a declarative semantic code (it transfers knowledge/facts, not procedure/skills), but other than that, it performs an identical task, linking nth-level semantic symbols to (n+1)th-level semantic supersymbols, or sentences. Lower down, at the level of phonetic coarticulation, an identical process occurs. Language is a virtual fractal.

Declarative coding creates terminal state attractor (TSA) basins

Figure 8 compares the declaratively coded data structures which exist at each of the TDE1, TDE2 and TDE3 levels. At the TDE1 level, the infant human learns to produce, inter alia, many sets of spontaneously produced joint angle sequences which seem to 'feel right', a process called 'motor babbling' - see figure 8(a). In behavioural terms, this process involves somatic 'angles' (homogeneous ccordinates, affine frames), but in language terms, this includes non-somatic reproductions, such as the coarticulated sequences of phones as shown in trilevel memory map figure 6(d)(e) & (f) . The reader must always remember that all cognition is linguistic- the term 'linguistic' must be broadened beyond its traditional communicative interpretation to connote a memory management system (philosophy and implementation):-  behaviour is language as much as language is behaviour**

** Skinner, B.F. (1959) Linguistic Behaviour - Chomsky was only partially correct in his demolishment of Skinner's position. 

Declarative coding automatically generates ambiguity trees

Figure 8 depicts the 'nuts and bolts' of the TDE embodied cognition mechanism.  The difference between conventional data structures and the Terminal State Attractor system derive from declarative coding. Declarative codes are 'ends' based, not 'means' based. In a typical 'before' to 'after' scenario, they depend only on the 'after' values. Therefore each 'after' value creates an attractor 'basin' of potential incoming trajectories, where there are many potential 'before' value candidates. Figure 8 attempts to depict this mechanism at each TDE level. It is the cybernetic drive/setpoint mechanism which acts to constantly reduce error about the target value. Each Goal State (GS) is matched by a satisficing set of target configurations (TC's). There is a many to one relation between targets (TC's) and goals (GS's), reminiscent of the way that several desires (targets) can satisfy the same need (goal). This is an automatic consequence of cybernetics, which uses declarative codes.

TSA mechanism performs discrete feature interpolation

As well as acting about a setpoint to reduce static error, the cybernetic mechanism also acts dynamically about a (potentially iterated) saccadic interval. Consider another of Benjamin Libet's experiments, the one with the green and red lights which are switched on with a small time delay separating them*. Rather than perceiving the 'real' situation (green light illuminates, then red light illuminates), most people perceive a single light which starts to move across the gap, then changes color, apparently in mid-air.  There is only one plausible explanation which satisfies both Occam and engineering viewpoints. This explanation involves interpolation over the time interval of both spatial and chromatic values. Although they actually change in a stepwise, discrete manner, they are perceived as if they were generated by a continuous physical process, eg a moving light which changes color and position smoothly, behind a pair of fixed windows acting as a kind of mask. 

*Dennett, D. (1981) Consciousness Explained

Speaker and Listener have identical data processes, but swapped physical and virtual grounding roles.

Figure 9(a) depicts both types of catenae, ascending 9(a)(i) and descending 9(a)(ii). Figure 9(b) demonstrates how the listener's analysis and the speaker's synthesis are identical and interdependent algorithms. Clearly, a 'mirror neuron' style of information processing underpins all behavioural perception.  Figure 10 depicts ultimate integration between embodied, behavioural and subjective levels under the declarative coding paradigm.