GOLEM-NEODE Model Of Mind

brief history from inception in 2010 to theory (2020) - overall aim is to build qualia-governed computer

Preface

The exponential increase in information manipulation capability of machines has been historically viewed in terms of 'intelligence' (= ability to solve problems). This biases any meaningful comparisons against animal minds, leading to the widespread belief that most animals don't share our 'type' of consciousness. This is demonstrably false. Consciousness is evolution's gift to all animals, including humans. Language is evolution's gift to humans, higher apes, parrots and cetaceans.

The very idea that we can attribute thought itself to mental models is not new [13][14]. However, as the conceptual penetration of computers and software into society increases, so also does the associated belief that we are all dependent for day-to-day functioning on mental 'software'. Even environmental problems have been attributed to the mental models we entertain [15].

Our attempts to increase the intelligence of existing computing machines by well known techniques are collectively known as AGI (Artificial General Intelligence) [17]. Somewhat less common are attempts to make computing machines that have the same kinds of intelligence as we do (eg driven by emotions, demonstrating consciousness). This new category has been 'coined' ABI (Artificial Biological Intelligence) [11]. This is to avoid confusion caused by the use of its other label, Strong AI.

There are many reasons why we might wish to build a brain like ours from non-living components. Scientists have long dreamed of becoming immortal. But there are also less idealistic, more practical reasons. For example, we are living longer on average, so more of us fall victim to senile dementia, a spectrum of diseases characterised by cognitive decline attributable to progressive atrophy of brain tissue. Dementia is thought to be caused by beta-amyloid protein deposits in cortical ensembles, or cerebellar damage due to ethanol abuse. It would be amazing to be able to offer sufferers of these incurable conditions the possibility of 'new-for-old' replacement of neuroanatomic pathology.

Computer scientists often lament about the inherent limitations of the processing power of existing 'Von Neumann' (ie Princeton type) computer designs, limitations that don't seem to apply to living brains when asked to perform similar tasks. If we discovered the exact reason (or reasons) for the speed-up of natural systems over artificial ones, then computers could be designed whose number-crunching performance matches, or perhaps even exceeds the levels currently demonstrated by brains.

Ideally, we could get all the most dirty, boring and dangerous jobs done by new robots, by programming their brains according to the latest ABI paradigms. Orwellian-sounding laws guiding the permissible numbers and uses of robots in workplaces would, however, be found necessary, not just nice 'lip-service''. Ulrich-Hatfield Law of metamatics [8] - 'kill 'em all'

The final reason listed here for investigating ABI is to develop new theories of brain, mind and self. It is to this abstract aim that I dedicate this monograph. I have read extensively about other scientists, taking great interest in exactly how they discovered new facts about our worlds and about our selves. I have tried to reproduce their insights, to become a 'scientist's scientist'. 

New Terminology

The strange new terms GOLEM and NEODE are directly analogous to the classic software and hardware subdomains of computer science, respectively.  GOLEM theory is unique in that it claims that cognition (including emotionality and consciousness) can be not just understood, but implemented, using mainstream computer science (the GOF-CS101-view), ie our current knowledge of data structures and algorithms. The GOF-CS101-view that underpins GOLEM-NEODE theory brings belated clarity to the murky, self-serving philosophies of representation and intentionality. By allocating cause to emotions and effects to consciousness (animated semantic states), GOLEM theory has created an opportunity for cognitive philosophers to put their house in order.

For example, philosopher (and fellow Flinders University alumnus) David Chalmers has recently thrown down the gauntlet by placing consciousness in the 'too-hard' basket [4]. GOLEM theory throws down the other gauntlet [12] (remember, gauntlets come in pairs) and makes the counterclaim that the brain is 100% understandable by applying the principles of GOF-CS101 (good-old-fashioned computer science 101 = data structures + algorithms). Maintenance of scientific sanctity (and thus sanity) is ensured! 

Curiously, Chalmers early work is much more optimistic. In it, he identifies the essence of computer programs as follows- they are the cause-effect mechanics (ie semantics) of real-world situations in symbolically coded form.  Yet, within this seminal essay of his, we see the conceptual seeds of his ultimate downfall [6]. Chalmers most frequent argument for the 'hard' problem requires the reader to imagine that zombies are possible. Chalmers claims that zombies are conceivable only if one already believes that structure and function of mental states are insufficient to explain consciousness. Chalmers therefore believes there is a hard problem of consciousness. Whoops, he has forgotten the role of cognitive plasticity.

While it is true that the structure and function of mental states are insufficient to explain consciousness, there is another option than the zombie argument. In addition to structure and function, living creatures also need knowledge (ie stored learning and memorised facts gained from experiences) to form the semantic grounding (SG) [7] which is a necessary prerequisite of all meaningful mental states (eg memories, reasoning = Block's a-type consciousness), not just conscious ones like perceptions (Block's p-type consciousness). This idea is derived from the same general idea that computers must be programmed before they can do useful work. Unlike purely mechanical contraptions, whose feedforward axiomatization is inherent in the physical layout and arrangement of their parts, computers are plastic by design, metaphorical 'lumps of clay', shaped by the 'potter's wheel' of experience. 

SG makes it possible for animals to understand the situations in which they find themselves, for the same reason that knowing the English language is the only thing that makes it possible to understand stories written in that language. Cognitive semantics is the common enabling factor which 'powers' all of the mental states which subvene BOTH a-type experiences (time-shifted subjectivity- memory, reason) and p-type ones (contemporaneous subjectivity - perceptions). It is a combination of these past and present mental states that make us able to extrapolate into the near future, construct hierarchical data structures that can in and of  themselves conceive of the abstract notion of a persistent, trans-episodic mnemonic self. It is the notion of the irreversible (and therefore spatio-temporally unique) self that entails the belief that we are both the agents and the targets (a.k.a. recipients, patients) of our experiences, able to act voluntarily and wilfully in pursuit of individualised versions of evolutionary (ie species-wide) agendas.

GOLEM

(1) Goal-Orientation- Conventional computers are driven by a FETCH-EXECUTE loop. In contrast to this, biological systems such as brains are driven by homeostatic drive states, each one corresponding to the pursuit of a metabolic goal. For example, the drive state known as hunger has as its goal the pursuit of a target item called food. Humans have meta-drive states which behave in a similar manner, but can be redirected at arbitrarily abstract terminal/target states. High functioning people often seek artistic integrity while simultaneously striving for vegan diet and gender-free lifestyle, without feeling confused, conflicted or goal-bound in any way.

 
(2) Linguistic-Cognition - Noam Chomsky used a 'poverty of stimulus' argument to demonstrate the inherent linguistic capability of the newborn human brain. Later, Steven Pinker[1] coined the term 'language acquisition device' (LAD), and tentatively located its function to Broca's and Wernicke's areas, an adjacent pair of anatomical regions in the left hemisphere. Damage to Broca's Area impairs the ability to produce speech (output channel) while damage to Wernicke's Area impairs the ability to understand its meaning (input channel). Subsequent experiments (see Asoulin [16]) demonstrated that the brain's linguistic functions previously believed to have  evolved for external communication between early humans is actually more suited to function as an internal basis for cognition. However, the realisation that cognition is inherently linguistic is itself problematic, since there is as yet no mainstream consensus about the true mechanism  of language. GOLEM theory therefore hypothesises a novel mechanism of linguistic cognition. 

GOLEM theory relies upon common-sense based architectonics, consisting of a simplistic-yet-satisficing two channel (dorsal input vs ventral output) model to help explain both mundane and the mysterious aspects of subject-oriented cognition. It divides cognolinguistic functionality between situated behaviour (implemented as semantic transitions in the output channel a.k.a. differential Δ-semantics) and embodied adaptation (implemented as semantic states within the input channel  a.k.a. cumulative Σ-semantics).  GOLEM is clearly distinguishable from other theories because of its reliance upon the two-channel paradigm.

(3) Emulation of Mind. Emulation of any system (e. g. environmental, biological, or technical) not only imitates the target system with respect to its behaviour and appearance, as with a simulation, but also closely models it in substructure detail and underlying causal mechanism [5]. In the contemporary computer science context, simulation operates at the level of software programming, while emulation operates at the firmware (or ROM) level. GOLEM is an emulation of mind because it consists of (non-living) processes which are functional analogs of those (living) emotional and conscious states which contribute to the production of human intelligence. GOLEM theory is wholly successful in providing a mechanistic solution to Chalmer's so-called 'hard' problem of consciousness (see section 3). A rough-and-ready way to distinguish emulations from simulations is that the former type of modelling is real-time predictive- the model can run in parallel with the reality. One can substitute for the other. For example, a flight simulator undergoes oscillatory motions which attempt to fool the inner ear's balance organs of the learner pilot into believing that they are actually experiencing aerobatic manouvres. The simulator is fixed to the floor of the  hangar- it is not actually going anywhere, unlike the aircraft whose spatial dynamics it is cleverly designed to mimic, an aircraft which is clearly a form of high-speed vehicular transport. A computer emulator by way of comparison is capable of running 'apps' designed for the target architecture in real-time, in a way that cannot be detected interactively, by the normal user of that app, on that platform.

(4) Two-channel 'duplex' model of mind. The GOLEM model extends the normal heuristic use of Ockam's Razor (OZ) by using it as a mandatory design tool at each stage of model development. For example: at the spinal cord, it is apparent that most, though not all, input data flows through dorsal (rear-facing) tracts, while most (...) output data flows through ventral (front-facing) tracts. Therefore OZ mandates that the brain adopt the same two-channel construction as the spinal cord. This leads immediately to a straightforward interpretation of linguistic cognition in which we allocate i-syntax (= behaviour planning) to the output channel, and i-semantics (=behaviour interpretation) to the input channel. 

(5) Synchrononous/asynchronous transitions in each channel. By applying OZ again to the two-channel model, we further subdivide its observable functions into two parts. The discrete thresholded transitions within each channel can be divided into synchronous events above and asynchronous events below the division. Further application of OZ suggests the following distinctions between the channels be made-
(i) that conscious mental qualia be represented in the upper, synchronous region of the input channel (INCH).
(ii) voluntary mental qualia be represented in the upper, synchronous region of the output channel (OUCH).
The added implication is of course that their complementary (ie unconscious and involuntary) non-qualiate states are represented in the lower regions of the input and output channels respectively. This situation is depicted in the upper diagram of the figure below. The lower diagram depicts a putative mapping of the two channel model to the gross neuroanatomical features in each of the cerebral hemispheres. The brain is conceptualised as an interface between the purely physical aspects of objective [9] existence below, and the purely phenomenal aspects of subjective experience above. Note that the cerebellar cortex and basal ganglia of one side is functionally related to the cerebral lobes and cortex of the opposite side via the contralateral decussation. The function of the cerebellum and basal ganglia (according to GOLEM theory) is therefore to provide a means of synchronising the qualiate aspects of the mind, ie to support consciousness.

NEODE

The second major discovery of this research is the NEODE (not an acronym). NEODE is the name for a model of the neuron which is more biologically plausible than the current model, both at the individual unit neuron level, and at the level of neural network adaptation or plasticity. The NEODE gets the first three letters of its name from Neocybernetics, an extension of conventional cybernetics, in which the concept of homeostasis is augmented by the superveiling concept of heterostasis. NEODE circuits use deliberate variation of setpoint values to implement learning and behaviour. 

Neodes (and by implication, real neurons) work on two levels, circuit switching and network balancing.
(a) Circuit Switching. Neuronal output can be off as well as on, depending on the difference between the sum of its dendritic inputs and the membrane's threshold. Warren McCulloch understood this in 1949 [17]. He claimed that it can be approximated by a binary state. Therefore it behaves just like any other digital computer element, implying that, at the unit coding level, brains and computers are directly comparable. 
(b) Network Balancing. While the membrane threshold value is a constant (negative) voltage, the sum of dendritic inputs can be made up of any combination of excitatory and inhibitory values. Indeed, positive dendritic connections are termed 'excitatory' because they tend to cause existing neuronal outputs to increase, or indeed, create neuronal output where there was none. This is due to the excitatory nature of the type of neurotransmitter emitted from the axon's terminal button, which is the 'key' to an excitatory post-synaptic 'lock' whose physical form is a receptor molecule in the post-synaptic cell membrane. Similarly, negative dendritic connections are termed 'inhibitory' because they tend to cause existing neuronal outputs to decrease, or indeed, shut down existing neuronal output. This is due to the inhibitory nature of the type of neurotransmitter emitted from the axon's terminal button, which interacts with an inhibitory post-synaptic receptor molecule in the post-synaptic cell membrane. In the general case, therefore, neuronal cell membranes are equipped with both excitatory and inhibitory receptor molecules. If this were not the case, they could not function as network balancing elements. 

Neodes (ie neurons, if the theory is correct) are sometimes capable of local (ie direct) cybernetic self-regulation. An example is a zeitgeber neuron which functions like a bio-analogue of the crystal oscillator in a digital device. For the vast majority of neodes, their output must be routed through the circuit appropriately, as to act both in a globally meaningful way, as well as promoting stable circuit activations. The stability criterion applied to heterostatic modes of the entire organism has two aspects- (i) embodied activation and (ii) situated behaviour .

As hinted at above, each neode/neuron has a mixture of excitatory and inhibitory inputs, which are themselves outputs coming from other neodes/neurons, and so forth. This seems like an impossibly complex array of values just to keep track of, let alone sensibly govern. The key insight of GOLEM-NEODE THEORY (GNT) is to understand that every governance task, whether biologically or artificially based, can be divided into two clearly separable (i) feedforward (axiomatic, programmatic, command, imperative) and (ii) feedback (automatic, operational, control, comparative) components. The feedforward part (think of a computer program) is essentially static in nature, an information structure, while the feedback part (think of a computer program's execution, ie its running processes) is just as necessarily dynamic in nature, an informational behaviour.

By tonic variation of setpoints (ie setpoint biasing), static learning is achieved. This is how computational/FSM states are implemented in neocybernetic (ie biological) machines. By phasic variation of setpoints (ie setpoint offsetting), dynamic behaviour is produced. This is how computational/FSM transitions are implemented in neocybernetic (ie biological) machines.

Neocybernetics (discovered by M.C. Dyer) incorporates Equilibrium Point Theory (independently discovered by Anatol Feldman). Both researchers (a) were struck by the wrongheadedness of the concept of efference copy (b) realised that voluntary muscular action without self-resistance (the stated aim of efference copy theory ) could instead be achieved by direct manipulation of homeostatic setpoint values. This direct method is simpler, and is therefore more preferable according to OZ. 

Both Feldman's EPT and Dyer use of phasic setpoint offsets are both successful solutions to the problem of voluntary action. Feldman [2] puts it in the following way.."by shifting balance in spatial coordinates, the nervous system converts posture-stabilizing (a static mechanism) to movement-production (a dynamic mechanism), thus solving the classical posture-movement problem". 

However, Dyer alone has understood that the underlying basis of this technique is Perceptual Common Coding [3] itself. Using this generalised insight, Dyer extended the concept by using tonic setpoint biases to solve the problem of implementing neural adaptation ( a.k.a. neuroplasticity). The insight goes as follows. Movement production (a.k.a. behaviour production) is posture change which occurs over a short time scale, whereas sensorimotor adaptation is posture change which occurs over long time scales. But if posture stabilizing mechanisms can be used to implement the former (behaviour), then they are equally well suited to implement the latter (adaptation). The only stipulation is that the neuroplastic changes needed to accurately reproduce successful behaviours may need to be stored separately from those needed to reliably repeat successful adaptation. Again, we note that divergence between neural state and transitions seems very computer-like.

Dyer's theory implicates non-Schwann neuroglia such as stellate and basket cells in the provision of externalised methods of parameterising neural plasticity. This method stands in contrast to existing mechanisms for neural network learning, which are all internally focussed- ie they locate ionic conductance variations within the synaptic cleft, while glibly ignoring the vexed issue of adaptation 'logic' (ie how to turn external error signals into internal circuit parameter adjustments). One of the principles underlying this research (and with an obvious morphosyntactic nod to Occam's Razor) has been coined 'Holmes' Shroud' (SH). It states that if you eliminate the impossible (ie 'call' any deaths, by throwing a logical 'shroud' over the propositional 'corpse') then whatever possibilities remain must include the truth. There is no practical method of finding every input conductance value for each neural soma. How would you allocate the individual conductances within each sum? Its the same problem as being asked to give four numbers that add up to 30; there is no unique solution, just a family of solution sets. Such 'internal' solutions are practically uncountable. While the 'patient' may not yet be dead, (ie a solution may be possible in theory) they are in a coma they will probably never wake from, if you will. Therefore, the truth lies in 'external' solutions, ie those in which there is only one value that needs to be found for each neuron, its output signal. 

1. Pinker, S. (1994) The Language Instinct.

2. Mullick, A.A. · Turpin, N.A. · Hsu, S-C· Subramanian, S.K. Feldman, A.G. · Levin, M.F. (2017) Referent control of the orientation of posture and movement in the gravitational field. J. Experimental Brain Research: Springer-Verlag

3. First brought to general scientific notice by William James in the 19th Century

4. Specifically, he claims that there is an aspect of consciousness which he classifies as a 'hard problem' - beyond the ken of current philosophical analysis.

5. Precise definitions of simulation, emulation and the distinction between them should be possible, but somehow are elusive. I rely upon the reader's common-sense.

6. Carruthers, G. Schier, E. (2012). "Dissolving the hard problem of consciousness". Consciousness Online fourth conference.

7. Garagnani, M., & Pulvermüller, F. (2016). Conceptual grounding of language in action and perception: A neurocomputational model of the emergence of category specificity and semantic hubs. European Journal of Neuroscience, 43(6), 721-737.

8. Metamatics is the putative science of automating automation itself - AI done by machines for machines. Sci Fi today is business as usual tomorrow. 

9. As is the case with so many other examples of terminology in this topic area, the term 'objective' has problematic semantics. In this context, it denotes the state of being on the outside, looking in, a viewpoint that is arguably intersubjective. The seminal Estonian semiotician Jakob Von Uexkull used the german term 'umbegung' to describe this class of informational access, reserving 'umwelt' for the intrasubjective viewpoints, and 'innenwelt' for infrasubjective viewpoints. It is not the words one uses, but the concepts and distinguishing features for each category that are critically important to a non-magical grasp of phenomenal (ie qualiate) matters. Tulving's choice of autonoetic, noetic and anoetic could be used to replace Uexkull's germanisms. 

10. I am fully aware that the title of Skinner's book is 'Linguistic Behaviour'. Maybe we should all give Burrhus a second hearing...

11. ABI seems a reasonable label to use, and comparison of AGI vs ABI (methods and viewpoint) constitutes a reasonable basis upon which to distinguish the two types of scientific endeavour. 

12. Gauntlets were originally the gloves in suits of armour. The term later came to denote a separate glove with specially strengthened cuff or sleeve worn over the wrist to stop injury by dagger or sword.

13. Craik, K. J. W. (1943). The nature of explanation. Cambridge University Press, Cambridge, UK.

14. Johnson-Laird, Philip N (1983). Mental Models: Toward a Cognitive Science of Language, Inference and Consciousness. Harvard University Press. ISBN 978-0-674-56882-2.

15. Jones, N. A., Ross, H., Lynam, T., Perez, P. and Leitch, A. (2011). Mental models: an interdisciplinary synthesis of theory and methods. Ecology and Society 16(1)

16. Asoulin, E. (2016). Language as an instrument of thought. Glossa: a journal of general linguistics 1(1): 46. 1-23

17. Although AGI has been referred to as strong AI, some academic sources (like this website) reserve the term "strong AI" for computer programs that experience sentience or consciousness.

18. McCulloch, W. & Pitts, W. (1943) A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys. 5: 115-133.