Discussion 9a: The enigma of Emergence
Although this discussion really should have been Discussion 1, there is a reason why this dialog on artificial intelligence first defined the lower tier elements of the four orders of perception and only later identified the top-tier elements of adaptive behaviors. We cannot begin with those top-tier elements and break them down into constituent sub-elements, then further decompose the sub-elements into even finer components of functionality because the top-tier behaviors are what are characterized as emergent phenomenon.
Without a doubt, it is devilishly difficult to engineer even the simplest of emergent behaviors. And because it is even more difficult to engineer emergent behavior of any complexity, our design process cannot be a top-down, decomposition effort, but must be a “build-up” affair which starts with the lowest elements and creates a functionality that by its very nature demonstrates the top-level emergent behaviors we desire.
After decades of studying the singular example of natural intelligence available and comparing it to the voluminous examples of AI research to date, the design of the Organon Sutra is left with the unwavering conclusion that hierarchical, top-down engineering will not produce emergent intelligence, so whether it is the implementation of biological neurons, or neural networks, or Von Neumann program algorithms, this design is left with only the possibility of providing a conglomeration of low level asynchronous assemblies whose subsequent behavior exhibits emergent intelligence.
And our artificial agent must start out as a stupid agent because its top-level behaviors cannot fully develop until the agent is exposed to the environment it will be adapting to. But we have several advantages that Nature did not have when she evolved natural intelligence, advantages that will allow our agent to develop intelligent behaviors in many orders of magnitude less exposure cycles than human intelligence required. The nature of the first advantage is that we have prior access to the rich body of cultural knowledge already assimilated. And the nature of the second advantage was hinted at when the dialog asked a seminal question at the very end of Discussion 6. Since we can only design the core, native functionality, the dialog asked “At what point can we say that we have engineered a sufficient amount of built-in capability?”
Nature required billions of years and millions of experiments to answer that question, but we have the advantage of hindsight, allowing us to be able to quickly get very close to the critical mass of processes that will ignite the emergent behaviors that we seek, without the need for so much messy experimentation.
But to realize this, we must start with a comprehension of the real nature of massively asynchronous assemblies, which begins with a basic design requirement that stems from the very definition of asynchronous signaling itself. How do we start with the simple activities of low level assemblies who by definition have no higher level temporal ordering between the inputs and outputs of individual signaling elements, and get to higher processes and behaviors that we ultimately desire to engineer?
The answer to that is that process flow must start at the lowest tiers of the systematic and proceed to the higher tiers, at which time the flow reciprocally descends back to the lower levels, in a closed, cyclic fashion. But as most readers will observe immediately, that answer leaves a tremendous amount of unanswered questions. However, at least we now have a starting design foundation which can be broken up into manageable parts.
And this design approach must be balanced with another sometimes antagonistic design concern. How do we avoid the inherent propensity of massively asynchronous assemblies to degenerate into dynamic chaos, assemblies who again by definition have no higher level temporal ordering between inputs and outputs?
An understanding of massively asynchronous assemblies is necessary because of a fundamental aspect of neurophysiology. As Semir Zeki pointed out in his foundational book A Vision of the Brain, “Wherever one looks in the cerebral cortex, one finds certain anatomical rules which are ubiquitous. The first of these rules is that there is no cortical region which is only recipient-all cortical areas have outputs as well as inputs.”
This property of neuroanatomy has been recognized for some time. Sir Charles Sherrington acutely observed this in his landmark treatise The Integrative Action of the Nervous System in 1906, when he wrote “A simple reflex is probably a purely abstract conception, because all parts of the nervous system are connected together and no part of it is probably ever capable of reaction without affecting and being affected by various other parts…”
What this means is that there is no higher level which forms the top of a hierarchy of lower levels, no final destination of processed, abstracted data. Since there is no single cortical area to which all of the specialized processing areas report to, how do all of these individual, seemingly lower-level processes come together and contribute to the singular experience we call consciousness?
It is no wonder that Nature took so long to rationalize these conflicting design considerations. And unfortunately, the powerful concepts expressed in the theories of Cybernetics remain largely silent on the subject of asynchronous assemblies, but we can look to Natures’ solution as the best example to illuminate our engineering path. And the best place to start with is the human neocortex, which we shall examine as we gain an understanding of how the design for our artificial agent will realize its particular massively asynchronous architecture.
Although most AI researchers are familiar with the term neocortex, this dialog will provide a somewhat simplified definition for the interested reader because it hopes to make a discrimination regarding the neocortex that is not always pointed out.
For the purposes of discussion, let’s look at the evolution of the human brain as having developed in four successive stages. (There is an ongoing conversation that a fifth stage is currently underway, a conversation that will be expanded upon in a later part of the dialog, as it will allow the discussion to close the circle of concepts it is outlining throughout the essay into a unified picture of artificial intelligence).
The first stage of this evolution was the development of what is referred to as the reptilian brain, which evolved about 250 million years ago, and as a component of the human central nervous system is now called the brain stem.
The reptilian brain stopped changing relatively soon after it developed, and so the human brain stem is functionally the same as in most reptiles. Although this brain structure is very inflexible in its behaviors and does not learn from experience, the reptilian brain had to be capable, in its own inelegant, hardwired way, of merely surviving in a relatively stable environment. However, more complex structures capable of handling the same basic functions but possessing more sophistication to develop adaptive behaviors were needed.
For this, the second stage of the evolutionary path saw the development of what is referred to as the limbic or mammalian brain, a structure which evolved about 60 million years ago. Like its reptilian counterpart, it has essentially ceased its evolutionary change, but unlike the reptilian brain, this structure exhibits feelings and demonstrates simple mechanisms that learn from experience.
As Nature experimented with the hardwired reptilian brain and the limited learning capabilities of the limbic structure, it also produced a third type of neural assembly whose architecture allowed far more plasticity than the prior two nervous systems. This neural structure is referred to as cerebral cortex and most extant reptiles and mammals are endowed with varying portions of cortical structures. Indeed, as a component of vertebrate nervous systems, the cerebral cortex is still very much evolving.
Therefore, the third stage in the evolution of the human brain is considered to be the pronounced expansion of cerebral cortex in species, culminating with the primates and a select number of other mammalian species.
This process of increasing the brains’ complexity by evolving different neural organizations was accomplished by Natures’ unique form of encephalization. Rather than eliminating the older, less complex organizations, the newer structures simply took control of the older ones and absorbed their functionality, not by turning them on when needed, but by inhibiting their actions while the newer structures dominated the focus of attention.
For many researchers, the cerebral cortex as it developed in humans is considered to be the neocortex (“new cortex”), but for very specific purposes the Organon Sutra dialog is placing the human cerebral cortex into a fourth stage of central nervous system evolution. The cerebral cortex in humans has been divided up into four so-called lobes, named after the bony plates of the cranium which covers each lobe. The four lobes are (from back to front) the occipital lobe, the parietal lobe, the temporal lobe and the frontal lobe. These divisions are mentioned only to point out the contention that it is the pronounced development of the frontal lobe that is unique to humans, and it is this unique development that provided the crucial catalyst in the emergence of intelligence in man.
For this reason, the dialog places the marked development of the frontal lobe into a fourth stage of evolution for the human brain, and it is important to comprehend how this section of the cerebral cortex acts as a catalyst for intelligent behavior, and how this type of structure contributes to artificial intelligence in a massively asynchronous systematic.
It is not the intention of this dialog to get too deep into neuroanatomy, but the interested reader must be able to picture a certain degree of cortex organization in order to comprehend the dynamics in a massively asynchronous assembly. And, even though a certain amount of unfamiliar terminology is being presented, the dialog will attempt to connect the dots for all of the relevant terms in the most straightforward manner possible, so bear with the discussion and you will be rewarded.
Although the dialog introduced the cerebral cortex by first dividing it up into its four constituent lobes, let us step back from that for now and talk about the structure as a whole. (To be anatomically correct, it should first have been pointed out that the cerebral cortex is actually two separate hemispheres, but again, for the sake of discussion, the dialog will simplify the anatomy as much as possible).
The neocortex is organizationally two flat sheets of neurons that together in humans is about a quarter square meter in area and three to four millimeters thick on average. It gets is convoluted and furrowed appearance because the sheet of grey matter has been folded into a very compact form to fit within the cranium.
After the conformal division by lobes, the neocortex is typically parceled into areas of specialization, such as the motor cortex, the somatosensory cortex, and the visual cortex, but the dialog will spare those jargon averse readers from these definitions initially, and discuss the organization of the neocortex from a different perspective.
If we study its histological structure, the neocortex is remarkable in its relative uniformity. So unfold and flatten out this neural carpet in your minds’ eye, and if you could peer close enough, you might see that throughout the entire sheet, the neocortex can be resolved into a laminar organization composed of six distinct layers, all running parallel with the surface of the sheet, with each layer comprised of a predominant neural cell type and arrangement. Various other specific neural cell types are interspersed with the predominant cell types of a given layer also. To be sure, there are exceptions to this uniformity, not all layers are apparent in all areas and even when a layer is present, its thickness and neural organization may vary somewhat. For example, what is typically labeled as the fourth layer is greatly diminished in parts of the primary motor cortex, and greatly enlarged in the somatosensory and primary auditory and visual areas of the neocortex. But even these variations do not bear on the conversation as we examine the neocortex from the perspective of massively asynchronous organizations.
There are a few anatomical aspects which are relevant however. As we mentioned, one of the steps in the evolution of the human brain was the enlargement of the cerebral cortex in species, but it should be pointed out that this enlargement occurred mainly by the expansion of the surface area without a comparable increase in thickness. This detail is pertinent because it tells us that whatever neural organization was responsible for the cytoarchitectural thickness of the cortex, Nature found it acceptable for cerebral functionality in most every species, whether it is a tiny tree shrew or giant ape. And certainly, it is this organization that should be the focus of our investigation for asynchronous systematics.
Perhaps we can see the significant patterns that Nature found to conserve as we examine the cortex further. Beyond the six layers and the minor variations brought on by area specializations, the neural organization throughout the cortex is remarkably homogenous. Once we resolve the layered cytoarchitecture, we find that the only other histologic arrangement or grouping is a columnar organization to the cortical neurons. Once you have pictured the six predominant neural type layers, then envision this laminar structure being honeycombed with invisible cylinders, perpendicular to the layers and the cortex surface. Each cylinder is roughly 300 micrometers wide and spans the thickness of the cortical sheet, and typically hosts about 4000 to 5000 neurons. And with a diameter of about 300 micrometers, there are about 2 million of these structures honeycombed throughout the cortex sheet.
These columnar modules are characteristic in that each column contains all of the major cortical neural cell types, interconnected in the columnar axis dimension. Now, at the risk of getting too detailed, the interested reader should envision these columnar modules as being further sub-divided into mini-columns (again with column axes perpendicular to the surface of the cortex sheet), with each mini-column typically comprised of about 80 to 120 neurons. And it would not be too much of a stretch to say that these mini-columns form the basic functional unit of the cerebral cortex, as a histological examination of these columns reveals essentially more homogeneity. But this definition is advanced only hesitantly, as the dialog does not want to prejudice any model the interested reader might be forming regarding the organizational nature of the neocortex, because as we shall see, definitions of functional specialization in the cerebral cortex will require a fluid, shifting perspective at all times.
Certainly, it was this homogeneity and modularity in its structure that provided the cerebral cortex with the plasticity that evolution was moving toward. But now, envisioning this laminated and honeycombed organization, how do we get this mostly homogenous neural structure of the cerebral cortex to perform the myriad different and hugely complex processes that we observe the human brain performing?
GO TO NEXT DISCUSSION
GO TO TOP OF DISCUSSION
Copyright © 2019 All rights reserved