Despite the cacophony of disagreement surrounding the definition of intelligence, artificial or otherwise, we cling to the illusion that it is something we can detect and even measure. The Flynn Effectis the observation that measured human intelligence is steadily rising over time. It is the hominal version of Moore's Law, which does not purport to measure intelligence, per se, but rather the increasing density of transistors in a given swath of silicon real estate over time. The increases due to Moore's Law are relatively easy to determine because we can measure the number of transistors per economic unit for a specific processor and point to the producing fabrication facilities as the source. Nothing as definitive can be said of human intelligence, except that we appear to be getting smarter on average, if IQ tests are an appropriate measure.
There seems to be an increasing level of intelligence—this thing we cannot adequately define but are very capable of testing for various purposes. Whether silicon based or human, the general level of intelligence is rising. The difference between our digital offspring and ourselves is one of degree, not direction. The following questions, posed in the first part of this series, must be considered:
- Where exactly is "intelligence," artificial and otherwise, headed?
- Are we heir to a "teleology of intelligence," or simply stumbling forward in a drunkard's random walk?
- Is there a discernible design or purpose that propels us forward?
The answer to that set of questions will fundamentallydetermine how we approach the artisan/artilect dichotomy we may face as Moore's Law continues to unfold:
- How will we respond to the emergence of entities that have the potential to dramatically surpass human intelligence, individually and collectively?
- If such entities are possible, how long do we have before we come face-to-interface with them?
- What might they think of us?
- Are we now at the dawn of that epoch?
To even begin contemplating a response, we require a presuppositionalframework, a "world view" of some sort. To play the game, we must presume something in the way of a universe external to our own personal, limited scope of direct experience. To that end, it should be clear that increasing intelligence is either anatural outcome of the universe or it is not. We may posit that the increasing intelligence we are witness to—whether it is biological or digital, whether one necessarily leads to the other—is an emergent property of the universe in which we live. Alternatively, we may dismiss such notions and retreat to the random-walk explanation for all manifestations of complexity, of which increasing intelligence is simply another example, and therefore no more remarkable than any other evolutionary development. Regardless of presuppositional framework, we have two choices when it comes to the teleology of intelligence: Either intelligence is natural, emergent, and inevitable, or it is random, accidental, and ultimately entropic.
One temptation we should avoid is a retreat to a "random emergence" explanation when it comes to discussions of the teleology of intelligence, as it again unnecessarily muddies the water. If we cannot acknowledge an intimate relationship between physical laws and the phenomenon of emergence, then we will not find safe harbor in either science or faith. When it comes to complexity, either emergence is a function of the underlying physical laws, or the physical laws are a function of the underlying emergence; therefore, a "random emergence" explanation is a non-starter. In this series I refer to the new phase of artificial intelligence (AI) exploration as the "Hawkins/de Garis epoch," after Palm founder Jeff Hawkins and professor Hugo de Garis, both of whom are pursuing general-purpose, human-competitive intelligence devices. To illustrate the two opposing approaches, Hawkins represents the artisan and de Garis the artilect.
Generally speaking, all classic AI systems fall into one or both of these two categories:
- Darwinian: Systems that learn or evolve based on some selection criteria
- Rules-based: Systems that model the knowledge domain and infer a set of rules based on the model
An a priori inference model is required in either case. With rules-based systems, the knowledge domain must be modeled in advance. With Darwinian systems, selection criteria (in other words, a teleology) is required. The inherent problem, therefore, is one of building a better model. Even Darwinian systems face this problem; the tuning of selection criteria in order to augment outcome is a common practice—indeed, augmentation in order to influence outcome itself betrays a teleological bias. All this is to say that the artisan ethos is alive and well. The fact that we have certain human awareness regarding truth and beauty, which leads us to favor one outcome over another, should not be a surprise when it comes to any endeavor, including work in AI. But then doesn't this awareness strongly indicate a presuppositional framework?
Human intelligence is different from any theory or attempt at AI so far. The artisan understands this difference on a gut level. Hawkins' theory of human intelligence, while compelling, does not address this visceral aspect of the human experience, relying instead on a pure neocortex explanation for human intelligence.
According to Hawkins, the hominal approach to data processing consists of four major organizational elements, all aspects of the structure of the neocortex, which:
- Stores sequences of patterns: The order in which memories flow is based on sequential (temporal) storage structures. A good example of this is the flow of the melody in a song; we remember melodies in sequence as opposed to reverse order, or some other method. All human memory, according to Hawkins, is sequentially stored and accessed.
- Recalls patterns auto-associatively: Memories are stored in the synaptic connections between neurons. While a vast amount of information is stored in the neocortex, only a few things are actively remembered at one time, because auto-associative memory ensures that only the particular part of the memory that is immediately relevant to the current context is activated. On the basis of these activated memory patterns, predictions are made—without us being aware of it—about what will happen next.
- Stores patterns in an invariant form: The neocortex constantly receives sequences of patterns of information, which it stores by creating invariant representations. These are memories that are independent of salient details that handle variations in the world automatically. For instance, you can recognize your friend's face after he has shaved his beard.
- Stores patterns in a hierarchy: The hierarchical structure of the neocortex plays an important role in perception and learning. Low regions in the structure of the neocortex make low-level predictions (information like color, time, and tone) regarding what is expected next, while higher-level regions make higher-level predictions about more abstract things. We "understand" something when the neocortex's prediction fits with the new sensory input. Whenever neocortex patterns and sensory patterns conflict, there is confusion and our attention is drawn to the anomaly, the details of which are then dispatched to higher neocortex regions to check if the situation can be understood on a higher level. If there are patterns somewhere else in the neocortex that more accurately fit the current sensory input, the anomaly is resolved. Otherwise, cognitive dissonance ensues and appropriate corrective measures (such as learning) are taken.
Learning requires the building of internal models. During repetitive learning, memories of the world first form in higher regions of the cortex, but as we learn they are reformed in lower parts of the cortical hierarchy. Well-learned patterns are represented low in the cortex while new information is sent to higher parts. Slowly, the neocortex builds in itself a representation of the world it encounters. According to Hawkins, "The real world's nested structure is mirrored by the nested structure of your cortex."
The Hawkins brain architecture provides a plausible explanation for the efficiency and great speed of the human brain while dealing with complex tasks of a familiar kind. It is interesting to note, however, that if Hawkins is correct, then it follows that we do not actually see or hear precisely what is occurring at any point in time. When people are talking, by definition we don’t fully listen to what they say. Rather, we are constantly predicting what they will say next, and as long as there seems to be a fit between prediction and incoming sensory information, our attention remains rather low. Only when they say something that actively conflicts with our prediction do we pay attention. It is therefore the unexpected that provides opportunities for learning—sort of like stumbling across a well-made watch in a verdant meadow on an uninhabited island.
This metaphor of discovering and explaining the unexpected watch in the meadow was first proposed by biologist Richard Dawkins and used in his 1986 book to illustrate his view of evolution and complexity. The blind watchmaker thesis posits two basic elements: variation (that is, mutation) and cumulative natural selection. Variation is the randomly occurring changes, at a genetic level, produced by spurious events such as DNA errors in the cell reproduction process. Generally, such mutations are either neutral or harmful in effect. Sometimes, however, such changes slightly improve an organism's ability to survive and reproduce. Under these favorable mutation conditions, the stage is set for the second element of the thesis with is cumulative natural selection.
Biological organisms generally produce more offspring than can or will survive to maturity. Offspring that possess mutative advantages can be expected to produce more viable descendants, all things being equal, than their non-mutant peers. The advantageous mutative trait eventually spreads throughout the species and becomes the norm for further cumulative improvements in succeeding generations. Given sufficient time and advantageous mutations, enormously complex organs and patterns of adaptive behavior can eventually be produced in tiny cumulative steps, without the assistance of any preexisting intelligence. All this can happen, that is, if the theory is true.
"One way or another, Darwinists meet the question 'Is Darwinism true?' with an answer that amounts to an assertion of power: 'Well, it is science, as we define science, and you will have to be content with that.' Some of us are not content with that, because we know that the empirical evidence for the creative power of natural selection is somewhere between weak and non-existent. Artificial selection of fruit flies or domestic animals produces limited change within the species, but tells us nothing about how insects and mammals came into existence in the first place. In any case, whatever artificial selection achieves is due to the employment of human intelligence consciously pursuing a goal. The whole point of the blind watchmaker thesis, however, is to establish what material processes can do in the absence of purpose and intelligence. That Darwinist authorities continually overlook this crucial distinction gives us little confidence in their objectivity."
(The Religion of the Blind Watchmaker by Phillip E. Johnson)
The self-consistent castle of belief, which subscribers to Dawkins' blind watchmaker thesis is built on, rests on a foundation that is no less a leap of faith than any other religion; it is simply less honest. Religions at least admit to a basis in faith. Dawkins, a self-described atheist, seems to not even be aware that his own presuppositional framework is also bound by a belief that does not rest on a logical proof or material evidence.
When I asked for his views on the teleology of intelligence, Ben Goertzel, who, like Hawkins and de Garis, also happens to be working on general-purpose human-competitive systems, responded, "I don't really believe in teleology, but I feel intuitively that there may be a tendency in the universe for the amount of pattern to increase over time—if so, then increasing intelligence would be a part of this...".
Intuition? Increasing "pattern?" Okay. So what is that?
Much of Goertzel's own AI work predates and foreshadows observations in the Hawkins opus in terms of the nature of human intelligence (although Goertzel goes well beyond Hawkins with speculations regarding the universal nature of consciousness itself). If anything, Hawkins' On Intelligence falls short of the mark that Goertzel's prolific theoretical work has established with respect to a complete proposal for a general-purpose, human-competitve machine. (See the Artificial General Intelligence Research Institute site as well as Goertzel's home page for a threshold view of his world.) Indeed, the four salient aspects of human intelligence that Hawkins articulates are also reflected in Goertzel's own work, though presented in different terminology. While Goertzel has come to some similar conclusions with respect to the nature of human intelligence, he goes beyond Hawkins by identifying other characteristics of human intelligence, including the twin complexity-era concepts of attractorsand emergence, neither of which is addressed by Hawkins in what is otherwise a reductionistapologetic.
But for Goertzel to side-step the notion of a teleology of intelligence with concepts like "intuition" and a universal tendency toward "pattern increase" is a little awkward, to say the least. It seems that when evidence fails to fit our models and cannot be rationally explained otherwise, it becomes part of that "other" category, or an intuitive understanding that people simply have, requiring no further examination. Ironically, this too betrays a presuppositional framework that humans have an intuitive modality for knowledge acquisition that we cannot or will not explain, but will effectively ignore when it comes to defining and understanding the knowledge creation process.
But then Goertzel also genuflects at the altar of emergence, which itself now seems to be a conceptual attractor for a growing audience of reformed reductionists. See Nobel Laureate Robert Laughlin's recent book A Different Universe - Reinventing Physics from the Bottom Downfor a wonderful discussion of this emerging Age of Emergence. The gnarly problem with a full-frontal embrace of emergence, however, is the fact that the label simply masks entire swaths of unexplained phenomena with a term that simply allows us to move on without further consideration. By definition, the emergent processes or behaviors cannot be explained otherwise and are therefore emergent. Punt. Declare scientific victory and move on.
Upon closer examination, a growing array of assumptions on which much of our current understanding of the universe is based rests on feet of emergent clay. Solid matter itself, for example, is an emergent property of self-organized molecules that do not individually reflect any of the properties of this emergent feature. Emergence also applies to (that is, masks our complete understanding of) capital 'L' Life, consciousness, human thought, and the entire field of evolutionary studies.
Perhaps Dawkins' watchmaker is not as blind as he would have us believe, with belief itself being the critical factor. The leap of faith that Dawkins must embrace in order to codify his religion of the Blind Watchmaker is no less than the one required to accept more of a Nearsighted Watchmaker; one who sees very well the very, very small, and who understands the essence of the inherent emergence above based on properties so well designed below. Perhaps, in the end, it will always reduce to a matter of belief.
If the artilect were given to laughter, I am sure the previous paragraph would have elicited a bit. But the artilect, like much of AI itself, is entirely inscrutable from our perspective. As such, we cannot know what makes it laugh or even if it does. Although according to Hugo de Garis, the unknowable nature of much of classic AI, not to mention the wildly improved Artilect version 2.0 to come, isn't of concern, as the artilect will far surpass human intelligence in fairly short order—just as soon as "... neuroscience understands why intelligence levels of human beings differ because of differences in their neural structures, [then] it will be possible to create an intelligence theory, which can be used by neuro-engineers to increase the intelligence of the artificial brains they build."
Are we clear on that? Once we fully understand human intelligence, which is an emergent (that is, we can't actually explain it) property of the underlying neural structures, we will be able to engineer around those principles in order to build better artificial brains. Makes perfect sense, doesn't it?
In all fairness, both Hawkins and de Garis would probably not assert that human intelligence is an emergent property of the underlying neural structures; we can give credit to Goertzel and Laughlin for those insights. Which is to say that the Hawkins/de Garis epoch is probably the last chapter of reductionist thinking in the AI space—and the end of research in classic AI, at least insofar as the creation of a general-purpose, human-competitive machine is concerned.
The real Artilect War is not one that de Garis has imagined; it is not some Bill Joy nightmare that pits man against machine or man against machine builder. The real Artilect War is the clash of wills that inevitably occurs when human civilizations collide. The quest for general-purpose, human-competitive machines is no less a measure of civilization than any other technical achievement. And how we view the world is our presuppositional framework for civilization itself. It is that collision of civilizations, through the looking glass of AI, that is chronicled in the final chapter in this series.