| Next: | 2.7: The deliberation level | Bookmark | |
| Up: | 2: Part II: Levels of organization in deliberative general intelligence | Monolithic | |
| Prev: | 2.5: The concept level |
Concepts are combinatorial learned complexity. Concepts represent regularities that recur, not in isolation, but in combination and interaction with other such regularities. Regularities are not isolated and independent, but are similar to other regularities, and there are simpler regularities and more complex regularities, forming a metaphorical "ecology" of regularities. This essential fact about the structure of our low-entropy universe is what makes intelligence possible, computationally tractable, evolvable within a genotype, and learnable within a phenotype.
The thought level lies above the learned complexity of the concept level. Thoughts are structures of combinatorial concepts that alter imagery within the workspace of sensory modalities. Thoughts are the disposable one-time structures implementing a non-recurrent mind in a non-recurrent world. Modalities are wired; concepts are learned; thoughts are invented.
Where concepts are building blocks, thoughts are immediate. Sometimes the distance between a concept and a thought is very short; bird is a concept, but with little effort it can become a thought that retrieves a bird exemplar as specific mental imagery. Nonetheless, there is still a conceptual difference between a brick and a house that happens to be built from one brick. Concepts, considered as concepts, are building blocks with ready-to-use concept kernels. A thought fills in all the blanks and translates combinatorial concepts into specific mental imagery, even if the thought is built from a single concept. Concepts reside in long-term storage; thoughts affect specific imagery.
The spectra for "learned vs. invented", "combinatorial vs. specific",
"stored vs. instantiated", and "recurrent vs. nonrecurrent" are conceptually
separate, although deeply interrelated and usually correlated. Some
cognitive content straddles the concept and thought levels. "Beliefs"
(declarative knowledge) are learned, specific, stored, and recurrent.
An episodic memory in storage is learned, specific, stored, and nonrecurrent.
Even finer gradations are possible: A retrieved episodic memory is
learned, specific, and immediate; the memory may recur as mental content,
but its external referent is nonrecurrent. Similarly, a concept which
refers to a specific external object is learned, specific, stored, and
"semi-recurrent" in the sense that it may apply to more than one sensory
image, since the object may be encountered more than once, but still referring
to only one object and not a general category.
| Modalities: | Concepts: | Thoughts: | ||
| Source: | Wired | Learned | Invented | |
| Degrees of freedom: | Representing | Combinatorial | Specific | |
| Cognitive immediacy: | (Not applicable.) | Stored | Instantiated | |
| Regularity: | Invariant | Recurrent | Nonrecurrent | |
| Amount of complexity: | Bounded | Open-ended | Open-ended |
The archetypal examples of "thoughts" (invented, specific, instantiated, nonrecurrent) are the sentences mentally "spoken" and mentally "heard" within the human stream of consciousness. We use the same kind of sentences, spoken aloud, to communicate thoughts between humans.
Words are the phonemic tags (speech), visual tags (writing), gestural tags (sign language), or haptic tags (Braile) used to invoke concepts. Henceforth, I will use speech to stand for all language modalities; "auditory tag" or "phonemic tag" should be understood as standing for a tag in any modality.
When roughly the same concept shares roughly the same phonemic tag within a group of humans, words can be used to communicate concepts between humans, and sentences can be used to communicate complex imagery. The phonemes of a word can evoke all the functionality of the real concept associated with the auditory tag. A spoken sentence is a linear sequence of words; the human brain uses grammatical and syntactical rules to assemble the linear sequence into a structure of concepts, complete with internal and external targeting information. "Triangular lightbulb", an adjective followed by a noun, becomes "triangular" targeting "light bulb". "That is a telephone", anaphor-verb-article-noun, becomes a statement about the telephone-ness of a previously referred-to object. "That" is a backreference to a previously invoked mental target, so the accompanying cognitive description ("is a telephone") is imposed on the cognitive imagery representing the referent of "that".
The cognitive process that builds a concept structure from a word sequence combines syntactic constraints and semantic constraints; pure syntax is faster and races ahead of semantics, but semantic disharmonies can break up syntactically produced cognitive structures. Semantic guides to interpretation also reach to the word level, affecting the interpretation of homophones and ambiguous phonemes.
For the moment I will leave open the question of why we hear "mental sentences" internally - that is, the reason why the transformation of concept structures into linear word sequences, obviously necessary for spoken communication, also occurs internally within the stream of consciousness. I later attempt to explain this as arising from the coevolution of thoughts and language. For the moment, let it stand that the combinatorial structure of words and sentences in our internal narrative reflects the combinatorial structure of concepts and thoughts.
The complexity of the thought level of organization arises from the cyclic interaction of thoughts and mental imagery. Thoughts modify mental imagery, and in turn, mental imagery gives rise to thoughts.
Mental imagery exists within the representational workspace of sensory modalities. Sensory imagery arises from environmental information (whether the environment is "real" or "virtual"); imaginative imagery arises from the manipulation of modality workspace through concept imposition and memory retrieval.
Mental imagery, whether sensory or imaginative, exhibits holonic organization: from the "pixel" level into objects and chunks; from objects and chunks into groups and superobjects; from groups and superobjects into mental scenes. In human vision, examples of specific principles governing grouping are proximity, similarity of color, similarity of size, common fate, and closure [Wertheimer23]; continuation [Moore98]; common region and connectedness [Palmer94]; and collinearity [Lavie96]. Some of the paradigms that have been proposed for resolving the positive inputs from grouping principles, and the negative inputs from detected conflicts, into a consistent global organization, include: Holonic conflict resolution (described earlier), computational temperature [Mitchell93], Prägnanz [Koffka35], Hopfield networks [Hopfield85], the likelihood principle [Helmholtz67]; [Lowe85], minimum description length [Hochberg57], and constraint propagation [Kumar92].
Mental imagery provides a workspace for specific perceptions of concepts and concept structures. A chunk of sensory imagery may be mentally labeled with the concept structure "yellow box", and that description will remain bound to the object - a part of the perception of the object - even beyond the scope of the immediate thought. Learned categories and learned expectations also affect the gestalt organization of mental imagery [Zemel02].
Mental imagery is the active canvas on which deliberative thought is painted - "active canvas" implying a dynamic process and not just a static representation. The gestalt of mental imagery is the product of many local relations between elements. Because automatic cognitive processes maintain the gestalt, a local change in imagery can have consequences for connected elements in working imagery, without those changes needing to be specified within the proximate thought that caused the modification. The gestalt coherence of imagery also provides feedback on which possible changes will cohere well, and is therefore one of the verifying factors affecting which potential thoughts rise to the status of actuality (see below).
Imagery supports abstract percepts. It is possible for a human to reason about an object which is known to cost $1000, but for which no other mental information is available. Abstract reasoning about this object requires a means of representing mental objects that occupy no a priori modality; however, this does not mean that abstract reasoning operates independently of all modalities. Abstract reasoning might operate through a modality-level "object tracker" which can operate independently of the modalities it tracks; or by borrowing an existing modality using metaphor (see below); or the first option could be used routinely, and the second option when necessary. Given an abstract "object which costs $1000", it is then possible to attach concept structures that describe the object without having any specific sensory imagery to describe. If I impose the concept "red" on the existing abstract imagery for "an object which costs $1000", to yield "a red object which costs $1000", the "red" concept hangs there, ready to activate when it can, but not yielding specific visual imagery as yet.
Similarly, knowledge generalized from experience with concept-concept relations can be used to detect abstract conflicts. If I know that all penguins are green, I can deduce that "a red object which costs $1000" is not a penguin. It is possible to detect the conflict between "red" and "green" by a concept-level comparison of the two abstract descriptions, even in the absence of visualized mental imagery. However, this does not mean that it is possible for AI development to implement only "abstract reasoning" and leave out the sensory modalities. First, a real mind uses the rich concept-level complexity acquired from sensory experience, and from experience with reasoning that uses fully visualized imaginative imagery, to support abstract reasoning; we know that "red" conflicts with "green" because of prior sensory experience with red and green. Second, merely because some steps in reasoning appear as if they could theoretically be carried out purely on the concept level does not mean that a complete deliberative process can be carried out purely on the concept level. Third, abstract reasoning often employs metaphor to contribute modality behaviors to an abstract reasoning process.
The idea of "pure" abstract reasoning has historically given rise to AI pathologies and should be considered harmful. With that caution in mind, it is nonetheless possible that human minds visualize concepts only to the extent required by the current train of thought, thus conserving mental resources. An early-stage AI is likely to be less adept at this trick, meaning that early AIs may need to use full visualizations where a human could use abstract reasoning.
Abstract reasoning is a means by which inductively acquired generalizations can be used in deductive reasoning. If empirical induction from an experiential base in which all observed penguins are green leads to the formation of the belief "penguins are green", then this belief may apply abstractly to "a red object which costs $1000" to conclude that this object is probably not a penguin. In this example, an abstract belief is combined with abstract imagery about a specific object to lead to a further abstract conclusion about that specific object. Humans go beyond this, employing the very powerful technique of "deductive reasoning". We use abstract beliefs to reason about abstract mental imagery that describes classes and not just specific objects, and arrive at conclusions which then become new abstract beliefs; we can use deductive reasoning, as well as inductive reasoning, to acquire new beliefs. "Pure" deductive reasoning, like "pure" abstract reasoning, should be considered harmful; deductive reasoning is usually grounded in our ability to visualize specific test cases and by the intersection of inductive confirmation with the deductive conclusions.
Imagery supports tracking of reliances, a cognitive function which is conceptually separate from the perception of event causation. Another way of thinking about this is that perceived cognitive causation should not be confused with perceived causation in real-world referents. I may believe that the sun will rise soon; the cause of this belief may be that I heard a rooster crow; I may know that my confidence in sunrise's nearness relies on my confidence in the rooster's accuracy; but I do not believe that the rooster crowing causes the sun to rise.
Imagery supports complex percepts for "confidence" by tracking reliances on uncertainty sources. Given an assertion A with 50% confidence that "object X is blue", and a belief B with 50% confidence that "blue objects are large", the classical deduction would be the assertion "object X is large" with 25% confidence. However, this simple arithmetical method omits the possibility, important even under classical logic, that A and B are both mutually dependent on a third uncertainty C - in which case the combined confidence is greater than 25%. For example, in the case where "object X is blue" and "blue objects are large" are both straightforward deductions from a third assertion C with 50% confidence, and neither A nor B have any inherent uncertainty of their own, then "object X is large" is also a straightforward deduction from C, and has confidence 50% rather than 25%.
Confidence should not be thought of as a single quantitative probability; confidence is a percept that sums up a network of reliances on uncertainty sources. Straightforward links - that is, links whose local uncertainty is so low as to be unsalient - may be eliminated from the perceived reliances of forward deductions: "object X is large" is seen as a deduction from assertion C, not a deduction from C plus "object X is blue" plus "blue objects are large". If, however, the assertion "object X is blue" is contradicted by independent evidence supporting the inconsistent assertion "object X is red", then the reliance on "object X is blue" is an independent source of uncertainty, over and above the derived reliance on C. That is, the confidence of an assertion may be evaluated by weighing it against the support for the negation of the assertion [Tversky94]. Although the global structure of reliances is that of a network, the local percept of confidence is more likely derived from a set of reliances on supporting and contradicting assertions whose uncertainty is salient. That the local percept of confidence is a set, and not a bag or a directed network, accounts for the elimination of common reliances in further derived propositions and the preservation of the global network structure. In humans, the percept of confidence happens to exhibit a roughly quantitative strength, and this quantity behaves in some ways like the mathematical formalism we call "probability".
Confidence and probability are not identical; for humans, this is both an advantage and a disadvantage. Seeing an assertion relying on four independent assertions of 80% confidence as psychologically different from an assertion relying on a single assertion of 40% confidence may contribute to useful intelligence. On the other hand, the human inability to use an arithmetically precise handling of probabilities may contribute to known cases of non-normative reasoning, such as not taking into account Bayesian priors, overestimating conjunctive probabilities and underestimating disjunctive probabilities, and the other classical errors described in [Tversky74]. See however [Cosmides96] for some cautions against underestimating the ecological validity of human reasoning; an AI might best begin with separate percepts for "humanlike" confidence and "arithmetical" confidence.
Imagery interacts with sensory information about its referent. Expectational imagery is confirmed or violated by the actual event. Abstract imagery created and left hanging binds to the sensory percept of its referent when and if a sensory percept becomes available. Imagery interacts with Bayesian information about its referent: assertions that make predictions about future sensory information are confirmed or disconfirmed when sensory information arrives to satisfy or contradict the prediction. Confirmation or disconfirmation of a belief may backpropagate to act as Bayesian confirmation or disconfirmation on its sources of support. (Normative reasoning in these cases is generally said to be governed by the Bayesian Probability Theorem.) The ability of imagery to bind to its referent is determined by the "matching" ability of the imagery - its ability to distinguish a sensory percept as belonging to itself - which in turn is a property of the way that abstract imagery interacts with incoming sensory imagery on the active canvas of working memory. A classical AI with a symbol for "hamburger" may be able to distinguish correctly spelled keystrokes typing out "hamburger", but lacks the matching ability to bind to hamburgers in any other way, such as visually or olfactorily. In humans, the abstract imagery for "a red object" may not involve a specific red image, but the "red" concept is still bound to the abstract imagery, and the abstract imagery can use the "red" kernel to match a referent in sensory imagery.
Imagery may bind to its referent in different ways. A mental image may be an immediate, environmental sensory experience; it may be a recalled memory; it may be a prediction of future events; it may refer to the world's present or past; it may be a subjunctive or counterfactual scenario. We can fork off a subjunctive scenario from a descriptive scene by thinking "What if?" and extrapolating, and we can fork off a separate subjunctive scenario from the first by thinking "What if?" again. Humans cannot continue the process indefinitely, because we run out of short-term memory to track all the reliances, but we have the native tracking ability. Note that mental imagery does not have an opaque tag selected from the finite set "subjunctive", "counterfactual", and so on. This would constitute code abuse: directly programming, as a special case, that which should result from general behaviors or emerge from a lower level of organization. An assertion within counterfactual imagery is not necessarily marked with the special tag "counterfactual"; rather, "counterfactual" may be the name we give to a set of internally consistent assertions with a common dependency on an assertion that is strongly disconfirmed. Similarly, a prediction is not necessarily an assertion tagged with the opaque marker "prediction"; a prediction is better regarded as an assertion with deductive support whose referent is a future event or other referent for which no sensory information has yet arrived; the prediction imagery then binds to sensory information when it arrives, permitting the detection of confirmation or disconfirmation. The distinction between "prediction", "counterfactual", and "subjunctive scenario" can arise out of more general behaviors for confidence, reliance, and reference.
Mental imagery supports the perception of similarity and other comparative relations, organized into complex mappings, correspondences, and analogies (with Copycat being the best existing example of an AI implementation; see [Mitchell93]). Mental imagery supports expectations and the detection of violated expectations (where "prediction", above, refers to a product of deliberation, "expectations" are created by concept applications, modality behaviors, or gestalt interactions). Mental imagery supports temporal imagery and the active imagination of temporal processes. Mental imagery supports the description of causal relations between events and between assertions, forming complex causal networks which distinguish between implication and direct causation [Pearl00]. Mental imagery supports the binding relation of "metaphor" to allow extended reasoning by analogy, so that, e.g., the visuospatial percept of a forking path can be used to represent and reason about the behavior of if-then-else branches, with conclusions drawn from the metaphor (tentatively) applied to the referent [Lakoff99]. Imagery supports annotation of arbitrary objects with arbitrary percepts; if I wish to mentally label my watch as "X", then "X" it shall be, and if I also label my headphones and remote control as "X", then "X" will form a new (though arbitrary) category.
Thoughts are the cognitive events that change mental imagery. In turn, thoughts are created by processes that relate to mental imagery, so that deliberation is implemented by the cyclic interaction of thoughts modifying mental imagery which gives rise to further thoughts. This does not mean that the deliberation level is "naturally emergent" from thought. The thought level has specific features allowing thought in paragraphs and not just sentences - "trains of thought" with internal momentum, although not so much momentum that interruption is impossible.
At any one moment, out of the vast space of possible thoughts, a single thought ends up being "spoken" within deliberation. Actually, "one thought at a time" is just the human way of doing things, and a sufficiently advanced AI might multiplex or multithread deliberation, but this doesn't change the basic question: Where do thoughts come from? I suggest that it is best to split our conceptual view of this process into two parts; first, the production of suggested thoughts, and second, the selection of thoughts that appear "useful" or "possibly useful" or "important" or otherwise interesting. In some cases, the process that invents or suggests thoughts may do most of the work, with winnowing relatively unimportant; when you accidentally rest your hand on a hot stove, the resulting bottom-up event immediately hijacks deliberation. In other cases, the selection process may comprise most of the useful intelligence, with a large number of possible thoughts being tested in parallel. In addition to being conceptually useful, distinguishing between suggestion and verification is useful on a design level if "verifiers" and "suggesters" can take advantage of modular organization. Multiple suggesters can be judged by one verifier and multiple verifiers can summate the goodness of a suggestion. This does not necessarily imply hard-bounded processing stages in which "suggestion" runs, terminates and is strictly followed by "verification", but it implies a common ground in which repertoires of suggestion processes and verification processes interact.
I use the term sequitur to refer to a cognitive process which suggests thoughts. "Sequitur" refers, not to the way that two thoughts follow each other - that is the realm of deliberation - but rather to the source from which a single thought arises, following from mental imagery. Even before a suggested thought rises to the surface, the suggestion may interact with mental imagery to determine whether the thought is interesting and possibly to influence the thought's final form. I refer to specific interactions as resonances; a suggested thought resonates with mental imagery during verification. Both positive resonances and negative resonances (conflicts) can make a thought more interesting, but a thought with no resonances at all is unlikely to be interesting.
An example of a sequitur might be noticing that a piece of mental imagery satisfies a concept; for a human, this would translate to the thought "X is a Y!" In this example, the concept is cued and satisfied by a continuous background process, rather than being suggested by top-down deliberation; thus, noticing that X is a Y comes as a surprise which may shift the current train of thought. How much of a surprise - how salient the discovery becomes - will depend on an array of surrounding factors, most of which are probably the same resonances that promoted the candidate suggestion "concept Y matches X" to the real thought "X is a Y!". (The difference between the suggestion and the thought is that the real thought persistently changes current mental imagery by binding the Y concept to X, and shifts the focus of attention.)
What are the factors that determine the resonance of the suggestion "concept Y matches X" or "concept Y may match X" and the salience of the thought "X is a Y"? Some of these factors will be inherent properties of the concept Y, such as Y's past value, the rarity of Y, the complexity of Y, et cetera; in AI, these are already-known methods for ranking the relative value of heuristics and the relative salience of categories. Other factors are inherent in X, such as the degree to which X is the focus of attention.
Trickier factors emerge from the interaction of X (the targeted imagery), Y (the stored concept that potentially matches X), the suggested mental imagery for Y describing X, the surrounding imagery, and the task context. A human programmer examining this design problem naturally sees an unlimited range of potential correlations. To avoid panic, it should be remembered that evolution did not begin by contemplating the entire search space and attempting to constrain it; evolution would have incrementally developed a repertoire of correlations in which adequate thoughts resonated some of the time. Just as concept kernels are not AI-complete, sequiturs and resonances are not AI-complete. Sequiturs and resonances also may not need to be human-equivalent to minimally support deliberation; it is acceptable for an early AI to miss out on many humanly obvious thoughts, so long as those thoughts which are successfully generated sum to fully general deliberation.
Specific sequiturs and resonances often seem reminiscent of general heuristics in Lenat's EURISKO [Lenat83] or other AI programs intended to search for interesting concepts and conjectures [Colton00]. The resemblance is further heightened by the idea of adding learned associations to the mix; for example, correlating which concepts Y are frequently useful when dealing with imagery described by concepts X, or correlating concepts found useful against categorizations of the current task domain, bears some resemblance to EURISKO trying to learn specific heuristics about when specific concepts are useful. Similarly, the general sequitur that searches among associated concepts to match them against working imagery bears some resemblance to EURISKO applying a heuristic. Despite the structural resemblance, sequiturs are not heuristics. Sequiturs are general cognitive subprocesses lying on the brainware level of organization. The subprocess is the sequitur that handles thoughts of the general form "X is a Y"; any cognitive content relating to specific Xs and Ys is learned complexity, whether it takes the form of heuristic beliefs or correlative associations. Since our internal narrative is open to introspection, it is not surprising if sequiturs produce some thoughts resembling the application of heuristics; the mental sentences produced by sequiturs are open to introspection, and AI researchers were looking at these mental sentences when heuristics were invented.
Some thoughts that might follow from "X is a Y!" (unexpected concept satisfaction) are: "Why is X a Y?" (searching for explanation); or "Z means X can't be a Y!" (detection of belief violation); or "X is not a Y" (rechecking of a tentative conclusion). Any sequence of two or more thoughts is technically the realm of deliberation, but connected deliberation is supported by properties of the thought level such as focus of attention. The reason that "Why is X a Y?" is likely to follow from "X is a Y!" is that the thought "X is a Y" shifts the focus of attention to the Y-ness of X (the mental imagery for the Y concept binding to X), so that sequitur processes tend to focus selectively on this piece of mental imagery and try to discover thoughts that involve it.
The interplay of thoughts and imagery has further properties that support deliberation. "Why is X a Y?" is a thought that creates, or focuses attention on, a question - a thought magnet that attracts possible answers. Question imagery is both like and unlike goal imagery. (More about goals later; currently what matters is how the thought level interacts with goals, and the intuitive definition of goals should suffice for that.) A goal in the classic sense might be defined as abstract imagery that "wants to be true", which affects cognition by affecting the AI's decisions and actions; the AI makes decisions and takes actions based on whether the AI predicts those decisions and actions will lead to the goal referent. Questions primarily affect which thoughts arise, rather than which decisions are made. Questions are thought-level complexity, a property of mental imagery, and should not be confused with reflective goals asserting that a piece of knowledge is desirable; the two interrelate very strongly but are conceptually distinct. A question is a thought magnet and a goal is an action magnet. Since stray thoughts are (hopefully!) less dangerous than stray actions, question-ness (inquiry) can spread in much more unstructured ways than goal-ness (desirability).
Goal imagery is abstract imagery whose referent is brought into correspondence with the goal description by the AI's actions. Question imagery is also abstract imagery, since the answer is not yet known, but question imagery has a more open-ended satisfaction criterion. Goal imagery tends to want its referent to take on a specific value; question imagery tends to want its referent to take on any value. Question imagery for "the outcome of event E" attracts any thoughts about the outcome of event E; it is the agnostic question "What, if anything, is the predicted outcome of E?" Goal imagery for "the outcome of event E" tends to require some specific outcome for E.
The creation of question imagery is one of the major contributing factors to the continuity of thought sequences, and therefore necessary for deliberation. However, just as goal imagery must affect actual decisions and actual actions before we concede that the AI has something which deserves to be called a "goal", question imagery must affect actual thoughts - actual sequiturs and actual verifiers - to be considered a cognitively real question. If there is salient question imagery for "the outcome of event E", it becomes the target of sequiturs that search for beliefs about implication or causation whose antecedents are satisfied by aspects of E; in other words, sequiturs searching for beliefs of the form "E usually leads to F" or "E causes F". If there is open question imagery for "the cause of the Y-ness of X", and a thought suggested for some other reason happens to intersect with "the cause of the Y-ness of X", the thought resonates strongly and will rise to the surface of cognition.
A similar and especially famous sequitur is the search for a causal belief whose consequent matches goal imagery, and whose antecedent is then visualized as imagery describing an event which is predicted to lead to the goal. The event imagery created may become new goal imagery - a subgoal - if the predictive link is confirmed and no obnoxious side effects are separately predicted (see the discussion of the deliberation level for more about goals and subgoals). Many classical theories of AI, in particular "theorem proving" and "planning" [Newell63], hold up a simplified form of the "subgoal seeker" sequitur as the core algorithm of human thought. However, this sequitur does not in itself implement planning. The process of seeking subgoals is more than the one cognitive process of searching for belief consequents that match existing goals. There are other roads to finding subgoal candidates aside from backward chaining on existing goals; for example, forward reasoning from available actions. There may be several different real sequiturs (cognitive processes) that search for relevant beliefs; evolution's design approach would have been "find cognitive processes that make useful suggestions", not "constrain an exhaustive search through all beliefs to make it computationally efficient", and this means there may be several sequiturs in the repertoire that selectively search on different kinds of causal beliefs. Finding a belief whose consequent matches goal imagery is not the same as finding an event which is predicted to lead to the goal event; and even finding an action predicted to lead to at least one goal event is not the same as verifying the net desirability of that action.
The sequitur that seeks beliefs whose consequents match goal imagery is only one component of the thought level of organization. But it is a component that looks like the "exclamation mark of thought" from the perspective of many traditional theories, so it is worthwhile to review how the other levels of organization contribute to the effective intelligence of the "subgoal seeker" sequitur.
A goal is descriptive mental imagery, probably taking the form of a concept or concept structure describing an event; goal-oriented thinking uses the combinatorial regularities of the concept layer to describe regularities in the structure of goal-relevant events. The search for a belief whose consequent matches a goal description is organized using the category structure of the concept layer; concepts match against concepts, rather than unparsed sensory imagery matching against unparsed sensory imagery. Searching through beliefs is computationally tractable because of learned resonances and learned associations which are "learned complexity" in themselves, and moreover represent regularities in a conceptually described model rather than a raw sensory imagery. Goal-oriented thinking as used by humans is often abstract, which requires support from properties of mental imagery; it requires that the mind maintain descriptive imagery which is not fully visualized or completely satisfied by a sensory referent, but which binds to specific referents when these become available. Sensory modalities provide a space in which all this imagery can exist and interprets the environment from which learned complexity is learned. The feature structure of modalities renders learning computationally tractable. Without feature structure, concepts are computationally intractable; without category structure, thoughts are computationally intractable. Without modalities there are no experiences and no mental imagery; without learned complexity there are no concepts to structure experience and no beliefs generalized from experience. In addition to supporting basic requirements, modalities contribute directly to intelligence in any case where referent behaviors coincide with modality behaviors, and indirectly in cases where there are valid metaphors between modality behaviors and referent behaviors.
Even if inventing a new subgoal is the "exclamation mark of thought" from the perspective of many traditional theories, it is an exclamation mark at the end of a very long sentence. The rise of a single thought is an event that occurs within a whole mind - an intact reasoning process with a past history.
Beliefs - declarative knowledge - straddle the division between the concept level and the thought level. In terms of the level characteristics noted earlier, beliefs are learned, specific, stored, and recurrent. From this perspective beliefs should be classified as learned complexity and therefore a part of the generalized concept level. However, beliefs bear a greater surface resemblance to mental sentences than to individual words. Their internal structure appears to resemble concept structures more than concepts; and beliefs possess characteristics, such as structured antecedents and consequents, which are difficult to describe except in the context of the thought level of organization. I have thus chosen to discuss beliefs within the thought level1.
Beliefs are acquired through cognitive processes that fall into two major classes, inductive and deductive, respectively referring to generalization over experience, and reasoning from previous beliefs. The strongest beliefs have both inductive and deductive support: deductive conclusions with experiential confirmation, or inductive generalizations with causal explanations.
Induction and deduction can intersect because both involve abstraction. Inductive generalization produces a description containing categories that act as variables - abstract imagery that varies over the experiential base and describes it. Abstract deduction takes several inductively or deductively acquired generalizations, and chains together their abstract antecedents and abstract consequents to produce an abstract conclusion, as illustrated in the earlier discussion of abstract mental imagery. Even completely specific beliefs confirmed by a single experience, such as "New Year's Eve of Y2K took place on a Friday night", are still "abstract" in that they have a concept-based, category-structure description existing above the immediate sensory memory, and this conceptual description can be more easily chained with abstract beliefs that reference the same concepts.
Beliefs can be suggested by generalization across an experiential base, and supported by generalization across an experiential base, but there are limits to how much support pure induction can generate (a common complaint of philosophers); there could always be a disconfirming instance you don't know about. Inductive generalization probably resembles concept generalization, more or less; there is the process of initially noticing a regularity across an experiential base, the process of verifying it, and possibly even a process producing something akin to concept kernels for cueing frequently relevant beliefs. Beliefs have a different structure than concepts; concepts are either useful or not useful, but beliefs are either true or false. Concepts apply to referents, while beliefs describe relations between antecedents and consequents. While this implies a different repertoire of generalizations that produce inductive beliefs, and a different verification procedure, the computational task of noticing a generalization across antecedents and consequents seems strongly reminiscent of generalizing a two-place predicate.
Beliefs are well-known in traditional AI, and are often dangerously misused; while any process whatever can be described with beliefs, this does not mean that a cognitive process is implemented by beliefs. I possess a visual modality that implements edge detection, and I possess beliefs about my visual modality, but the latter aspect of mind does not affect the former. I could possess no beliefs about edge detection, or wildly wrong beliefs about edge detection, and my visual modality would continue working without a hiccup. An AI may be able to introspect on lower levels of organization (see Part III), and an AI's cognitive subsystems may interact with an AI's beliefs more than the equivalent subsystems in humans (again, see Part III), but beliefs and brainware remain distinct - not only distinct, but occupying different levels of organization. When we seek the functional consequences of beliefs - their material effects on the AI's intelligence - we should look for the effect on the AI's reasoning and its subsequent decisions and actions. Anything can be described by a belief, including every event that happens within a mind, but not all events within a mind are implemented by the possession of a belief which describes the rules governing that event.
In formal, classical terms, the cognitive effect of possessing a belief is sometimes defined to mean that when the antecedent of a belief is satisfied, its consequent is concluded. I would regard this as one sequitur out of many, but it is nonetheless a good example of a sequitur - searching for beliefs whose antecedents are satisfied by current imagery, and concluding the consequent (with reliances on the belief itself and on the imagery matched by the antecedent). However, this sequitur, if applied in the blind sense evoked by classical logic, will produce a multitude of useless conclusions; the sequitur needs to be considered in the context of verifiers such as "How rare is it for this belief to be found applicable?", "How often is this belief useful when it is applicable?", or "Does the consequent produced intersect with any other imagery, such as open question imagery?"
Some other sequiturs involving beliefs: Associating backward from question imagery to find a belief whose consequent touches the question imagery, and then seeing if the belief's antecedent can be satisfied by current imagery, or possibly turning the belief's antecedent into question imagery. Finding a causal belief whose consequent corresponds to a goal; the antecedent may then become a subgoal. Detecting a case where a belief is violated - this will usually be highly salient.
Suppose an AI with a billiards modality has inductively formed the belief "all billiards which are 'red' are 'gigantic'". Suppose further that 'red' and 'gigantic' are concepts formed by single-feature clustering, so that a clustered size range indicates 'gigantic', and a clustered volume of color space indicates 'red'. If this belief is salient enough, relative to the current task, to be routinely checked against all mental imagery, then several cognitive properties should hold if AI really possesses a belief about the size of red billiards. In subjunctive imagery, used to imagine non-sensory billiards, any billiard imagined to be red (within the clustered color volume of the 'red' concept) would need to be imagined as being gigantic (within the clustered size range of the 'gigantic' concept). If the belief "all red billiards are gigantic" has salient uncertainty, then the conclusion of gigantism would have a reliance on this uncertainty source and would share the perceived doubt. Given external sensory imagery, if a billiard is seen which is red and small, this must be perceived as violating the belief. Given sensory imagery, if a billiard is somehow seen as "red" in advance of its size being perceived (it's hard to imagine how this would happen in a human), then the belief must create the prediction or expectation that the billiard will be gigantic, binding a hanging abstract concept for 'gigantic' to the sensory imagery for the red billiard. If the sensory image is completed later and the concept kernel for 'gigantic' is not satisfied by the completed sensory image for the red billiard, then the result should be a violated expectation, and this conflict should propagate back to the source of the expectation to be perceived as a violated belief.
Generally, beliefs used within subjunctive imagery control the imagery directly, while beliefs used to interpret sensory information govern expectations and determine when an expectation has been violated. However, "sensory" and "subjunctive" are relative; subjunctive imagery governed by one belief may intersect and violate another belief - any imagery is "sensory" relative to a belief if that imagery is not directly controlled by the belief. Thus, abstract reasoning can detect inconsistencies in beliefs. (An inconsistency should not cause a real mind to shriek in horror and collapse, but it should be a salient event that shifts the train of thought to hunting down the source of the inconsistency, looking at the beliefs and assertions relied upon and checking their confidences. Inconsistency detections, expressed as thoughts, tend to create question imagery and knowledge goals which direct deliberation toward resolving the inconsistency.)
Why is the transformation of concept structures into linear word sequences, obviously necessary for spoken communication, also carried out within the internal stream of consciousness? Why not use only the concept structures? Why do we transform concept structures into grammatical sentences if nobody is listening? Is this a necessary part of intelligence? Must an AI do the same in order to function?
The dispute over which came first, thought or language, is ancient in philosophy. Modern students of the evolution of language try to break down the evolution of language into incrementally adaptive stages, describe multiple functions that are together required for language, and account for how preadaptations for those functions could have arisen [Hurford99]. Functional decompositions avoid some of the chicken-and-egg paradoxes that result from viewing language as a monolithic function. Unfortunately, there are further paradoxes that result from viewing language independently from thought, or from viewing thought as a monolithic function.
From the perspective of a cognitive theorist, language is only one function of a modern-day human's cognitive supersystem, but from the perspective of an evolutionary theorist, linguistic features determine which social selection pressures apply to the evolution of cognition at any given point. Hence "coevolution of thought and language" rather than "evolution of language as one part of thought". An evolutionary account of language alone will become "stuck" the first time it reaches a feature which is adaptive for cognition and preadaptive for language, but for which no independent linguistic selection pressure exists in the absence of an already-existent language. Since there is currently no consensus on the functional decomposition of intelligence, contemporary language evolution theorists are sometimes unable to avoid such sticking points.
On a first look DGI might appear to explain the evolvability of language merely by virtue of distinguishing between the concept level and the thought level; as long as there are simple reflexes that make use of learned category structure, elaboration of the concept level will be independently adaptive, even in the absence of a humanlike thought level. The elaboration of the concept level to support cross-modality associations would appear to enable crossing the gap between a signal and a concept, and the elaboration of the concept level to support the blending or combination of concepts (adaptive because it enables the organism to perceive simple combinatorial regularities) would appear to enable primitive, nonsyntactical word sequences. Overall this resembles Bickerton's [Bickerton90] picture of protolanguage as an evolutionary intermediate, in which learned signals convey learned concepts and multiple concepts blend, but without syntax to convey targeting information. Once protolanguage existed, linguistic selection pressures proper could take over.
However, as [Deacon97] points out, this picture does not explain why other species have not developed protolanguage. Cross-modal association is not limited to humans or even primates. Deacon suggests that some necessary mental steps in language are not only unintuitive but actually counterintuitive for nonhuman species, in the same way that the Wason Selection Test is counterintuitive for humans. Deacon's account of this "awkward step" uses a different theory of intelligence as background, and I would hence take a different view of the nature of the awkward step: my guess is that chimpanzees find it extraordinarily hard to learn symbols as we understand them because language, even protolanguage, requires creating abstract mental imagery which can hang unsupported and then bind to a sensory referent later encountered. The key difficulty in language - the step that is awkward for other species - is not the ability to associate signals; primates (and rats, for that matter) can readily associate a perceptual signal with a required action or a state of the world. The awkward step is for a signal to evoke a category as abstract imagery, apart from immediate sensory referents, which can bind to a referent later encountered. This step is completely routine for us, but could easily be almost impossible in the absence of design support for "hanging concepts in midair". In the absence of thought, there are few reasons why a species would find it useful to hang concepts in midair. In the absence of language, there are even fewer reasons to associate a perceptual signal with the evocation of a concept as abstract imagery. Language is hard for other species, not because of a gap between the signal and the concept, but because language uses a feature of mental imagery for which there is insufficient design support in other species. I suspect it may have been an adaptive context for abstract imagery, rather than linguistic selection pressures, which resulted in the adaptation which turned out to be preadaptive for symbolization and hence started some primate species sliding down a fitness gradient that included coevolution of thought and language.
If, as this picture suggests, pre-hominid evolution primarily elaborated the concept layer (in the sense of elaborating brainware processes that support categories, not in the sense of adding learned concepts as such), it implies that the concept layer may contain the bulk of supporting functional complexity for human cognition. This does not follow necessarily, since evolution may have spent much time but gotten little in return, but it is at least suggestive. (This paper's section on the concept level is, in fact, the longest section.) The above picture also suggests that the hominid family may have coevolved combinatorial concept structures that modify mental imagery internally (thoughts) and combinatorial concept structures that evoke mental imagery in conspecifics (language). It is obvious that language makes use of many functions originally developed to support internal cognition, but coevolution of thought and language implies a corresponding opportunity for evolutionary elaboration of hominid thought to coopt functions originally evolved to support hominid language.
The apparent necessity of the internal narrative for human deliberation could turn out to be an introspective illusion, but if real, it strongly suggests that linguistic functionality has been coopted for cognitive functionality during human evolution. Linguistic features such as special processing of the tags that invoke concepts, or the use of syntax to organize complex internal targeting information for structures of combinatorial concepts, could also be adaptive or preadaptive for efficient thought. Only a few such linguistic features would need to be coopted as necessary parts of thought before the "stream of consciousness" became an entrenched part of human intelligence. This is probably a sufficient explanation for the existence of an internal narrative, possibly making the internal narrative a pure spandrel (emergent but nonadaptive feature). However, caution in AI, rather than caution in evolutionary psychology, should impel us to wonder if our internal narrative serves an adaptive function. For example, our internal narrative could express deliberation in a form that we can more readily process as (internal) sensory experience for purposes of introspection and memory; or the cognitive process of imposing internal thoughts on mental imagery could coopt a linguistic mechanism that also translates external communications into mental imagery; or the internal narrative may coopt social intelligence that models other humans by relating to their communications, in order to model the self. But even if hominid evolution has coopted the internal narrative, the overall model still suggests that - while we cannot disentangle language from intelligence or disentangle the evolution of thought from the evolution of language - a de novo mind design could disentangle intelligence from language.
This in turn suggests that an AI could use concept structures without serializing them as grammatical sentences forming a natural-language internal narrative, as long as all linguistic functionality coopted for human intelligence were reproduced in non-linguistic terms - including the expression of thoughts in an introspectively accessible form, and the use of complex internal targeting in concept structures. Observing the AI may require recording the AI's thoughts and translating those thoughts into humanly understandable forms, and the programmers may need to communicate concept structures to the AI, but this need not imply an AI capable of understanding or producing human language. True linguistic communication between humans and AIs might come much later in development, perhaps as an ordinary domain competency rather than a brainware-supported talent. Of course, human-language understanding and natural human conversation is an extremely attractive goal, and would undoubtedly be attempted as early as possible; however, it appears that language need not be implemented immediately or as a necessary prerequisite of deliberation.
| Next: | 2.7: The deliberation level |
| Up: | 2: Part II: Levels of organization in deliberative general intelligence |
| Prev: | 2.5: The concept level |