Greg Detre
Tuesday, July 02, 2002
contemporary symbolic AI systems are now too constrained to be able to deal with exceptions to rules, or to exploit fuzzy, approximate, or heuristic fragments of knowledge. Partly in reaction to this, the connectionist movement initially tried to develop more flexible systems, but soon came to be imprisoned in its own peculiar ideology---of trying to build learning systems endowed with as little architectural structure as possible, hoping to create machines that could serve all masters equally well. The trouble with this is that even a seemingly neutral architecture still embodies an implicit assumption about which things are presumed to be "similar."
'specialties'---for the the very act of naming a specialty amounts to celebrating the discovery of some model of some aspect of reality, which is useful despite being isolated from most of our other concerns
Thus, the present-day systems of both types show serious limitations. The top-down systems are handicapped by inflexible mechanisms for retrieving knowledge and reasoning about it, while the bottom-up systems are crippled by inflexible architectures and organizational schemes. Neither type of system has been developed so as to be able to exploit multiple, diverse varieties of knowledge.
Researchers in Artificial Intelligence have devised many ways to do this, for example, in the forms of: Rule-based systems. Frames with Default Assignments. Predicate Calculus. Procedural Representations. Associative data bases. Procedural representations. Semantic Networks. Object Oriented Programming. Conceptual Dependency. Action Scripts. Neural Networks Natural Language.
In the 1960s and 1970s, students frequently asked, "Which kind of representation is best," and I usually replied that we'd need more research before answering that. But now I would give a different reply: "To solve really hard problems, we'll have to use several different representations." This is because each particular kind of data-structure has its own virtues and deficiencies, and none by itself seems adequate for all the different functions involved with what we call "common sense." Each have domains of competence and efficiency, so that one may work where another fails. Furthermore, if we rely only on any single "unified" scheme, then we'll have no way to recover from failure.
section 6.9 of The_Society_of_Mind "The secret of what something means lies in how it connects to other things we know. That's why it's almost always wrong to seek the "real meaning" of anything. A thing with just one meaning has scarcely any meaning at all."
"Mary gave Jack the book" this will produce in you, albeit unconsciously, many different kinds of thoughts (see SOM 29.2)---that is, mental activities in such different realms as: A visual representation of the scene. Postural and Tactile representations of the experience. A script-sequence of a typical script-sequence for "giving." Representation of the participants' roles. Representations of their social motivations. Default assumptions about Jack, Mary and the book. Other assumptions about past and future expectations.
we rarely use a representation in an intentional vacuum, but we always have goals
we must also take into account the functional aspects of what we know, and therefore we must classify things (and ideas) according to what they can be used for, or which goals they can help us achieve
The further a feature or difference lies from the surface of the chosen representation, the harder it will be to respond to, exploit, or adapt to it---and this is why the choice of representation is so important. In each functional context we need to represent particularly well the heuristic connections between each object's internal features and relationships, and the possible functions of those objects. That is, we must be able to easily relate the structural features of each object's representation to how that object might behave in regard to achieving our present goals. This is further discussed in sections 12.4, 12.5, 12.12, and 12.13 of SOM.
But what do we mean by "close" or "near." Decades of research on different forms of that question have produced theories and procedures for use in signal processing, pattern recognition, induction, classification, clustering, generalization, etc., and each of these methods has been found useful for certain applications, but ineffective for others.
It is time to stop arguing over which type of pattern classification technique is best--- because that depends on our context and goal. Instead, we should work at a higher level of organization, discover how to build managerial systems to exploit the different virtues, and to evade the different limitations, of each of these ways of comparing things. Different types of problems, and representations, may require different concepts of similarity. Within each realm of discourse, some representation will make certain problems and concepts appear to be more closely related than others. To make matters worse, even within the same problem domain, we may need different notions of similarity for: Descriptions of problems and goals. Descriptions of knowledge about the subject domain. Descriptions of procedures to be used.
the languages of many sciences, not merely those of Artificial Intelligence and of psychology, are replete with attempts to portray families of concepts in terms of various sorts of spaces equipped with various measures of similarity
Why have logic-based formalisms been so widely used in AI research? I see two motives for selecting this type of representation. One virtue of logic is clarity, its lack of ambiguity. Another advantage is the pre-existence of many technical mathematical theories about logic. But logic also has its disadvantages. Logical generalizations apply only to their literal lexical instances, and logical implications apply only to expressions that precisely instantiate their antecedent conditions. No exceptions at all are allowed, no matter how "closely" they match
Logic theorists seem to have forgotten that in actual life, any expression like "For all X$, P(X)"--that is, in any world which we find, but don't make---must be seen as only a convenient abbreviation for something more like this: "For any thing X being considered in the current context, the assertion P(X) is likely to be useful for achieving goals like G, provided that we apply in conjunction with certain heuristically appropriate inference methods."
It has become particularly popular, in AI logic programming, to restrict the representation to expressions written in the first order predicate calculus. This practice, which is so pervasive that most students engaged in it don't even know what "first order" means here, facilitates the use of certain types of inference, but at a very high price: that the predicates of such expressions are prohibited from referring in certain ways to one another. This prevents the representation of meta-knowledge, rendering those systems incapable, for example, of describing what the knowledge that they contain can be used for.
Furthermore, it must be obvious that in order to apply our knowledge to commonsense problems, we need to be able to recognize which expressions are similar, in whatever heuristic sense may be appropriate.
Indeed, we can think about much of Artificial Intelligence research in terms of a tension between solving problems by searching for solutions inside a compact and well-defined problem space (which is feasible only for prototypes)---versus using external systems (that exploit larger amounts of heuristic knowledge) to reduce the complexity of that inner search.
How can we make Formal Logic more expressive, given that each fundamental quantifier and connective is defined so narrowly from the start. This could well be beyond repair, and the most satisfactory replacement might be some sort of object-oriented frame-based language
While logical representations have been used in popular research, rule- based representations have been more successful in applications. In these systems, each fragment of knowledge is represented by an IF-THEN rule so that, whenever a description of the current problem-situation precisely matches the rule's antecedent IF condition, the system performs the action described by that rule's THEN consequent.
What if several rules match equally well? Of course, we could choose the first on the list, or choose one at random, or use some other superficial scheme---but why be so unimaginative? In SOM, we try to regard conflicts as opportunities rather than obstacles---an opening that we can use to exploit other kinds of knowledge. �
Connectionist - In some instances, this has occurred without any external supervision; furthermore, some of these systems have also performed acceptably in the presence of incomplete or noisy inputs---and thus correctly recognized patterns that were novel or incomplete. This means that the architectures of those systems must indeed have embodied heuristic connectivities that were appropriate for those particular problem-domains.
distribution may oppose parallelism: the more distributed a system is---that is, the more intimately its parts interact---the fewer different things it can do at the same time.
When we simultaneously activate two distributed representations in the same network, they will be forced to interact. In favorable circumstances, those interactions can lead to useful parallel computations, such as the satisfaction of simultaneous constraints. But that will not happen in general; it will occur only when the representations happen to mesh in suitably fortunate ways.
Such problems will be especially serious when we try to train distributed systems to deal with problems that require any sort of structural analysis in which the system must represent relationships between substructures of related types---that is, problems that are likely to demand the same structural resources
For these reasons, it will always be hard for a homogeneous network to perform parallel "high-level" computations
only answer is providing more hardware. More generally, it seems obvious that without adequate memory-buffering, homogeneous networks must remain incapable of recursion, so long as successive "function calls" have to use the same hardware
Each connectionist net, once trained, can do only what it has learned to do.
tend to forget how computationally massive a fully connected neural network is
argues against full connectivity for large, common sense systems
perceptrons: The problem is that it is usually easy to make isolated recognitions by detecting the presence of various features, and then computing weighted conjunctions of them. Clearly, this is easy to do, even in three-layer acyclic nets. But in compound scenes, this will not work unless the separate features of all the distinct objects are somehow properly assigned to those correct "objects." For the same kind of reason, we cannot expect neural networks to be generally able to parse the tree-like or embedded structures found in the phrase structure of natural-language.
This will surely need additional architecture to represent that structural analysis of, for example, a visual scene into objects and their relationships, by protecting each mid-level recognizer from seeing inputs derived from other objects, perhaps by arranging for the object-recognizing agents to compete to assign each feature to itself, while denying it to competitors. This has been done successfully in symbolic systems, and parts have been done in connectionist systems (for example, by Waltz and Pollack) but there remain many conceptual missing links in this area--- particularly in regard to how another connectionist system could use the output of one that managed to parse the scene
Most serious of all is what we might call the Problem of Opacity: the knowledge embodied inside a network's numerical coefficients is not accessible outside that net. This is not a challenge we should expect our connectionists to easily solve. I suspect it is so intractable that even our own brains have evolved little such capacity over the billions of years it took to evolve from anemone-like reticulae. Instead, I suspect that our societies and hierarchies of sub-systems have evolved ways to evade the problem, by arranging for some of our systems to learn to "model" what some of our other systems do
The problem of opacity grows more acute as representations become more distributed
It also makes it harder to learn, past a certain degree of complexity, because it is hard to assign credit for success, or to formulate new hypotheses (because the old hypotheses themselves are not "formulated").
Thus, distributed learning ultimately limits growth, no matter how convenient it may be in the short term, because "the idea of a thing with no parts provides nothing that we can use as pieces of explanation"
But unless a distributed system has enough ability to crystallize its knowledge into lucid representations of its new sub-concepts and substructures, its ability to learn will eventually slow down and it will be unable to solve problems beyond a certain degree of complexity.
Just as networks sometimes solve problems by using massive combinations of elements each of which has little individual significance, symbolic systems sometimes solve problems by manipulating large expressions with similarly insignificant terms, as when we replace the explicit structure of a composite Boolean function by a locally senseless canonical form. Although this simplifies some computations by making them more homogeneous, it disperses knowledge about the structure and composition of the data---and thus disables our ability to solve harder problems. At both extremes---in representations that are either too distributed or too discrete---we lose the structural knowledge embodied in the form of intermediate-level concepts
"thinking," requires facilities for temporarily storing partial states of the system without confusing those memories.
One answer is to provide, along with the required memory, some systems for learning and executing control scripts, as suggested in section 13.5 of SOM. To do this effectively, we must have some "Insulationism" to counterbalance our "connectionism". Smart systems need both of those components, so the symbolic-connectionist antagonism is not a valid technical issue, but only a transient concern in contemporary scientific politics.
Consequently, I expect that the future art of brain design will have to be more like sculpturing than like our present craft of programming. It will be much less concerned with the algorithmic details of the sub-machines than with balancing their relationships
Consider a few of the wonderful bugs that still afflict even our own grand human brains:
Obsessive preoccupation with inappropriate goals.
Inattention and inability to concentrate.
Bad representations.
Excessively broad or narrow generalizations.
Excessive accumulation of useless information.
Superstition; defective credit assignment schema.
Unrealistic cost/benefit analyses.
Unbalanced, fanatical search strategies.
Formation of defective categorizations.
Inability to deal with exceptions to rules.
Improper staging of development, or living in the past.
Unwillingness to acknowledge loss.
Depression or maniacal optimism.
Excessive confusion from cross-coupling.
�Furthermore, there will be a tendency for newly acquired skills to develop from the relatively few that are already well developed and this, again, will bias the largest scale connections toward evolving into recursively clumped, rather than uniformly connected arrangements. A different tendency to limit connectivities is discussed in section 20.8, which proposes a sparse connection-scheme that can simulate, in real time, the behavior of fully connected nets---in which only a small proportion of agents are simultaneously active. This method, based on a half-century old idea of Calvin Mooers, allows many intermittently active agents to share the same relatively narrow, common connection bus. This might seem, at first, a mere economy, but section 20.9 suggests that this technique could also induce a more heuristically useful tendency, if the separate signals on that bus were to represent meaningful symbols.�
�The statistical theories tend to uniformly weight all instances, for lack of systematic ways to emphasize the types of situations of most practical interest. But the AI systems of the future, like their human counterparts, will normally prefer to satisfy rather than optimize---and we don't yet have theories that can realistically portray those mundane sorts of requirements.�