  |
VAN applications can be broadly categorized into two types (not orthogonal): Making sense of a stream of heterogeneous data, and on-line data (document) classification based on examples and user-defined rules. Both types culminate in ontology development.
Conventionally, formation of ontological models is conceptualized as interplay between inferential mechanisms of induction and deduction. Deduction derives properties of class members from a property asserted on the entire class, while induction derives class property from those of the class members. This conceptualization overlooks the fact that the classes must be already formed, at least hypothetically, for the deduction and induction to operate. That is, ontological hypotheses precede induction and deduction rather than emerge as a result of these inferential mechanisms. John Stuart Mill’s methods of inductive reasoning will illustrate the contention.
According to Mill’s "Logic", there are four inductive methods: the method of agreement, the method of difference, the method of residues, and the method of concomitant variables (Mill also discussed a combination of the first two). All four methods are centered on the notion of causality: "cause of phenomenon as …the antecedent, or concurrence of antecedents, on which it is invariable and unconditionally consequent." Correspondingly, if A under circumstances BC is followed by abc while under circumstances DE it is followed by ade, then A cannot be the cause of either bc or de since they sometimes do not occur when A occurs. However, since a occurs under both sets of conditions, a could be caused by A (method of agreement). Method of difference entails assertion as to whether something other than A can be the cause of a. For that assertion, one verifies occurrence of a without A. If disconfirmed, A is the cause or part of the cause of a, so far as the evidence goes.
Mill’s methods stem from an implicit assumption that a relatively small and well-circumscribed set of observables (A, B, C, D, E and a, b, c, d, e) has been already separated from the manifold of other observables. Clearly, absence of such separation entails combinatorial explosion rendering Mills’ methods impossible. Simply stated, in order to apply Mills methods to a set of observables, one must begin by cutting this set out from the universe of possible observables. In Mill’s "Logic", this crucial first step is overlooked, or remains implicit. Charles Sanders Peirce recognizes this deficit of conventional logic in the concept of abduction:
"The first asserting of a hypothesis and the entertaining of it, whether as a simple interrogation or with any degree of confidence, is an inferential step of abduction (or retroduction). This will include a preference for any one hypothesis over others which would equally explain the facts … I call such inference abduction because its legitimacy depends on altogether different principles from those of other kinds of inference" (Peirce, 1955, p.150).
The VAN model agrees with Peirce that abduction, although preceding logical analysis, by itself is not a logical process but “depends on altogether different principles,” and asserts that those principles are precisely those of dynamic integration and partitioning in the universe of observables implemented in the model.
In summary, the VAN process paves the way to applications of logical calculi, making them computationally feasible. It does so by partitioning large sets of observables into small subsets.
The overall characteristics of VAN computation include the following:
- The process is extremely very fast, based on the proprietary methods of dynamic network partitioning.
- The process is self-improving, that is, user feedback regarding the quality of the output will cause quicker process convergence and error reduction in the subsequent processing.
- VAN system is flexible, so that new entries (e.g., observation sources) can be incorporated within minimal time.
- User complexity is very low, entailing minimal training requirements.
- The process yields accelerated return, as opposed to diminishing return in other AI methods. That is, VAN works faster and more reliably when the size of the data set increases, as opposed to diminishing speed and reliability seen by conventional methods.
|
|
