OpenCogPrime:TheoreticalJustification

From OpenCog
Jump to: navigation, search

Toward a Theoretical Justification of the OpenCogPrime Design

This chapter addresses a conceptual and theoretical question: Why should anyone believe the OpenCogPrime design should actually work, in the strong sense of leading to general intelligence at the human level or ultimately beyond?

The question is addressed here from two different, closely related perspectives.

OpenCogPrime:EssentialSynergies pursues a qualitative argument regarding the power of the different OpenCogPrime:KnowledgeCreation algorithms in OCP to dramatically palliate each others' internal combinatorial explosions.

OpenCogPrime:PropositionsAboutMOSES and OpenCogPrime:PropositionsAboutOpenCogPrime take a somewhat different angle (building on the same conceptual basis), and pursue the question of how a mathematical study of the viability of the OpenCogPrime approach might be made. The mathematical study is left unfinished, and a number of interesting mathematical and quasi-mathematical propositions are formulated but not proved, so this aspect remains a tantalizing speculation. However, I (Ben Goertzel) feel it sheds conceptual light on aspects of the design, even without the (very significant) additional corroboration of mathematical proof.

We begin in OpenCogPrime:PropositionsAboutMOSES, not with OpenCogPrime in general, but by making some theoretical propositions regarding MOSES, the form of Probabilistic Evolutionary Learning used in the current OpenCogPrime system. The MOSES propositions are easier to grasp onto than the OpenCogPrime ones, and represent the same style of theorizing. Following these we proceed to OpenCogPrime:PropositionsAboutOpenCogPrime, which contains some more general propositions pertinent to the overall OpenCogPrime design.

A Note on Mathematical Theory

We now make a few comments on the role of mathematical formulation and proof in AGI.

The task of AGI design admits not only a multitude of technical approaches, but also a multitude of methodological approaches. Even within the scope of "computer science" approaches (as opposed to, e.g., "human brain emulation" approaches) a large degree of methodological variation is possible. For instance there is the distinction between "theory-based" and "pragmatics-based" approaches.

In a theory-based approach, one would begin with a mathematical theory of intelligence, and use this to arrive at an AGI design, and then use the mathematics to justify (either via rigorous proofs or hand-waving-laded physicist-style mathematical demonstrations) the design, and ideally to estimate its degree of intelligence given various amounts of computational resources.

In a pragmatics-based approach, on the other hand, one begins with an AGI design, convinces oneself that it looks plausible, and then builds it and experiments with it. Theories of its behavior may be constructed but the emphasis is on practice, and theory is generally used to resolve specific micro-level points regarding the behavior of the practical system.

The path to OpenCogPrime (via Novamente Cognition Engine, via the Webmind AI Engine, via various toy predecessor systems) began with an attempt to take the theory-based approach, but when this approach proved extremely slow and difficult, a shift to a more pragmatic approach was made. The OpenCogPrime design — like the NCE design to which it bears significant resemblance — was created via a synthesis of ideas from computer science, cognitive science, neuroscience, philosophy of mind and other areas, and the argument for its plausibility and practical viability is a complex one including some appeals to intuition along with more rigorous argumentation.

However, the more rigorous theoretical justification of designs in the OCP/NCE family is still a topic of interest, and does not seem an utter impossibility. We don't consider this kind of justification as necessary for the practical success of OpenCogPrime as an AGI system, but, we do consider it as potentially valuable in terms of fine-tuning the design, understanding which parts are essential and which are less so, and identifying possible shortcomings of the design.


Initial author: Ben Goertzel Note:This wiki document was originally created by Ben Goertzel. Being a wiki page, however, it is subject to revision by others, and Ben Goertzel hence can't be held responsible for whatever it may turn into (for worse or, hopefully, for better). For those who may be curious, the original version of the page is available in the directory http://goertzel.org/original_ocp_wiki/ . Please do not remove this note from this page. — Ben G, June 4, 2008