The Promise of Artificial Intelligence by Brian Cantwell Smith

Author:Brian Cantwell Smith
Language: eng
Format: epub
Publisher: MIT Press

______________

1. The articulative capacities listed in the sidebar are interdependent, both syntactically and semantically. Some of the relations among them are theorized in cognitive science and philosophy under notions of productivity (the fact that cognitive production and comprehension are unbounded), systematicity (the fact that the meanings of whole sentences and whole thoughts are systematically related to the meanings of the words or tokens they are made up of), and compositionality (the fact that the meaning of a complex sentence or thought is determined by its grammatical structure and by the meanings of its constituents). The significance of these combinatoric relations for human cognition is bluntly suimmarized by Jerry Fodor, but the point would generalize to any computational system aiming at achieve general intelligence: “Human cognition exhibits a complex of closely related properties—including systematicity, productivity and compositionality—which a theory of cognitive architecture ignores at its peril. If you are stuck with a theory that denies that cognition has these properties, you are dead and gone.” Jerry Fodor, “Connectionism and the Problem of Systematicity (Continued): Why Smolensky’s Solution Still Doesn’t Work,” Cognition 62, no. 1 (1997): 109–119.

2. These examples are considered nonconceptual because the sense of redness pointed to in the first is likely not redness in general, but the particular red of the sofa—a shade that the speaker is unlikely to possess adequate conceptual resources to describe, and because the patches of wall signified by “here” and “there,” in the second, are unlikely to be well-defined regions with determinate boundaries.

3. John Haugeland, “Analog and Analog,” Philosophical Topics 12, no. 1 (1981): 213–225.

4. See, for example, Gary Marcus, The Algebraic Mind: Integrating Connectionism and Cognitive Science (Cambridge, MA: MIT Press, 2001); the papers in Joe Pater, “Generative Linguistics and Neural Networks at 60: Foundation, Friction, and Fusion,” plus comment articles (Language, 95:1, 2019); and Hector Levesque, Common Sense, the Turing Test, and the Quest for Real AI: Reflections on Natural and Artificial Intelligence (Cambridge, MA: MIT Press, 2017).

One challenge for such projects will be to determine whether helpful insights will emerge in the experience and techniques that have been developed in contemporary machine learning systems trained on the outputs of human conceptualization, such as Wikipedia articles, Twitter feeds, etc. Along with the evident issues of bias, prejudice, etc., discussed in the previous chapter, a serious issue that would need to be addressed, explored in the remainder of this book, is that the data sources these systems use as training sets are not held systematically accountable to the highly variable conceptions (registration schemes) in terms of which they have been formulated. Including the results of any such “data mining” without evaluating the registrational practice underlying each and every one of them (every article, every post) would constitute exactly the sort of “gluing together” that would defeat the aim of the project.

I suggested that this first integrative goal would be easier to address than the second, but at a deeper level doing it properly may depend on successful treatment of the second as well.

5.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.