Daniel M. Berry

Cheriton School of Computer Science
University of Waterloo
Waterloo, ON, Canada

Abstract:

This talk concerns building an artificial intelligence (AI) to do a non-algorithmic task that requires real intelligence. The literature and practice of AI development does not clarify what is a requirements specification (RS) of an AI that allows determining whether an implementation of the AI is correct. Arguing from the idea that an AI must mimic human beings, the talk shows, with two related AIs for stop sign recognition, how traditional measures to evaluate AIs can serve as a major part of an RS for an AI. The talk suggests that (1) these evaluation measures, (2) criteria for acceptable values of these measures, and (3) information about an AI's context informing tradeoffs in these measures constitute an RS of the AI. The talk concludes with some open questions that will be the subject of future work.