(These are excerpts from, and additions to, my book "Intelligence is not Artificial")
The Intelligence of Neural Networks
Nobody has a viable definition of "intelligence" but there are different levels at which a machine can simulate what we do. The "Infinite Monkey Theorem" that the French mathematician Emile Borel discussed in his book "Statistical Mechanics and Irreversibility" (1913) probably represents the lowest level: let a monkey type randomly on a typewriter for millions of years and it will eventually produce all the books ever written by humankind.
Ross Ashby in his paper "Design for an Intelligence Amplifier" (1956) calculated that the Brownian motion of the molecules inside a cubic centimeter of air produces the correct binary code
for a trigonometric formula 100,000 times a second. He concluded that a child doodling all day long for long enough would eventually write down the same algebraic formula.
A slightly more "intelligent" (or, at least, more feasible) simulation of our intelligence is the "Chinese Room" imagined by the philosopher John Searle. If you lock in a room a person, who knows absolutely nothing of the Chinese language, and provide her with the list of all possible answers in Chinese to all possible questions in Chinese (no matter how complex the question), a master of Chinese language asking questions in Chinese and reading her answers in Chinese would conclude that the person in the room is a highly literate Chinese scholar when in fact the person locked in the room knows absolutely nothing about Chinese (maybe not even that it is Chinese).
Whether you call "intelligent" Borel's monkey and Searle's room is up to you. It really depends on your definition of "intelligence".
Neural networks fall somewhere in between. Like Borel's monkey they perform a lot of random "typing" (except that here "typing" is really "number crunching") and like Searle's Chinese room they have been provided "rules" (although here they are in the form of "trained" configurations).
A deep-learning practitioner will probably tell you that building a neural network to simulate human behavior is a lot more feasible than setting up Searle's Chinese room. However, no neural network has ever come even remotely close to being able to answer complex questions in Chinese (or any other language).
Back to the Table of Contents
Purchase "Intelligence is not Artificial"