Information Theory

Information Theory, pioneered at Bell Telephone, is the study of information turned into quantifiable units of data.

In Claude Shannon's formulation of information, the more improbable a piece of data is, the more information or meaning it has. While is seems contradictory to most definitions of signal and noise, the fact is that noise is highly meaningful (representing, for example, a star's explosion 100 million years ago), but generally this unwanted "signal" is considered noise, because while it is improbable, it is generally undesirable data.

So the AI researcher, just like biology, presents carefully separated data, like visual inputs, separated from audio inputs, separated from finance reports (a data base sense organ one could build), the complete works of Shakespeare (a text input sense organ, one could build), etc.

The AI researcher must design their AIs body. Will it perceive bidirectionally-encoded light to make a primitive visual system? Will you add another such system to allow depth cuing? How about audio? You probably want it to perceive audio data in the same dimension of time that your speaking in, so you'll probably build a time-domain 1 (amplitude) or 2 (frequency) dimensional input stream, perhaps a temperature sensor, a tactile surface, etc.

But that's just using similarity to the body. Once you understand the nature of the AI, you can build information networks with senses humans have never had: light polarity, stock market tickers, radio signals of various frequencies, perhaps water platelets.

In every case, it is the consistent presentation of the information to the AI that allows it to create meaningful structures of knowledge. "Meaningful" here, epistemologically, is as related to our own, human structures. One could say that one is not created any new knowledge, therefore, as it arises out of mirroring our own, yet the computation engine applied as it is with electrical power, allows a sort of quasi-human slave...

Okay, enough of this