Human Intelligence

designs

Many Simple Computations

from which emerges

Machine Intelligence

This is pretty much what emergent systems try (and arguably succeed) to do. However, a lot of 'emergent systems' are just a tool for creating a complicated computation system. Sometimes we have to tinker with the inputs, but it's not random - we have reasonable expectations about what needs to be tweaked. So some of our intelligence is going into the system. In some of the genetic annealing programs that generate code, we can analyze and understand the output. I think these are just convoluted transfer intelligence systems.

So we can't design away the transfer intelligence problem. I think actual machine intelligence will be more like a discovery. The inspiration for this comes from Wolfram's A New Kind of Science. I've heard mixed reviews of this book, but I haven't gotten into the parts where he applies his cellular automata to science. Which is good, because all I want to talk about is randomness.

Wolfram shows that a lot of systems that exhibit seemingly random output are driven by random input. These are basically 'transfer randomness' systems. But somehow you still need to get randomness into these systems. Where does that randomness come from?

For Wolfram, that randomness emerges from cellular automata number 30. This CA is extremely simple (8 basic rules) however it produces uncorrelated random behaviour with any non-zero amount of input. The output simply cannot be predicted based on the input unless you actually run the CA. I'd also say that it is impossible to 'tweak' the system to make a small change in output behavior. In my mind, this systems exhibits true emergence.

I think this same kind of emergence exists for machine intelligence. But I don't really need to call it 'machine' intelligence - I'll just call it intelligence. I think that there are some programs that exhibit true emergent intelligence. We won't be able to predict or tweak the output, and we won't understand how it works - it just will. It definitely won't be transfer intelligence - any sort of input intelligence beyond defining the system's rules will be unnecessary.

Wolfram also found evidence for something like a minimum threshold for complexity in cellular automata. If the rules are too simple, complexity does not emerge. I think a similar sort of minimum threshold for intelligence exists in emergent systems. Obviously we haven't reached it yet.

Wolfram showed that a lot of other systems mapped directly to celluar automata. So I don't have to specify which particular kind of emergent system will be the source of intelligence. Rob and I did an interesting exercise - we mapped his particular evolutionary system on to an existing evolutionary system. I suspect there is a set of basic properties for emergent systems.

( "Suspect" ? I sound like Wolfram.... )

So how do we discover intelligence? I think maybe we've been aiming too low. As far as I can see, the emergent systems in development are all designed to be simple and small enough that we can understand them. The goals are always very simple - push a block across a plane, or something like that. That's not how intelligence exists in the real world. Even the simplest creatures in nature that exhibit intelligence have thousands of sensory inputs. We can't reasonably expect to pick a few specific inputs and get the same level of intelligent output. We need to throw it all into the mix and see what happens.

Now, you probably won't be able to do this on your PC. In fact, I think PC's are kind of holding AI back. Everyone wants to run these simulations on the desktop, but to do so the scope of the experiment has to be severly cut back. Basically we make intelligent guesses as to what might work. I think maybe Wolfram has the right idea - years of brute-force searching in the simulation space. Because if we really have to discover it, intelligent guesses aren't going to work. Serious hardware is called for.

Anyway. The End.