Inductive inference is like lift


Inductive inference is like lift

Hi! I'm Tim Tyler, and this is a video which compares machine intelligence to powered flight, and inductive inference to lift.

Aviation comparison

Around the turn of the last century, humans had a basic understanding of flight. We were able to construct paper planes, and could build kites and balsa gliders. However, what we could not do is lift and transport significant cargoes.

Today, I think we are at a similar position with respect to machine intelligence. We are able to do some light work - but have not yet mastered the heavy lifting.

Lift

In the case of powered flight, probably the main technical problem that needed to be solved involved figuring out how best to generate lift. We didn't really understand how lift was generated initially - and it turned out not to be obvious. One we had figured that out, flying machines soon followed.

Inductive inference

In machine intelligence, I think we face a similar situation - though this time the problem is inductive inference. Inductive inference is what allows prediction of the future from the past. It is technically equivalent to stream compression. Our key problem is that we do not yet know how to effectively compress streams. It isn't the only problem we face - any more than lift was the only problem facing the Wright brothers. However, it does seem as though mastering inductive inference is the key problem that prevents most machine intelligence projects from getting off the ground.

Strategy

In aeronautical engineering, generating lift wasn't the only problem. We also needed to learn how to make light, strong frames, and build enduring, efficient motors, construct reliable ejector seats - and so on. However, many of those problems were not so specific to flight - and typically there were other people working on them.

If you look at the strategy employed by the Wright brothers one of the things they did in 1901 was to construct a six-foot wind tunnel and conducted extensive tests on miniature wings, to determine which designs would produce the most lift. The small models allowed a rapid build-test cycle and were inexpensive to construct.

I think machine intelligence research could benefit from a similar kind of effort, targetted at inductive inference.


Tim Tyler | Contact | http://matchingpennies.com/