The virtues of generality
Most machine intelligence systems today are expert systems - they operate in a narrow domain.
If you look at the world of compressors, there are lots of specific-purpose compressors - compressing images, videos and text mostly. These do most of the world's heavy lifting in the area. However, there are also a range of general-purpose compressors.
These are not, typically, a collection of specific-purpose compressors clumped together, with adaptive selection of the algorithm depending on the data type. Rather they just have fewer preconceptions about the form of the data.
A second path to machine intelligence
I think this illustrates a "second path" to machine intelligence - besides the
obvious one of making lots of expert systems in different fields and allowing
their scope to broaden.
General-purpose compressors are interesting - in part - because if you have a good one, you can use it to build specific-purpose compressors of any kind you like.
General-purpose compressors may not be very competitive on particular tasks - when compared with specific-purpose compressors today - but their generality means that they can still find a niche in the marketplace.
The significance of general purpose stream compression for machine intelligence suggests that such systems may become more important in the future.
Tim Tyler |