Status Updates From Associative Engines: Connec...

“What simple networks lack… is the ability to learn structure-transforming generalizations.”(73) A structure-preserving generalization takes a stem and adds -ed to form past tense forms. A structure-transforming one can deal with an already formed past-tense form and (take the initiative to) delete the -ed to exhibit the stem, still applying the knowledge derived from structure-preserving generalizations

Because the network doesn’t work on symbols, it is not possible to alter it locally when moving it to a systematically altered environment. There is no discrete stretch of code to apply the modification to. So the network has to be retrained in a “global and nonobvious fashion” (72). What was a source of flexibility compared to symbolism is also a source of inflexibility

So the idea is that whatever the system can be described as doing counts for nothing when we want to know what it does. This is precisely what some would say is not the case with intentional behaviour, but where does intentionality come from?

A competence theory is a description of the effects of the computations, not of the computations themselves. “How the knowledge of spelling-sound correspondence is actually encoded and used is critical, not the fact that it is easier to summarize what the model does in a language that abstracts away from those characteristics.” (Seidenberg 1989: 67)

What there is in the network are just transient responses to input that “reflect the local context in their very structure” (32). There is context-sensitivity all the way down because there is no single thing to “contextualize”. The illusion that there is is like the illusion that words exist.

The context-insensitivity is the eventual result of a process of abstraction, not a basic feature of supposed “symbols”.

“The representation of the knowledge is set up in such a way that the knowledge necessarily influences the course of processing. Using knowledge in processing is no longer a matter of finding the relevant information in memory and bringing it to bear: it is part and parcel of the processing itself.” (McClelland, Rumelhart & Hinton 1986: 32)

Mind as text vs mind as process. Networks’ representations of an item build in their structures (activation patterns) that of its prototype: “the prototyping is a natural effect of the way in which such systems store information” (22). But knowledge is not stored: all there is is the processing. “Changes in the knowledge base and in the representing characteristics thus go hand in hand.” (39)