The 'wiki' section of the site is the progress of being migrated from an old site. There are some rendering issues and many of the links don't work yet.

Long-Term Memory Searches

Discussion - Fire-and-Forget Style

A lot of processes within the humain brain seem to operate on a “fire-and-forget-and-wait” model:

  1. send a one-way request into the abyss
  2. wait for the abyss to mysteriously contain information that can be used for a next step

There’s some subtle implications here for my architecture:

  • memory doesn’t return the ‘result’ directly back to the processor that asked for it; instead, it pushes the result straight into working-memory – perhaps like a sense.
  • the result does not link back to the request (ie: Event.references() doesn’t contain a link to the request)
  • when the result is a list of possible matches, it shouldn’t be contained in some sort of single wrapping MemoryListEvent; instead, each individual item should be pushed into WorkingMemory.
  • the requesting processor, must identify useful results within workingmemory itself, without any direct ‘result’ indication.

This approach seems inefficient and overly loose, but it also fits with another phenomenon:

  • We seem to have to “learn” how to drive our brain. There are many behaviours for which which there is no hard-wiring of an ability to know how to work with them, but rather we discover those behaviours over time and learn how to use them.
  • Thus, the level of separation between request and response fits that. Over time we learn that a response may ‘just appear’ if we concentrate on something.

Advantages:

  • Flat representation model.
  • Feeds into the idea that LTM searches just happen based on whats in WM, continuously in the background, rather than always requiring an explicit request.
  • Easier for other processors to integrate into the process. So seems more biologically plausible.

Disadvantages:

  • Harder to implement the ability to try multiple alternatives
  • WM filled up with noise.
  • Needs the [[Parallel Thought]] model to work well, so that the events from LTM go into WM with low priority, but can still make it in.

Rebuttal:

  • Because Conscious Feedback only gives us a small window into what’s held in WM, humans may well get a single MemoryResponse event into WM, wrapping the list of results. ** But that does require a further processor to extract the results (probably one at a time) in order to make them available for other processors to examine. And how does this MemoryResponseSelectionProcessor decide which item to pick next, when it doesn’t know what the other processors are looking for. The whole thing just seems less biologically plausible.
  • If WM was suddenly filled up with the noise of the results, you’d expect it to come into the CF loop. ** But CF only shows the single top strongest event, so we wouldn’t necessarily know.

Discussion - Memory as a Sense

‘Memory’ could be a sense too. For humans, it appears to behave quite differently to other senses; primarily because it doesn’t behave as a stream, like the other senses. So we perceive it in a very different way, and attribute very different rationalizations to its makeup.

But, if it could be quite reasonable to consider it a sense like any other, and that it feeds in in just the same way.

That would mean:

  • It comes in as an input, not directly back into a processor as a ‘response’ or directly into working memory.
  • Interceptors pick up its data and pre-process them before they are fed into processors and/or directly into working memory, as chosen by the attention-attenuator. Basically, STM and LTM start working almost like processors themselves, with an active component that injects data into the system – as a ‘sense’ input.

It operates in response to triggers, which may be explicit requests, or just a strong correlation to something in working memory at the time. – ie: memories that pop up out of the blue.

(Added 2019-10-27. Labels: work-in-progress)