- It seems that when people are asked to retrieve information about
an episode, they can't yank all the data out in one go. It seems that
they try to reconstruct it (page 284), often by working out what must have happened.
- The model should describe: (a) the organisation of memory; (b) the
processes that retrieve from it; (c) how the organisation is maintained
or changed as new information is added. So it should focus on
representation and on processes.
- So which representations support fast retrieval? Vance can't be
searching lists, at least not if it's a serial search. (The paper
doesn't add the latter qualifier, but it's necessary.)
Also, if experiences were stored in lists, it would probably be easy to
enumerate them. (I say probably, because it's logically possible for the
data to be there, but for the retrieval mechanism to have no access to
it.) But they can't (protocol about museums on page 287).
- Instead, people construct lists of experiences on the fly, as they
narrow in on a description of the event being remembered. This seems to
involve trying to find features (small museum, in London,...) and
then get to experiences from them.
- So there's something important about features. Which
representations make heavy use of features? One is the kind of semantic
net I showed last week. For museums, this might look like the diagram on
page 289, a taxonomy of museum visits.
- But such a network would allow people to recall all
relevant experiences, just by following the links. No need for
false starts (``Was I in a museum in Oxford? No, I was only there for
two hours, and I only had time to see the outside of the buildings'').
- Also, let's assume it's an advantage, in a memory system, if it
allows data to be retrieved by feature. The network above can only
handle this by exhaustive search. That's undesirable, because it slows
down as the number of items increases.
Combining the idea of retrieval-by-feature with human performance, it
seems that we have a network of concepts, with links from features to
these concepts. Retrieval entails gradually narrowing the set of
possible features until it uniquely identifies one item. See diagram on
page 291.
Such an organisation may seem perverse (because of false starts, etc),
but we must realise that any organisation will entail some
computation. It's better to have a small amount of computation to find
relevant features, than a large amount of computation in doing
exhaustive search.
- Given the type of organisation proposed in the diagram, we can see
that the index (types of feature e.g. place, and values of feature e.g.
Europe) should be chosen so that it gives good differentiation between
items. For museum visits, no point in having a feature-type ``window
shape'' with value ``rectangular''!
This implies that indexes should be built as new experiences come in, so
as to give good differentiation. Concepts like ``visits in Europe'' will
be built by generalisation, and so learning will entail discovering how
different episodes resemble one another, and what the significant
differences are.