[From Bill Powers (990807.1911 MDT)]
Marc Abrams (990807.1337)--
I have a small problem with Long-term/Short-term memory.
1) When does something become "long-term" vs. "Short-term"?
2) How would memory "know" which features might be useful in the future? It
would seem that the knowledge is either present or not to be used later or.
First we have to understand memory itself, and especially how specific
memories are retrieved.
Suppose you recorded all conversations that took place anywhere in your
house, 24 hours per day. If you faithfully made the recordings on 1-hour
tapes and threw the tapes in a bag for ten years, all the information you
might want would be there in the bag -- but if you haven't labelled the
tapes it might as well be lost. It would take you as long to find a
specific recorded conversation as it would take to listen to half of the
tapes, on the average -- five years. And even then, how would you recognize
the conversation you want if you didn't already know what it was? And if
you already knew what it was, why would you need the recordings?
In B:CP I spoke about the problem of addressing memories. This is a problem
for any kind of memory, conventional or associative, in the brain or in a
collection of tapes. If information is stored, I assume it is stored for a
very long time. But retrieving a specific memory is another matter.
This, in fact, is one of the first problems I had in understanding computer
memories, back in the mid-fifties. If you don't know what is stored in a
given memory unit, how can you ask to get back some information that is
stored in it? There seems to be a paradox: if you already know what
information is stored in it, you don't need to query memory; but if you
don't know what the stored information is, you won't recognize it when you
find it.
The answer is that somehow that memory unit has to be labeled, so the
contents of the unit can be found by looking at the label. The label tells
you something about the stored information other than the information itself.
For example, if the label says 760612.1400, you know that whatever is
recorded on this cassette was recorded on June 12, 1976 from 2:00 PM to
2:59.999 PM. So if you know, from other information, that you recorded a
conversation with your lawyer on June 12, 1976 at about 2:30 PM, all you
have to do is find the tape with that label and play it to find out what
the conversation was about. You don't have to know in advance anything
about the conversation except the date and time it was recorded. You evoke
the "memory" of the conversation by, in this case, using its temporal address.
Where does the address come from? Well, you might keep a written log, in
which you find "Fatal conv. with Lwyr -- 760612.1430". To find the actual
conversation, you refer to a list in which the entries are much shorter
than the conversation, and give you only a pointer to where the whole
conversation is stored. This is known as an "index" and is a way of making
memory searches faster. To search ten years of indices might take you only
a tenth of a year instead of an average of five years. If the index entries
are categorized and sorted, so you can zip right to the L's and find
"Lawyer", then to the F's to find "fatal" then you only need to search the
entries under "L" and "F" and it might take only a hundredth or a
thousandth of a year to find when that fatal conversation was.
But suppose all that you know about the fatal conversation is that the
lawyer was, for some strange reason, wearing a cowboy hat during it. You
don't know _when_ it happened, but you do know part of the content of the
whole experience. Now the temporal index will do you no good. What you need
to do is search all the tapes -- video tapes now, of course -- for
instances in which there is a lawyer and a cowboy hat, with the lawyer
wearing the hat. Two configurations and one relationship.
If you really had to search all the recordings for this information, it
would still take at least 5 years, on the average, to search 10 years of
information, assuming you could look for three items of information at
once. But suppose that somehow you could broadcast these three items to
your library of tapes, and that each tape were connected to a little
librarian that could play that tape to see if it held a lawyer AND a cowboy
hat AND the relationship "HAT ON LAWYER". Each librarian who found a match
would flip a switch turning on a light if the items were found. After
issuing the request "Lawyer and hat and hat-on-lawyer," you would wait one
hour and then look to see what lights were lit. With one librarian per
tape, it would take just one hour to scan all ten years of data and light
all the lights for times when the requested combination was found. This
would require 87,660 librarians, tapes, and tape players, but who cares?
They're small.
Notice that the address we are now using has nothing specific to do with
the actual conversation we're trying to retrieve. It's just a few items
that were "associated" with the conversation for the simple reason that
they existed at the same time the conversation took place and were recorded
as part of the same overall experience -- uh, video tape. There are no
actual associative links, however -- those are formed at the time of
retrieval.
This is called "content-addressed memory." It has the advantage that no
preset addressing scheme is needed as long as everything that happens at
the same time is recorded in parallel. The disadvantage, of course, is that
you do not retrieve just one unique recording. The response of the
librarians gives you ALL the tapes on which the requested three items were
recorded. You then have to search them one at a time to see what they're
about, and there's no guarantee that you'll find the one you really need
first. However, this method is very fast, since all the librarians are
working at once and each deals only with a brief period of time. And with
care, one can truly form unique addresses which yield one unique result.
In B:CP I suggested that human memory may be of the content-addressed type.
If so, then forming useful addresses means making sure that something
unique is being recorded along with the information you may want to
retrieve later. Tricks for memory addressing, I noted, have been known for
a long time and are very effective.
Of course you then have to record the information to be used as an address
in another place, with its own addressing scheme, so when you want to
retrieve some kind of information you first must retrieve the addressing
scheme to be used. But as in the case of the index, above, what now must be
recorded is far less information than the information ultimately to be
retrieved; you don't need to storee the whole book again, but just the
table of contents or the index.
It now strikes me that this sort of arrangement smacks of a hierarchical
arrangement, with the "index" consisting of higher-level perceptions, and
the items to be retrieved consisting of stored lower-level perceptions. I
can't get any farther with that idea, but there it is.
If you like far-out ideas, here is one.
It's been suggested by various people through history that we record all
experiences and never lose any of them, although we might resist recalling
some of them. Certain examples of memory feats might make this suggestion
seem less impossibly profligate with recording equipment.
The problem is that it's very hard to distinguish between failure of the
addressing scheme and disappearance of a recording. Either way, the stored
information seems to have been forgotten. Of course we sometimes recall
experiences that we thought were forgotten, but this doesn't prove that all
apparent forgetting is due to failure of an addressing scheme, although
it's a hint in that direction.
But there is an alternative to either possibility: loss of recordings or
loss of addresses of stored perceptions. It is to say that nothing is
recorded at all, and that what all addressing schemes do is to make contact
with past perceptual signals _at the time they were actually happening_.
This, of course, would be very economical of neural facilities, since no
storage at all would be needed except for the addresses themselves.
Unfortunately, it calls for existence of a brain function which has not
been demonstrated to exist: the ability to form a synaptic connection
between a neuron in the present and a neuron in the past.
Did I ever mention that I used to write science fiction?
Anyhow, the distinction between long-term and short-term memory can be made
several ways. Short-term memory suggests a special-purpose function which
holds perceptions for a short time until memory recordings (and addresses)
can be formed. But that function might simply be a perceptual function
itself. If it takes a while for perceptual signals to die out after the
input is removed, that might serve for short-term memory.
The main difference between short-term and long-term memory is the way they
are addressed for retrieval. There is only one address for short-term
memory: "now." After a while, addresses are formed, and perhaps an index is
filled in, so occurrances farther in the past can be retrieved. It would be
interesting to know whether the critical time-delay is different for
low-level and high-level memories.
As to how memory knows what information might be usable in the future, I
don't think it does. Memory is a tool used by the hierarchy of control
systems. We can vary how we use it, and the attention we give to forming
unique memory addresses has much to do with what we will be able to
retrieve later, and what we will be unable to find later even if it's
really recorded.
The literature on memory phenomena is gigantic. I can speculate freely
because I know so little about it.
Best,
Bill P.
