- In January 1962 I conceived some model for the human memory . Now (in 1970) while writing down something about it, I feel strongly the inadequacy of the model .
- The reason that I nevertheless describe this model here is that it gives an approach to the problem of the functioning of the human brain that I did notyet find in this form in the literature . Perhaps the model is comparable with the models EPAM, WEPAM and SAL of Feigenbaum and Simon, W.H.Wynn and D.L.Hintzman (see e.g. Computer Reviews nr 13595 and 19032) . Only they use a completely different basic principle (searching is done along some single or multiple tree) from the principle used here .
- I will only describe a model that expresses the basic principle (= directly searching for resembling pictures); many substantial improvements of this model seem possible .
- The model is in terms of actual computing machines and uses the terminology of that, like "registers" and "flipflops" .
- With the currently (1970) available technical apparatus it seems not possible to simulate the concerned model such that each scanning lasts less than 0.1 second .
- See also the note below "How does a human being think ?" (MEMHUMAN) .
- The model.
- A number of, say, 10^12 (10 to the power 12) memory-units, each containing one flip-flop, are arranged in a matrix of 10^9 (10 to the power 9) registers of each 1000 memory-units . In addition there are, say, 5 central registers of each 1000 memory-units .
- The content of one register will be called a picture . The content of the flipflops that are "on" , are called the "points" of the picture .
- One after another, the central registers send out their content , to be called a scan-picture .
- A scan-picture is compared simultaneously with all pictures of the memory; if a picture has more than, say, 100 points in common with the scan-picture, then this picture is written into a central register .
- At any time a few of the registers are in "storing-state" , i.e. taking the current scan-picture as their new content . -- Remark. Logically it is indeed not necessary that the scan-picture is taken as the new content, it could be a picture sent out by quite another register, which could be called eg. a store-buffer-register -- . Which of the registers are in storing-state is a matter of chance; however a number of registers are only very seldom in storing state, e.g. once in 100 years (i.e. once in a lifetime ), while other registers are more often, maybe once a year, once a day, or once a minute ..
- Thus continually there are disappearing pictures from the memory. This is not disturbing because important pictures are recalled often, and also a number of pictures will be kept forever .
- Concluding remarks.
- The well-known association-principle can easily be described within this model .
- The memory should be accompanied by some set of transformational functions, e.g. translation, rotation, contraction, centering, and some in-and-output channels, using the central registers mentioned above .
- It might also be interesting to try to describe in the model more elaborate learning-activities, eg. methods of classification .
- Note from L. Horowitz, Aug. 9 , 1970 ,
- Conceived in spring 1962, written in March 1964, re-written Aug. 9 , 1970 , and typed out with some textual revisions on 13 March 2003 .
- The model is in terms of actual computing machines of 1964. See also the note " An attempt to devise a model for a functional structure of the human memory " (MEMMODEL) .
- How does a human being think ? I imagine this as follows . I am using current (1964) computing terminology .
- At each point in time some information is conscious; let us assume that this information is contained in a central register, always the same register, let us call this the consciousness-register . Suppose this register to be rather large, eg. 100000 (10^5) bits ; it contains at each point of time a simplified copy of a number of things, not only visual, but also motoric, sensoric etc.; such a simplified copy we will sometimes call a "picture" .
- The conscious information is continually changing. The question is: what are the rules for changing ? I imagine this as follows :
- At each point of time a part (maybe 1000 bits) of the conscious information is "active"; this "active part" is being compared with thing outside the consciousness-register, like the memory, impressions of perception etc. This comparing means that there is sought for parts in the memory and/or environment that resemble the "active part" , more precisely: that form a continuation of the "active part" .
- For example: someone hears the gong (bell to announce dinner) ; the sound comes into the consciousness-register; this is compared with the contents in memory (by "scanning" the memory); this recalls the picture of the ringing of a gong near a served table; this picture is brought into the consciousness-register .
- Next the picture "served table" is scanned and this recalls the picture "going downstairs" which comes into the consciousness-register .
- Now I imagine that, if a certain part of the picture has remained the same for about 1/20 second, then it leads to the execution of a command, in this case of a certain movement of the legs .
- So one can say that the conscious information is extending itself
continually. Then it is necessary that also continually parts of the
conscious information are disappearing in a short time, eg. some seconds
or less .
- A further question is: which part of the conscious information is "active" at a fixed point of time ? This can be a random part, so that the "stronger" parts of the conscious information are scanned more often .
- I imagine that the memory consists of a very large number of registers of each about 1000 bits, and that at each point of time contains an amount of information .
- The memory is refilled continually. Firstly with perceptions, secondly with "self-perceptions" , ie. parts of the conscious information, as follows . Continually there are formed from the conscious information new pictures by means of certain simple operations, eg. translation, smaller-making, centering the picture. These pictures are then brought into the memory .
- It will be necessary that continually there are disappearing pictures from memory-registers . This will never be disturbing because there will always be pictures in the memory that resemble to a certain extent the picture that disappears, if only the memory is large enough, like 10^9 (10 to the power 9) registers .
- It seems most simple to suppose that this goes on as follows: some picture must be memorized; chance decides in which register this picture will be stored; this register then is cleared, and filled with the new picture .
- If each memory-register should have the same probability to be refilled, then in a short time all pictures stored longer time ago , would have disappeared . But if eg. for half the memory-registers the average frequency of refilling is 100 years, and for the other half less, eg. some a month, some a day, some a minute, and some 3 seconds , then there are always a large number of pictures of longer time ago in the memory .
- In addition "important pictures" of course are often reconsidered eg. during sleep, and so often re-stored .
- One more remark: there might be as much as 10^8 registers that are refilled each 10 seconds . Then there would be at least 10^8/10 = 10^7 pictures per second that are stored . Maybe at the same time one picture is stored many times . This multiple storing seems interesting from other points of view .
- When will one picture "resemble" another picture ? Let us assume that this depends on the size of the common part of both pictures, more precisely: if the pictures have more than e.g. 100 points in common then the two pictures are called resembling.
- One may make a model of a "scanning-memory" as follows (compare the figure below).
memory (10^9 vertical lines)
|->--|--|--|--|-- ... --|--|--|--> scan-
scan- |->--|--|--|--|-- ... --|--|--|--> lines
input |->--|--|--|--|-- ... --|--|--|--> (10^3
(10^3 .. .. .. .. .. .. hori-
bits) .. .. .. .. .. .. zontal
|->--|--|--|--|-- ... --|--|--|--> lines)
|->--|--|--|--|-- ... --|--|--|-->
o o o o ... o o o assembling gates (10^9)
o o o o ... o o o memorizing gates (10^9)
- The memory we conceive as an array of 10^9 * 1000 points . Each 1000 points form a "memory-register" . With each memory-register is associated an "assembling gate" (used for scanning) and a "memorizing gate" . This memorizing gate is most of the time closed but once in a while it opens and sends a pulse to all the store-gates (see below) of the memory-register .
- With each of the 10^12 points of the array is associated: 1. a flip-flop to remember a 0 or a 1 . 2. a read-gate . The read-gate opens when it receives a pulse from the scan-input , then when the flip-flop contains a 1 it sends out a pulse downwards to the assembling gate . 3. an issuing-gate . The issuing gate opens when it receives a pulse upwards from the assembling-gate , then when the flip-flop contains a 1 it sends out a pulse in horizontal direction . 4. a store-gate . The store-gate opens when it receives a pulse from the memorizing-gate with the effect that the current pulse from the scanregister is stored in the flip-flop .
- The scan-input gives pulses in some of the horizontal wires to
the right . If the intersection-point of such a wire with a certain
memory-register is "loaded positive" , then there will be sent a pulse
downwards to the "assembling gate" of the register . If this "assembling
gate" has received e.g. 100 pulses, then this assembling gate opens and
there goes a pulse upwards back to all the issuing gates of the points of
the memory-register, with the subsequent effect that the picture which is
contained in the register is sent out in horizontal direction . This can
be considered as an output-picture of the current scanning-action .
- All output-pictures of the same scanning action will combine, eg. by simple union, to form the new contents of a 1000-bits-scan-register .
- Much more improvement of the above mechanism is possible . For example: use multiple scan-registers .
- An advantage of the above mechanism is that it does not matter if
some points of the memory are disturbed : there will always be parts of
memory that resemble the disturbed part and that can take over their
- A further advantage of this "scan-memory" is that pictures in a very short time (perhaps 1/1000 sec.) can be compared with memory-contents .
- A scan-memory has only significance if it is very large . Maybe that in the future some sort of "scan-memories" will become valuable devices ; I am thinking e.g. of searching in an information-store for an item that has a large number of given properties .
- But we are not yet so far . We know too little about many things .
And the technical possibilities e.g. of integrated circuits, are (in 1970)
not yet sufficient . However it will be fruitful to consider things as
above from time to time .
_______________________ - home