An Associative Memory Device


In my pamphlet on "intelligence" there is mentioned an associative memory based on circles of nerve impulses spreading, on fibres in expanding circles.

In this hypothetical device the active unit is based on the properties of acoustic waves in a peitzioelectric thin layer.

Imagine a flat layer of Aluminium coated with a thin layer of (say) barium titanate.  On the surface away from the aluminium there are two sets (A and B) of arrays of small thin pads of gold connected to a thin gold conductor.

When set A is electrically excited with a pulse pattern of active pads, some active others silent, acoustic-electric circles of wave will spread out from each active centre.

These spreading circles will reach pads of set B and stimulate them to different degrees depending on how many waves activate them.

I believe this pattern of output activation (on set B)  for one pattern of input activation (from set A) is a fourier transform of the input.  So pattern B is the fourier transform of set A.

Now (important) the degree of activation (amplitude of signal at array B) is an analogue function and if each pad remembered how much it was activated in could lower a threshold.

Once this threshold was low enough an electrical circuit linked to a pad on set B would give an pulse for a short time.  This array of pulses would constitute the output of the device.


This arrangement will have the interesting property that after an array of pulses is sent to set A and an output array obtained on set B, then a second, less complete array of pulses arriving at set A will generate are more complete of complete set of pulses out on set B.

An example of its use is the output of an array of photocells strobed at one instant would give a particular output and when the second occasion of a similar, but not identical, input occured this would be "recognised" and a more complete output will occur.

By appropriately gating the input array to the output array the logical difference of the two arrays could be constructed and this would be a further processing of the information.

It occurred to me that the acoustic-electric pulse would decay exponentially away from its source so that points on array B close to the origin on Array A will see bigger pulses.  To overcome this I suggest an additional array C that applies positive feedback to restore the pulse height to its original height.  In this way, with high density arrays the sensors away from the point of origin will see pulses as high as when it originated.  Multiple pulses need to be restored to their original multiple height by having a multiple level quantised trigger.