Electronic network for collective decision based on large number of connections between signals
AssigneesBell Telephone Laboratories, Incorporated, California Institute of Technology
InventorJohn J. Hopfield
Life of this patent
Estimating the term…
Loading patent…
AssigneesBell Telephone Laboratories, Incorporated, California Institute of Technology
InventorJohn J. Hopfield
Estimating the term…
Amplifiers, optionally having complementary (positive and negative) outputs, are connected to a matrix of input and output conductors (where the output conductors are each a pair for the case of complementary amplifier outputs). Each connection is implemented with a resistors R.sub.ij =R.sub.ji connecting the output(s) of amplifiers j to the input of amplifiers i, and vice versa, where i and j are the ith and jth amplifiers not necessarily in sequence. The value of each resistor is selected for the nature of the decisional operation intended to satisfy the following circuit equation of motion ##EQU1## where V.sub.j =g(u.sub.j) the output of amplifier j due to an input u.sub.i, C.sub.i is the input capacitance of amplifier i, and R.sub.i is the equivalent of p.sub.i and R.sub.ij according to the equation ##EQU2## and R.sub.ij =R.sub.ji. For the implementation of an associative memory, only positive (or negative) output terminals need be connected by resistors of unit value to input terminals of amplifiers i and j for the amplifier i in which a binary 1 is to be stored. (The amplifier j is one or more of the other amplifiers.) The outputs of the array of amplifiers will produce the entire word stored in response to a few bit-1 input signals I.sub.i to amplifiers so connected by resistors R.sub.ij. For problem solution, the resistance R.sub.ij =R.sub.ji is selected to have a value that, with appropriate signals at all input conductors (perhaps zero) the network will collectively drive to a stable state at the complementary output terminals which provide an output code word that is a very good solution to the problem.
A network of nonlinear devices having continuous input-output relations for parallel processing of input signals, having useful collective decisional properties to which all of the input signals make a contribution to some degree in the range from 0 to 100%, comprising a matrix of N input and N output conductors, where N is a positive integer greater than 1, and each of said N output conductors may be a pair of complementary positive and negative conductors, N amplifiers of high gain, each amplifier having at least a positive output terminal and optionally a complementary negative output terminal for the case of each output conductor being comprised of a pair, each input conductor of said array being connected to the input terminal of a separate one of said amplifiers, and each amplifier having its output terminal connected to a separate one of said output conductors, each amplifier of a series . . . i, j, k . . . having its output conductor connected to the input conductor of a separate selected amplifier, where each connection is implemented with a resistor R.sub.ij, and the value of each resistor is selected for the nature of the decisional operation intended to satisfy the following circuit equation of motion:
An electronic network for collective decision based on a large number N of signals comprising an array of N amplifiers each designated by a letter from a series . . . i, j, k, . . . connected to a matrix of N input and N output conductors, each connection implemented with equal resistors R.sub.ij and R.sub.ji connecting the outputs of amplifiers j and i to the inputs of amplifiers i and j, where i and j are the ith and jth amplifiers of said series not necessarily arranged in sequence, where the value of each resistor is selected for the nature of the decisional operation intended to satisfy the following circuit equation of motion ##EQU22## where V.sub.j =g(u.sub.i) is the output of amplifier j due to an input u.sub.i and an amplifier gain g, C.sub.i is the input capacitance of amplifier i, and R.sub.i is the equivalent of p.sub.i and R.sub.ij according to the equation ##EQU23## and R.sub.ij =R.sub.ji.
Dependencies are inferred from the claim text.