Weighted summation circuitry with digitally controlled capacitive structures

Abstract

An array of weighted summation circuits, N in number, each generate a weighted sum response to the same plurality of input signals, M in number. Each weighted summation circuit includes at least one corresponding capacitive element for determining the weighting of each of the input signals within that weighted summation circuit. At least one corresponding capacitive element is of a programmable type having its capacitance value determined in accordance with the bits of a digital word received at a control word port thereof. The array of weighted summation circuits are preferably constructed in integrated circuit form together with an interstitial memory having respective word storage elements for temporarily storing the digital words applied to the control word ports of nearby capacitive elements in the integrated circuitry.

Description

The invention generally concerns neural nets the processors of which use capacitors to perform weighted summation in accordance with Coulomb's Law.

BACKGROUND OF THE INVENTION

U.S. Pat. No. 5,039,871 issued 13 Aug. 1991 to W. E. Engeler, entitled "CAPACITIVE STRUCTURES FOR WEIGHTED SUMMATION AS USED IN NEURAL NETS" and assigned to General Electric Company is incorporated herein by reference. U.S. Pat. No. 5,039,870 issued 13 Aug. 1991 to W. E. Engeler, entitled "WEIGHTED SUMMATION HAVING DIFFERENT-WEIGHT RANKS OF CAPACITIVE STRUCTURES" and assigned to General Electric Company is incorporated herein by reference. These patents explain in their backgrounds of invention the desirability of using weighted summation networks in neural network layers, which weighted summation networks use capacitive elements to determine in accordance with Coulomb's Law the weighting accorded their synapse input signals. Y. P. Tsividis and D. Anastassion in a letter "Switched-Capacitor Neural Networks" appearing in ELECTRONICS LETTERS, 27th Aug. 1987, Vol. 23, No. 18, pages 958,959 (IEE) describe a switched capacitor method of implementing weighted summation in accordance with Coulomb's Law that is useful in analog sampled-data neural net systems. U.S. Pat. Nos. 5,039,870 and 5,039,871 describe methods of implementing weighted summation in accordance with Coulomb's Law that do not require capacitors to be switched.

It not only is desirable to use capacitive elements to determine in accordance with Coulomb's Law the weighting accorded analog synapse input signals, as described specifically in U.S. patent applications Ser. Nos. 366,838 and 366,839, it is desirable to use capacitive elements to determine in accordance with Coulomb's Law the weighting accorded digital synapse input signals. U.S. patent application Ser. No. 546,970 filed 1 Aug. 1990 by W. E. Engeler, entitled "NEURAL NETS SUPPLIED DIGITAL SYNAPSE SIGNALS DIGITAL ON A BIT-SLICE BASIS" and assigned to General Electric Company describes digital synapse input signals being so weighted. So does U.S. patent application Ser. No. 628,257 filed 14 Dec. 1990 by W. E. Engeler, entitled "DIGITAL CORRELATORS INCORPORATING ANALOG COMPUTER STRUCTURES OPERATED ON A BIT-SLICED BASIS" and assigned to General Electric Company.

SUMMARY OF THE INVENTION

Apparatus embodying the invention includes an array of weighted summation circuits, N in number, each weighted summation circuit generating a weighted sum response to the same plurality of input signals, M in number, and including at least one corresponding capacitive element for determining the weighting of each of the input signals within that weighted summation circuit. At least one corresponding capacitive element is of a programmable type having its capacitance value determined in accordance with the bits of a digital word received at a control word port thereof. In preferred embodiments of the invention the array of weighted summation circuits are constructed in integrated circuit form together with an interstitial memory having respective word storage elements for temporarily storing the digital words applied to the control word ports of nearby capacitive elements in the integrated circuitry.

BRIEF DESCRIPTION OF THE DRAWING

FIG. 1 is a schematic diagram of a neural net layer as described by W. E. Engeler in U.S. patent application Ser. No. 366,838 entitled "NEURAL NET USING CAPACITIVE STRUCTURES CONNECTING INPUT LINES AND DIFFERENTIALLY SENSED OUTPUT LINE PAIRS", which neural net layer uses capacitors to perform weighted summations of synapse signals to be subsequently sensed and non-linearly amplified to generate axon response signals, which capacitors may in accordance with the present invention have their capacitance values programmed in response to a digital signal.

FIGS. 2A and 2B together form a FIG. 2 that is a schematic diagram of a modification of the FIG. 1 neural net that can be made manifold times to provide in accordance with a further aspect of the invention, for the programmable weighting of the capacitances used in performing weighted summation of synapse signals.

FIG. 3 is a schematic diagram of a neural net layer as described by W. E. Engeler in U.S. patent application Ser. No. 366,839 entitled "NEURAL NET USING CAPACITIVE STRUCTURES CONNECTING OUTPUT LINES AND DIFFERENTIALLY DRIVEN INPUT LINE PAIRS", which neural net layer uses capacitors to perform weighted summations of synapse signals to be subsequently sensed and non-linearly amplified to generate axon response signals, which capacitors may in accordance with the present invention have their capacitance values programmed in response to a digital signal.

FIGS. 4A and 4B together form a FIG. 4 that is a schematic diagram of a modification of the FIG. 3 neural net that can be made manifold times to provide in accordance with a further aspect of the invention, for the programmable weighting of the capacitances used in performing weighted summation of synapse signals.

FIG. 5 is a schematic diagram illustrating one way of pulsing the non-linear output driver amplifiers, as may be used in a FIG. 1 neural net layer modified manifoldly per FIG. 2, or as may be used in a FIG. 3 neural net layer modified manifoldly per FIG. 4.

FIGS. 6A and 6B together form a FIG. 6 that is a schematic diagram of another neural net described by W. E. Engeler in U.S. patent application Ser. No. 366,838, which other neural net uses pairs of input lines driven by balanced input signals for connection to the pairs of differentially sensed output lines by weighting capacitors connected in quad configurations and operated as full bridges, which capacitors may in accordance with the present invention have their capacitance values programmed in response to a digital signal.

FIG. 7 is a schematic diagram of training apparatus described by W. E. Engeler in U.S. patent applications Ser. Nos. 15 366,838 and 366,839, which training apparatus can be used with the FIG. 1 neural net layer manifoldly modified per FIG. 2, with the FIG. 3 neural net layer manifoldly modified per FIG. 5, or with the FIG. 1 neural net layer manifoldly modified per FIG. 6.

FIG. 8 is a schematic diagram of a system described by W. E. Engeler in U.S. patent applications Ser. Nos. 366,838 and 366,839, which system has a plurality of neural net layers each constructed in accordance with FIG. 1 modified manifold times per FIG. 2, with FIG. 3 modified manifold times per FIG. 5, or with FIG. 1 modified manifold times per FIG. 6.

FIG. 9 is a conceptual schematic diagram of a pair of capacitors having capacitances that sum to a constant value and that are programmable responsive to digital words encoding weighting factor in two's complement arithmetic.

FIG. 10 is a table indicating the disposition of capacitive elements as components of the FIG. 9 pair of capacitors as determined by the digital words encoding weighting factor in two's complement arithmetic.

FIG. 11 is a conceptual schematic diagram of a pair of capacitors having capacitances that sum to a constant value and that are programmable responsive to digital words encoding weighting factor in one's complement arithmetic.

FIG. 12 is a table indicating the disposition of capacitive elements as components of the FIG. 11 pair of capacitors as determined by the digital words encoding weighting factor in one's complement arithmetic.

FIG. 13 is a schematic diagram showing in greater detail the electrical connections of one of the switched capacitive elements used in the FIG. 9 or FIG. 11 pair of capacitors.

FIG. 14 a view of a portion of the top surface of a monolithic integrated circuit in which reposes two of the switched capacitive elements and the single-bit storage elements for storing their respective control bits, which FIG. 14 has been labelled to indicate the location of circuit elements. This view can be analyzed in accordance with a normal procedure of designers of monolithic integrated circuits using stacked transparencies of different colors reproducing the masks of FIGS. 15, 16, 17, 18, 19, 20, 21, 22 and 23, respectively, all aligned by superposing their corresponding corner alignment keys. The masks of FIGS. 15, 16, 17, 18, 19, 20, 21, 22 and 23 are those used for constructing, in accordance with conventional complementary metal oxide semiconductor (CMOS) processing, each pair of capacitors that are adjustable in complementary way to a weighting word and for the word storage element to store that/Weighting word.

FIG. 15 is the mask defining the extent of the n-well region in a p-type silicon die, which n-well region underlies portions of a FIG. 14 double switched capacitor structure.

FIG. 16 is the mask defining the active area regions of both the n- and p-channel devices in a FIG. 14 double switched capacitor structure. These active area regions, covered by relatively thin gate oxide, are within rectangular boxes and include source, drain and channel regions. The region that surrounds each of these active area regions is a relatively thick field oxide area.

FIG. 17 is the mask defining the shape of the polycrystalline silicon conductors used in a FIG. 14 switched capacitor structure as gate electrodes in both the n- and p-channel field-effect transistors and for short-run conductors.

FIG. 18 is the mask defining the extent of the p+ implant used for establishing the source and drain regions of the p-channel field-effect transistors and for contacting to the p-type silicon substrate in a FIG. 14 double switched capacitor structure.

FIG. 19 is the mask defining the extent of the n+ contacts to n-well and of the n+ source and drain regions of the n-channel field-effect transistors in a FIG. 14 double switched capacitor structure.

FIG. 20 is the mask locating, in a FIG. 14 double switched capacitor structure, the position of contact openings between the first metallization layer and all of the following: the polycrystalline silicon conductors defined by the FIG. 17 mask, the p+ source and drain regions of the p-channel field-effect transistors, and the n+ source and drain regions of the n-channel field-effect transistors.

FIG. 21 is the mask defining the pattern of the first metallization layer in a FIG. 14 double switched capacitor structure.

FIG. 22 the mask locating the position of contact openings between the first and second metallization layers in a FIG. 14 double switched capacitor structure.

FIG. 23 is the mask defining the pattern of the second metallization layer in a FIG. 14 double switched capacitor structure.

FIG. 24 is a conceptual schematic diagram of a quad connection of four capacitors, the capacitances of which are programmable responsive to digital words encoding weighting factor in two's complement arithmetic.

FIG. 25 is a conceptual schematic diagram of a quad connection of four capacitors, the capacitances of which are programmable responsive to digital words encoding weighting factor in one's complement arithmetic.

FIG. 26 a schematic diagram of a neural net layer wherein weighting with higher bit resolution is provided without having to have as wide a range of weighting capacitor sizes.

Each of FIGS. 27 and 28 is conceptual schematic diagram of capacitor quads constructed in accordance with the invention to implement weighting with higher bit resolution in a neural net layer per FIG. 26.

DETAILED DESCRIPTION

FIG. 1 shows a neural net comprising a plurality, N in number, of non-linear amplifiers OD.sub.1, OD.sub.2, . . . OD.sub.(N-1), OD.sub.N. Each of a plurality, M in number, of input voltage signals x.sub.1, x.sub.2, . . . x.sub.(M-1), x.sub.M supplied as "synapse" signals is weighted to provide respective input voltages for the non-linear voltage amplifiers OD.sub.1, OD.sub.2, . . . OD.sub.(N-1 ), OD.sub.N, which generate respective "axon" responses y.sub.1, y.sub.2, . . . y.sub.(N-1), y.sub.N. This weighting is, as will be described in detail further on in this specification, done using capacitive structures in accordance with the invention.

M is a positive plural integer indicating the number of input synapse signals to the FIG. 1 neural net, and N is a positive plural integer indicating the number of output axon signals the FIG. 1 net can generate. To reduce the written material required to describe operation of the FIG. 1 neural net, operations using replicated elements will be described in general terms; using a subscript i ranging over all values one through M for describing operations and apparatuses as they relate to the (column) input signals x.sub.1, x.sub.2, . . . x.sub.(M-1), x.sub.M ; and using a subscript j ranging over all values one through N for describing operations and apparatus as they relate to the (row) output signals y.sub.1, y.sub.2, . . . y.sub.(N-1), y.sub.N. That is, i and j are the column and row numbers used to describe particular portions of the neural net.

Input voltage signal x.sub.i is applied to the input port of an input driver amplifier ID.sub.i that is a voltage amplifier which in turn applies its voltage response to an input line IL.sub.i. Respective output lines OL.sub.j and OL.sub.(j+N) connect to the non-inverting input port of output driver amplifier OD.sub.j and to its inverting input port. The non-linear output driver amplifier OD.sub.j is shown in FIG. 1 as simply being a differential-input non-linear voltage amplifier with the quiescent direct potential applied to its (+) and (-) input signal terminals via output lines OL.sub.j and OL.sub.(j+N) being adjusted by clamping to a desired bias voltage at selected times using a respective direct-current restorer circuit DCR.sub.j. (The direct-current restorer circuit DCR.sub.j is shown at the left of FIG. 1 for drafting reasons, but is normally associated with the input port of the output driver amplifier OD.sub.j.) Output driver amplifier OD.sub.j generates at its output port a non-linear voltage response to the cumulative difference in charge on that respective pair of output lines OL.sub.j and OL.sub.(j+N).

A respective capacitor C.sub.i,j connects each of the input lines IL.sub.i to each of the OL.sub.j output lines OL.sub.j, and a respective capacitor C.sub.i,(j+N) connects each of the input lines IL.sub.i to each of the output lines OL.sub.(j+N). Since at its output terminal the output driver amplifier OD.sub.j responds without inversion to x.sub.i input signal voltage applied to its non-inverting (+) input terminal via capacitor C.sub.i,j and responds with inversion to x.sub.i input signal voltage applied to its inverting (-) input terminal via capacitor C.sub.i,(j+N), respectively, the electrically equivalent circuit is x.sub.i signal voltage being applied to a single output line QL.sub.j by a capacitor having a capacitance that equals the capacitance of C.sub.i,j minus the capacitance of C.sub.i,(j+N). This technique of single-ended output signal drive to paired output lines that are differentially sensed avoids the need for switched-capacitance techniques in order to obtain inhibitory (or negative) weights as well as excitory (or positive) weights. Thus, this technique facilitates operating the neural net with analog signals that are continuous over sustained periods of time, if so desired.

FIG. 1 shows each of the input lines IL.sub.i as being provided with a respective load capacitor CL.sub.i to cause that capacitive loading upon the output port of the input driver amplifier ID.sub.i to be substantially the same as that upon each output port of the other input driver amplifiers. This is desirable for avoiding unwanted differential delay in responses to the input signals x.sub.i. Substantially equal capacitive loading can be achieved by making the capacitance of each of the input line loading capacitors, CL.sub.1 through CL.sub.M, very large compared to the total capacitance of the capacitors C.sub.i,j connecting thereto. Preferably, however, this result is achieved by making the capacitance of each of the input line loading capacitors complement the combined value of the other capacitances connecting thereto. This procedure reduces the amount of line loading capacitance required. Where the voltages appearing on the output lines OL.sub.j and OL.sub.(i+N) are sensed directly by the non-linear output driver amplifiers OD.sub.1, . . . OD.sub.N, as shown in FIG. 1, this procedure makes the voltage division ratio for each input voltage x.sub.i, . . . x.sub.m independent of the voltage division ratios for the other input voltages.

FIG. 1 also shows each of the output lines OL.sub.j being loaded with a respective Icad capacitor CL.sub.(M+j) and each of the output lines OL.sub.(N+j) being loaded with a respective load capacitor CL.sub.(M+N+j). This is done so that the total capacitance on each output line remains substantially the same as on each of the other output lines. This can be done by choosing CL.sub.(M+j) to be much larger than other capacitances to output line OL.sub.j, and by choosing CL.sub.(M+N+j) to be much larger than other capacitances to output line OL.sub.N+j). Alternatively, this can be done by choosing CL.sub.(M+j) and CL.sub.(M+N+j) to complement the combined value of the other capacitances connecting the same output line. The input voltage to output driver amplifier OD.sub.j will (to good approximation) have the following value, v.sub.j, in accordance with Coulomb's Law. ##EQU1## The generation of voltage v.sub.j can be viewed as the superposition of a plurality of capacitive divisions between, on the one hand, the effective capacitance (C.sub.(i,j) -C.sub.i,(j+N)) each input voltage has to output line OL.sub.j and, on the other hand, the total capacitance C.sub.j of the output line to its surroundings. That is, C.sub.j is the total capacitance on output line OL.sub.j or the total capacitance on output line OL.sub.(N+j), which capacitances should be equal to each other and fixed in value.

Each non-linear output driver amplifier OD.sub.j in the FIG. 1 neural net layer can be implemented using linear voltage amplifier circuitry followed by non-linear voltage amplifier circuitry. Each output driver amplifier can comprise a long-tailed pair connection of transistors having a current mirror amplifier load for converting their output signal voltage to single-ended form for application to an ensuing non-linear voltage amplifier. The long-tailed pair connection of transistors is a differential amplifier connection where their source electrodes have a differential-mode connection to each other and to a constant-current generator. An ensuing non-linear voltage amplifier can, as described in patent applications Ser. Nos. 366,838 and 366,839, comprise a cascade connection of two source-follower transistors, one an n-channel MOSFET and the other a p-channel MOSFET, each provided with a respective suitably-valued constant-current generator source load. Non-linearity of response in such a cascade connection comes about because (I) source-follower action of the n-channel MOSFET for positive-going excursions of its gate electrode potential becomes limited as its source potential approaches its drain potential V.sub.HI and (2) source-follower action of the p-channel MOSFET for negative-going excursions of its gate electrode potential becomes limited as its source potential approaches its drain potential V.sub.LO. At the source electrode of the output source-follower of the cascade connection, there is a sigmoidal response to a linear ramp potential applied to the gate electrode of the input source-follower of the cascade connection.

Alternatively, the difference in charge appearing on the output lines OL.sub.j and OL.sub.(i+N) can be sensed by fully differential charge-sensing amplifiers preceding the non-linear voltage amplifiers in the output driver amplifiers. In such case the output signals from the charge. sensing amplifiers will be balanced with reference to a reference VBIAS potential. This alternative will be described presently in connection with FIG. 2.

Consider now how neuron model behavior is exhibited by input driver amplifier ID.sub.i, capacitors C.sub.i,j and C.sub.i,(j+N), and non-linear output driver amplifier OD.sub.j for particular respective values of i and j. If the capacitance of capacitor C.sub.i,j is larger than the capacitance of capacitor C.sub.i,(j+N) for these particular values of i and j, then the output voltage y.sub.j for that j will exhibit "excitory" response to the input voltage x.sub.i. If the capacitances of C.sub.i,j and C.sub.i,(j+N) are equal for these i and j values, then the output voltage y.sub.j for that j should exhibit no response to the input voltage y.sub.j. If the capacitance of capacitor C.sub.i,j is smaller than the capacitance of capacitor C.sub.i(j+N) for those i and j values, then the output voltage y.sub.j for that j will exhibit "inhibitory" response to the input voltage x.sub.i.

In some neural nets using capacitors for weighting synapse signals, the capacitors C.sub.i,j and C.sub.i,(j+N) for all i and j may be fixed-value capacitors, so there is never any alteration in the weighting of input voltages x.sub.i where i=1, . . . M. However, such neural nets lack the capacity to adapt to changing criteria for neural responses--which adaptation is necessary, for example, in a neural network that is to be connected for self-learning. It is desirable in certain applications, then, to provide for altering the capacitances of each pair of capacitors C.sub.i,j and C.sub.i,(j+N) associated with a respective pair of values of i and j. This alteration is to be carried out in a complementary way, so the sum of the capacitances of C.sub.i,j and of C.sub.i,(j+N) remains equal to C.sub.k. In accordance with the invention, each of a set of component capacitors with capacitances related in accordance with powers of two is selected to be a component of one or the other of the pair of capacitors C.sub.i,j and C.sub.i,(j+N), the selecting being done by field effect transistors operated as transmission gates.

FIG. 2, comprising component FIGS. 2A and 2B, shows a representative modification that can be made to the FIG. 1 neural net near each set of intersections of output lines OL.sub.j and OL.sub.(j+N) with an input line IL.sub.i from which they receive with differential weighting a synapse input signal x.sub.i. Such modifications together make the neural net capable of being trained. Each capacitor pair C.sub.i,j and C.sub.i,(j+N) of the FIG. 1 neural net is to be provided by a pair of digital capacitors DC.sub.i,j and DC.sub.i,(j+N). The capacitances of DC.sub.i,j and DC.sub.i,(j+N) are controlled in complementary ways by a weighting factor and its one's complement as described by a digital word stored in a respective word-storage element WSE.sub.i,j of an array of such elements located interstitially among the rows of digital capacitors and connected to form a memory. This memory may, for example, be a random access memory (RAM) with each word-storage element WSE.sub.i,j being selectively addressable by row and column address lines controlled by address decoders. Or, by way of further example, this memory can be a plurality of static shift registers, one for each column j. Each static shift register will then have a respective stage WSE.sub.i,j for storing the word that controls the capacitances of each pair of digital capacitors DC.sub.i,j and DC.sub.i,(j+N).

The word stored in word storage element WSE.sub.i,j may also control the capacitances of a further pair of digital capacitors DC.sub.(i+M),j and DC.sub.(i+M),(j+N), respectively. The capacitors DC.sub.(i+M),j and DC.sub.(i+M),(j+N) connect between "ac ground" and output lines OL.sub.j and OL.sub.(j+N), respectively, and form parts of the loading capacitors CL.sub.(M+j). The capacitances of DC.sub.(i+2M,j) and DC.sub.i,j are similar to each other and changes in their respective values track each other. The capacitances of DC.sub.(i+M),(j+N) and DC.sub.i,(j+N) are similar to each other and changes in their respective values track each other. The four digital capacitors DC.sub.i,j, DC.sub.i,(j+N), DC.sub.(i+M),j and DC.sub.(ik+M),(j+N) are connected in a quad or bridge configuration having input terminals connecting from the input line IL.sub.i and from a-c ground respectively and having output terminals connecting to output lines OL.sub.j and OL.sub.(j+N) respectively. This configuration facilitates making computations associated with back-propagation programming by helping make the capacitance network bilateral insofar as voltage gain is concerned. Alternatively, where the computations for back-propagation programming are done by computers that do not involve the neural net in the computation procedures, the neural net need not include the digital capacitors DC.sub.(i+M),j and DC.sub.(i+M),(j+N).

When the FIG. 2 neural net is being operated normally, following programming, the .phi..sub.P signal applied to a mode control line MCL is a logic ZERO. This ZERO on mode control line MCL conditions each output line multiplexer OLM.sub.j of an N-numbered plurality thereof to select the output line OL.sub.j to the inverting input terminal of a respective associated fully differential amplifier DA.sub.j. This ZERO on mode control line MCL also conditions each output line multiplexer OLM.sub.(j+N) to select the output line OL.sub.(j+N) to the non-inverting input terminal of the respective associated fully differential amplifier DA.sub.j, which is included in a respective differential charge-sensing amplifier DQS.sub.j that performs a charge-sensing operation for output line OL.sub.i. A fully differential amplifier constructed of MOS field-effect transistors, as may serve for any one of the fully differential amplifiers DA.sub.j for j=1, 2, . . . N, is described on pages 255-257 of the book Analog MOS Integrated Circuits for Signal Processing by R. Gregorian and G. C. Temes, copyright 1986 , published by John Wiley & Sons, Inc., of New York, Chichester, Brisbane, Toronto and Singapore.

In furtherance of this charge-sensing operation, a transmission gate TG.sub.j responds to the absence of a reset pulse .phi..sub.R to connect an integrating capacitor CI.sub.j between the (+) output and (-) input terminals of amplifier DA.sub.j ; and a transmission gate TG.sub.(j+5N) responds to the absence of the reset pulse .phi..sub.R to connect an integrating capacitor CI.sub.(j+N) between the (-) output and (+)input terminals of amplifier DA.sub.j. With integrating capacitors CI.sub.j and CI.sub.(j+N) so connected, amplifier DA.sub.j functions as a differential charge amplifier. When .phi..sub.p signal on mode control line MCL is a ZERO, the input signal x.sub.i induces a total differential change in charge on the capacitors DC.sub.i,j and DC.sub.i,(j+N) proportional to the difference in their respective capacitances. The resulting displacement current flows needed to keep the input terminals of differential amplifier DA.sub.j substantially equal in potential requires that there be corresponding displacement current flow from the integrating capacitor CI.sub.j and CI.sub.(j+N) differentially charging those charging capacitors to place thereacross a differential voltage v.sub.j defined as follows. ##EQU2##

The half V.sub.j signal from the non-inverting (+) output terminal of amplifier DA.sub.j is supplied to a non-linear voltage amplifier circuit NL.sub.j which can be the non-linear voltage amplifier circuit using a cascade connection of p-channel and n-channel source-follower field effect transistors as previously described. The non-linear voltage amplifier circuit NL.sub.j responds to generate the axon output response y.sub.j. It is presumed that this non-linear voltage amplifier NL.sub.j supplies y.sub.j at a relatively low source impedance as compared to the input impedance offered by the circuit y.sub.j is to be supplied to--e.g. on an input line in a succeeding neural net layer. If this is so there is no need in a succeeding neural net layer to interpose an input driver amplifier ID.sub.i as shown in FIG. 1. This facilitates interconnections between successive neural net layers being bilateral. An output line multiplexer OLM.sub.j responds to the .phi..sub.P signal appearing on the mode control line MCL being ZERO to apply y.sub.j to an input line of a succeeding neural net layer if the elements shown in are in a hidden layer. If the elements shown in FIG. 2 are in the output neural net layer, output line multiplexer OLM.sub.j responds to the .phi..sub.P signal on the mode control line being ZERO to apply y.sub.j to an output terminal for the neural net.

From time to time, the normal operation of the neural net is interrupted; and, to implement dc-restoration a reset pulse .phi..sub.R is supplied to the differential charge sensing amplifier. DQS.sub.j. Responsive to .phi..sub.R, the logic complement of the reset pulse .phi..sub.R, going low when .phi..sub.R goes high, transmission gates TG.sub.j and TG.sub.(j+5N) are no longer rendered conductive to connect the integrating capacitors CI.sub.j and CI.sub.(j+N) from the output terminals of differential amplifier DA.sub.j. Instead, transmission gates TG.sub.(j+N) and TG.sub.(j+4N) respond to .phi..sub.R going high to connect to V.sub.BIAS the plates of capacitor CI.sub.j and CI.sub.(j+N) normally connected from those output terminals, V.sub.BIAS being the 2.5 volt intermediate potential between the V.sub.SS =0 volt and V.sub.DD =5 volt operating voltages of differential amplifier DA.sub.j. Other transmission gates TG.sub.(j+2N) and TG.sub.(j+3N) respond to .phi..sub.R going high to apply direct-coupled degenerative feedback from the output terminal of differential amplifier DA.sub.j to its input terminals, to bring the voltage at the output terminals to that supplied to its inverting input terminal from output lines OL.sub.j and OL.sub.(j+N). During the dc-restoration all x.sub.i are "zero-valued". So the charges on integrating capacitors CI.sub.j and CI.sub.(j+N) are adjusted to compensate for any differential direct voltage error occurring in the circuitry up to the output terminals of differential amplifier DA.sub.j. Dc-restoration is done concurrently for ail differential amplifiers DA.sub.j (i.e., for values of j ranging from one to N).

During training, the .phi..sub.P signal applied to mode control line MCL is a logic ONE, which causes the output line multiplexer OLM.sub.j to disconnect the output lines OL.sub.j and OL.sub.(j+N) from the (+) and (-) input terminals of differential amplifier DA.sub.j and to connect the output lines OL.sub.j and OL.sub.(j+N) to receive +.delta..sub.j and -.delta..sub.j error terms. These +.delta..sub.j and -.delta..sub.j error terms are generated as the balanced product output signal of a analog multiplier AM.sub.j, responsive to a signal .DELTA..sub.j and to a signal y'.sub.j which is the change in output voltage y.sub.j of non-linear amplifier NL.sub.j for unit change in the voltage on output line OL.sub.j. The term .DELTA..sub.j for the output neural net layer is an error signal that is the difference between y.sub.j actual value and its desired value d.sub.j. The term .DELTA..sub.j for a hidden neural net layer is also an error signal, which is of a nature that will be explained in detail further on in this specification.

Differentiator DF.sub.j generates the signal y'.sub.j, which is a derivative indicative of the slope of y.sub.j change in voltage on output line OL.sub.j, superposed on V.sub.BIAS. To determine the y'.sub.j derivative, a pulse doublet comprising a small positive-going pulse immediately followed by a similar-amplitude negative-going pulse is introduced at the inverting input terminal of differential amplifier DA.sub.j (or equivalently, the opposite-polarity doublet pulse is introduced at the non-inverting input terminal of differential amplifier DA.sub.j) to first lower y.sub.j slightly below normal value and then raise it slightly above normal value. This transition of y.sub.j from slightly below normal value to slightly above normal value is applied via a differentiating capacitor CD.sub.j to differentiator DF.sub.j.

Differentiator DF.sub.j includes a charge sensing amplifier including a differential amplifier DA.sub.(j+N) and an integrating capacitor CI.sub.(j+2N). During the time y.sub.j that is slightly below normal value, a reset pulse .phi..sub.S is applied to transmission gates TG.sub.(j+7N) and TG.sub.(j+8N) to render them conductive. This is done to drain charge from integrating capacitor CI.sub.(j+N), except for that charge needed to compensate for DA.sub.(j+N) input offset voltage error. The reset pulse .phi..sub.S ends, rendering transmission gates TG.sub.(j+7N) and TG.sub.(j+5N) no longer conductive, and the complementary signal .phi..sub.S goes high to render a transmission gate TG.sub.(j+6N) conductive for connecting integrating capacitor CI.sub.(j+2N) between the output and inverting-input terminals of differential amplifier DA.sub.(j+N).

With the charge-sensing amplifier comprising elements DA.sub.(j+N) and CI.sub.(j+N) reset, the small downward pulsing of y.sub.j from normal value is discontinued and the small upward pulsing of y.sub.j from normal value occurs. The transition between the two abnormal conditions of y.sub.j is applied to the charge-sensing amplifier by electrostatic induction via differentiating capacitor CD.sub.j. Differential amplifier DA.sub.(j+N) output voltage changes by an amount y'.sub.j from the V.sub.BIAS value it assumed during reset. The use of the transition between the two pulses of the doublet, rather than the edge of a singlet pulse, to determine the derivative y'.sub.j makes the derivative-taking process treat more similarly those excitory and inhibiting responses of the same amplitude. The doublet pulse introduces no direct potential offset error into the neural net layer.

Responsive to a pulse .phi..sub.T, the value y'.sub.j +V.sub.BIAS from differentiator DF.sub.j is sampled and held by row sample and hold circuit SH.sub.j for application to analog multiplier AM.sub.j as an input signal. This sample and hold procedure allows y.sub.j to return to its normal value, which is useful in the output layer to facilitate providing y.sub.j for calculating (y.sub.j -d.sub.j). The sample and hold circuit SH.sub.j may simply comprise an L-section with a series-arm transmission-gate sample switch and a shunt-leg hold capacitor, for example. The analog multiplier AM.sub.j accepts a first push-pull input signal between input terminals IN1 and IN1, accepts a second push-pull input signal between terminals IN2 and IN2, and supplies product output signal in balanced form at its output terminals POUT and POUT. The difference between y'.sub.j +V.sub.BIAS and V.sub.BIAS voltages is applied as a differential input signal to the analog multiplier AM.sub.j, which exhibits common-mode rejection for the V.sub.BIAS term. In U.S. patent applications Ser. Nos. 366,838 and 366,839 respectively entitled "NEURAL NET USING CAPACITIVE STRUCTURES CONNECTING INPUT LINES AND DIFFERENTIALLY SENSED OUTPUT LINE PAIRS" and "NEURAL NET USING CAPACITIVE STRUCTURES CONNECTING OUTPUT LINES AND DIFFERENTIALLY DRIVEN INPUT LINE PAIRS", the four-quadrant analog multiplier AM.sub.j is described as being a push-pull-output analog multiplier formed by modifying a single-ended-output analog multiplier described by K. Bultt and H. Wallinga in their paper "A CMOS Four-quadrant Analog Multiplier" appearing on pages 430-435 of the IEEE JOURNAL OF SOLID STATE CIRCUITS, Vol. SC-21, No. 3, June 1986, incorporated herein by reference.

During training, the .phi..sub.P signal applied to mode control line MCL is a logic ONE, as previously noted. When the FIG. 2 elements are in the output layer, the ONE on mode control line MCL conditions an output multiplexer OM.sub.j to discontinue the application of y.sub.j signal from non-linear amplifier NL.sub.j to an output terminal. Instead, the output multiplexer OM.sub.j connects the output terminal to a charge-sensing amplifier QS.sub.j. Charge sensing amplifier QS.sub.j includes a differential amplifier DA.sub.(j+2N) and an integrating capacitor CI.sub.(j+2N) and is periodically reset responsive to a reset pulse .phi..sub.U. Reset pulse .phi..sub.U can occur simultaneously with reset pulse .phi..sub.S, for example. Output signal .DELTA..sub.j from charge-sensing amplifier QS.sub.j is not used in the output layer, however. Analog multiplier AM.sub.j does not use .DELTA..sub.j +V.sub.BIAS and V.sub.BIAS as a differential input signal in the output layer, (y.sub.j -d.sub.j) being used instead.

When the FIG. 2 elements are in a hidden neural net layer, .phi..sub.P signal on the mode control line MCL being a ONE conditions output multiplexer OM.sub.j to discontinue the application of y.sub.j signal from non-linear amplifier NL.sub.j to the input line IL.sub.j of the next neural net layer. Instead, output multiplexer OM.sub.j connects the input line IL.sub.j to charge-sensing amplifier QS.sub.j. Charge-sensing amplifier QS.sub.j senses change in the charge on input line IL.sub.j during training to develop a .DELTA..sub.j error signal superposed on V.sub.BIAS direct potential. The difference between .DELTA..sub.j +V.sub.BIAS and V.sub.BIAS voltages is used as a differential input signal to analog multiplier AM.sub.j, which multiplier exhibits common-mode rejection for the V.sub.BIAS term.

Charge-sensing amplifier QS.sub.j employs differential-input amplifier DA.sub.(j+2N) and integrating capacitor CI.sub.(j+2N). Transmission gates TG.sub.(i+gN.sub.), TG.sub.(j+10N) and TG.sub.(j+11N) cooperate to provide occasional resetting of charge conditions on the integrating capacitor CI.sub.j+2N responsive to the reset pulse .phi..sub.U.

FIG. 3 shows a neural net in which the input driver amplifier ID.sub.i applies, in response to input voltage signal x.sub.i, not only a non-inverted voltage response from its (+) output port to an input line ILi, but also an inverted voltage response from its (-) output port to an input line IL.sub.(i+M). A respective degenerative feedback connection from its (+) output terminal to its (-) input terminal conditions each of the input driver amplifiers ID.sub.i in the FIG. 3 neural net to provide x.sub.i voltage-follower response at its (+) output terminal to x.sub.i signal applied to its (+) input terminal and to provide inverted, -x.sub.i response at its (-) output terminal. A respective single output line OL.sub.j connects to the input port of output driver amplifier OD.sub.j ,which generates at its output port a non-linear voltage response to the cumulative charge on that respective output line OL.sub.j.

The non-linear output driver amplifier OD.sub.j is shown in FIG. 3 as being just a non-linear voltage amplifier with the quiescent direct potential applied to its input signal terminal via output line OL.sub.j being adjusted by clamping to a desired bias voltage at selected times using a respective direct-current restorer circuit DCR.sub.j. A respective capacitor C.sub.i,j connects each of the input lines IL.sub.i to each of the output lines OL.sub.j, and a respective capacitor C.sub.(i+M),j connects to each of the output lines OL.sub.j the one of the input lines IL.sub.(i+M) paired with that IL.sub.i. Since the paired IL.sub.i and IL.sub.(i+M) input lines are driven with x.sub.i and -x.sub.i signal voltages respectively, the electrically equivalent circuit is x.sub.i signal voltage being applied to output line OL.sub.j by a capacitor having a capacitance that equals the capacitance of C.sub.i,j minus the capacitance of C.sub.(i+M),j. This balanced input signal drive to paired input lines technique avoids the need for switched-capacitance techniques in order to obtain inhibitory as well as excitory weights, and thus facilitates operating the neural net with analog signals that are continuous over sustained periods of time, if so desired.

FIG. 3 shows each of the input lines IL.sub.i or IL.sub.(i+M) as being provided with a respective load capacitor CL.sub.i or CL.sub.(i+M) to cause that capacitive loading upon each of the output terminals of the input driver amplifier ID.sub.i to be substantially the same as that upon each output port of the other input driver amplifiers. This is desirable for avoiding unwanted differential delay in responses to the input signals x.sub.i. Substantially equal capacitive loading can be achieved by making the capacitance of each of the input line loading capacitors CL.sub.1 -CL.sub.2M very large compared to the total capacitance of the capacitors C.sub.i,j or C.sub.(i+M),j connecting thereto. Preferably, however, this result is achieved by making the capacitance of each of the input line loading capacitors complement the combined value of the other capacitances connecting thereto. This procedure reduces the amount of line loading capacitance required. Where the voltage appearing on the output lines is sensed directly by the non-linear output driver amplifiers OD.sub.1, . . . OD.sub.N, as shown in FIG. 3, this preferable procedure makes the voltage division ratio for each input voltage x.sub.i, . . . x.sub.M independent of the voltage division ratios for the other input voltages. Where the charge appearing on the output lines is sensed by charge-sensing amplifiers preceding the non-linear output driver amplifiers, as will be described later on in this specification in connection with FIG. 4, this latter consideration is not as important.

FIG. 3 also shows each of the output lines OL.sub.j being loaded with a respective load capacitor CL.sub.(2M+j) to cause the total capacitance on that line to remain substantially the same as on each of the other output lines. Again, this can be done either by choosing CL.sub.(2M+j) to be much larger than other capacitances to output line OL.sub.j, or by choosing CL.sub.(2M+j) to complement the combined value of the other capacitances connecting thereto. The input voltage to output driver amplifier OD.sub.i will (to good approximation) have the following value, v.sub.j, in accordance with Coulomb's Law. ##EQU3## Here C.sub.j is the total capacitance on output line OL.sub.j. The generation of voltage v.sub.j can be viewed as the superposition of a plurality of capacitive divisions between, on the one hand, the effective capacitance (C.sub.(i,j) -C.sub.(i+M),j) each input voltage has to output line OL.sub.j and, on the other hand, the total capacitance C.sub.j of the output line to its surroundings.

Consider now how neuron model behavior is exhibited by input driver amplifier ID.sub.i, capacitors C.sub.i,j and C.sub.(i+M),j, and non-linear output driver amplifier OD.sub.j for particular respective values of i and j. The voltage responses input driver amplifier ID.sub.i applies to input lines IL.sub.i and IL.sub.(i+M) are the same in amplitude but are of opposing polarity as referred to a common-mode voltage that is designed to be nominally the same as a bias voltage V.sub.BIAS midway between the 0-volt V.sub.SS and +5-volt V.sub.DD supply voltages. If the capacitance of capacitor C.sub.i,j is larger than the capacitance of capacitor C.sub.(i+M),j for these particular values of i and j, then the output voltage y.sub.j for that j will exhibit "excitory" response to the input voltage x.sub.i. If the capacitances of C.sub.i,j and C.sub.(i+M),j are equal for these i and j values, then the output voltage y.sub.j for that j should exhibit no response to the input voltage y.sub.j. If the capacitance of capacitor C.sub.i,j is smaller than the capacitance of capacitor C.sub.(i+M),j for those i and j values, then the output voltage y.sub.j for that j will exhibit "inhibitory" response to the input voltage x.sub.i.

In some neural nets constructed in accordance with the invention the capacitors C.sub.i,j and C.sub.(i+M)j for all i and j may be fixed-value capacitors, so there is never any alteration in the weighting of input voltages x.sub.i where i=1, . . . M. However, such neural nets lack the capacity to adapt to changing criteria for neural responses--which adaptation is necessary, for example, in a neural network that is to be connected for self-learning. It is desirable in certain applications, then, to provide for altering the capacitances of each pair of capacitors C.sub.i,j and C.sub.(i+M),j associated with a respective pair of values of i and j. This alteration is to be carried out in a complementary way, so the sum of the capacitances of C.sub.i,j and of C.sub.(i+M),j remains equal to C.sub.k. In accordance with the invention, each of a set of component capacitors with capacitances related in accordance with powers of two is selected to be a component of one or the other of the pair of capacitors C.sub.i,j and C.sub.(i+M),j, the selecting being done by field effect transistors operated as transmission gates.

FIG. 4, comprising component FIGS. 4A and 4B, shows a representative modification that can be made to the FIG. 3 neural net near each set of intersections of an output line OL.sub.j with input lines IL.sub.i and IL.sub.(i+M) driven by opposite senses of a synapse input signal x.sub.i. Such modifications together make the neural net capable of being trained. Each capacitor pair C.sub.i,j and C.sub.(i+M),j of the FIG. 3 neural net is to be provided by a pair of digital capacitors DC.sub.i,j and DC.sub.(i+M),j. The capacitances of DC.sub.i,j and DC.sub.(i+M),j are controlled in complementary ways by a digital word, as drawn from a respective word-storage element WSE.sub.i,j in an array of such elements located interstitially among the rows of digital capacitors and connected to form a memory. This memory may, for example, be a random access memory (RAM) with each word-storage element WSE.sub.i,j being selectively addressable by row and column address lines controlled by address decoders. Or, by way of further example, this memory can be a plurality of static shift registers, one for each column j. Each static shift register will then have a respective stage WSE.sub.i,j for storing the word that controls the capacitances of each pair of digital capacitors DC.sub.i,j and DC.sub.(i+M),j.

The word stored in word storage element WSE.sub.i,j may also control the capacitances of a further pair of digital capacitors DC.sub.i,(j+N) and DC.sub.(i+M),(j+N), respectively. The capacitors DC.sub.i,(j+N) and DC.sub.(i+M),(j+N) connect between "ac ground" and input lines IL.sub.i and IL.sub.(i+M), respectively, and form parts of the loading capacitors CL.sub.i and CL.sub.(i+M), respectively. The capacitances of DC.sub.(i+M,(j+N) and DC.sub.i,j are similar to each other and changes in their respective values track each other. The four digital capacitors DC.sub.i,j, DC.sub.(i+M),j, DC.sub.i,(j+N) and DC.sub.(i+M),(j+N) are connected in a quad or bridge configuration having input terminals to which the input lines Il.sub.i and IL.sub.(i+M) respectively connect and having output terminals connecting to output line QL.sub.j and to ac ground respectively. The capacitances of DC.sub.i,(j+N) and DC.sub.(i+M),j are similar to each other and changes in their respective values track each other. This configuration facilitates making computations associated with back-propagation programming by helping make the capacitance network bilateral insofar as voltage gain is concerned. Alternatively, where the computations for back-propagation programming are done by computers that do not involve the neural net in the computation procedures, the neural net need not include the digital capacitors DC.sub.i,j +N and DC.sub.(i+M),(j+N). These digital capacitors DC.sub.i,(j+N) and DC.sub.(I+M),(j+N) are not needed either where very large loading capacitors are placed on the output lines OL.sub.j, but this alternative undesirably reduces sensitivity of the output driver amplifier OD.sub.j.

When the FIG. 4 neural net is being operated normally, following programming, the .phi..sub.P signal applied to a mode control line MCL.is a logic ZERO. This ZERO conditions a respective input line multiplexer ILMi to connect the non-inverting output port at each input driver amplifier ID.sub.i to input line IL.sub.i. The .phi..sub.P signal on mode control line MCL being a ZERO also conditions a respective input line multiplexer ILM.sub.(i+M) to connect the inverting output port of each input driver amplifier ID.sub.i to input line IL.sub.(i+M).

A ZERO on mode control line MCL also conditions each output line multiplexer OLM.sub.j of an n-numbered plurality thereof to select the output line OL.sub.j to the inverting input terminal of a respective associated differential-input amplifier DA.sub.j, included in a respective charge-sensing amplifier QS.sub.j that performs a charge-sensing operation for output line OL.sub.j. In furtherance of this charge-sensing operation, a transmission gate TG.sub.j responds to the absence of a reset pulse Q.sub.R to connect an integrating capacitor CI.sub.j between the output and inverting-input terminals of differential-input amplifier DA.sub.j. Amplifier DA.sub.j may be an operational amplifier of the conventional voltage amplifier type or may be an operational transconductance amplifier. With integrating capacitor CI.sub.j so connected, amplifier DA.sub.j functions as a charge amplifier. When .phi..sub.P signal on mode control line MCL is a ZERO, the input signal x.sub.i induces a total change in charge on the capacitors DC.sub.i,j and DC.sub.(i+M),j proportional to the difference in their respective capacitances. The resulting displacement current flow from the inverting input terminal of differential-input amplifier DA.sub.j requires that there be a corresponding displacement current flow from the integrating capacitor CI.sub.j charging that capacitor to place thereon a voltage v.sub.j defined as follows. ##EQU4##

The voltage V.sub.j is supplied to a non-linear voltage amplifier circuit NL.sub.j, which non-linear voltage amplifier circuit responds to generate the axon output response y.sub.j.

From time to time, the normal operation of the neural net is interrupted, and to implement dc-restoration a reset pulse .phi..sub.R is supplied to each charge sensing amplifier QS.sub.j. Responsive to .phi..sub.R the logic complement of the reset pulse .phi..sub.R, going low when .phi..sub.R goes high, transmission gate TG.sub.j is no longer rendered conductive to connect the integrating capacitor CI.sub.j from the output terminal of differential amplifier DA.sub.j. Instead, a transmission gate TG.sub.(j+N) responds to .phi..sub.R going high to connect to V.sub.BIAS the plate of capacitor CJ.sub.j normally connected from that output terminal, V.sub.BIAS being the 2.5 volt intermediate potential between the V.sub.SS =0 volt and V.sub.DD =5 volt operating voltages of differential amplifier DA.sub.j. Another transmission gate TG.sub.(j+2N) responds to .phi..sub.R going high to apply direct-coupled feedback from the output terminal of differential amplifier DA.sub.j to its inverting input terminal, to bring the voltage at the output terminal to that supplied to its inverting input terminal from output line OL.sub.j. During the dc-restoration all x.sub.i are "zero-valued". So the charge on integrating capacitor CI.sub.j is adjusted to compensate for any direct voltage error occurring in the circuitry up to the output terminal of differential amplifier DA.sub.j. DC-restoration is done concurrently for all differential amplifiers DA.sub.j (i.e., for values of j ranging from one to N).

During training, the .phi..sub.P signal applied to mode control line MCL is a logic ONE, which causes the output line multiplexer OLM.sub.j to disconnect the output line OL.sub.j from the inverting input terminal of differential amplifier DA.sub.j and to connect the output line OL.sub.j to receive a .delta..sub.j error term. This .delta..sub.j error term is generated as the product output signal of an analog multiplier AM.sub.j, responsive to a signal .DELTA..sub.j and to a signal y'.sub.j which is the change in output voltage y.sub.j of non-linear amplifier NL.sub.j for unit change in the voltage on output line OL.sub.j. The term .DELTA..sub.j is for the output neural net layer the difference between y.sub.j actual value and its desired value d.sub.j. The term .DELTA..sub.j is for a hidden neural net layer the .DELTA..sub.j output of the succeeding neural net layer during the back-propagation procedure.

Differentiator DF.sub.j generates the signal y'.sub.j, which is a derivative indicative of the slope of y.sub.j change in voltage on output line OL.sub.j, superposed on V.sub.BIAS. To determine the y'.sub.j derivative, a pulse doublet comprising a small positive-going pulse immediately followed by a similar-amplitude negative-going pulse is introduced at the inverting input terminal of differential amplifier DA.sub.j (or equivalently, the opposite-polarity doublet pulse is introduced at the non-inverting input terminal of differential amplifier DA.sub.j) to first lower y.sub.j slightly below normal value and then raise it slightly above normal value. This transition of y.sub.j from slightly below normal value to slightly above normal value is applied via a differentiating capacitor CD.sub.j to differentiator DF.sub.j.

Differentiator DF.sub.j includes a charge sensing amplifier including a differential amplifier DA.sub.(j+N) and an integrating capacitor CI.sub.(j+N). During the time that y.sub.j is slightly below normal value, a reset pulse .phi..sub.S is applied to transmission gates TG.sub.(j+4N) and TG.sub.(j+5N) to render them conductive. This is done to drain charge from integrating capacitor CI.sub.(j+N), except for that charge needed to compensate for DA.sub.(j+N) input offset voltage error. The reset pulse .phi..sub.S ends, rendering transmission gates TG.sub.(j+4N) and TG.sub.(j+5N) no longer conductive, and the complementary signal .phi..sub.S goes high to render a transmission gate TG.sub.(j+3N) conductive for connecting integrating capacitor CI.sub.(i+N) between the output and inverting-input terminals of differential amplifier DA.sub.(j+N).

With the charge sensing amplifier comprising elements DA.sub.(j+N) and CI.sub.(j+N) reset, the small downward pulsing of y.sub.j from normal value is discontinued and the small upward pulsing of y.sub.j from normal value occurs. The transition between the two abnormal conditions of y.sub.j is applied to the charge sensing amplifier by electrostatic induction via differentiating capacitor CD.sub.j. Differential amplifier DA.sub.(i+N) output voltage changes by an amount y'.sub.j from the V.sub.BIAS value it assumed during reset. The use of the transition between the two pulses of the doublet, rather than the edge of a singlet pulse, to determine the derivative y'.sub.j makes the derivative-taking process treat more similarly those excitory and inhibiting responses of the same amplitude. The doublet pulse introduces no direct potential offset error into the neural net layer.

Responsive to a pulse .phi..sub.T, the value y'.sub.j +V.sub.BIAS from differentiator DF.sub.j is sampled and held by (row) sample and hold circuit RSH.sub.j for application to the analog multiplier AM.sub.j as an input signal. This sample and hold procedure allows y.sub.j to return to its normal value, which is useful in the output layer to facilitate providing y.sub.j for calculating (y.sub.j -d.sub.j). The sample and hold circuit RSH.sub.j may simply comprise an L-section with a series-arm transmission-gate sample switch and a shunt-leg hold capacitor, for example. The difference between y.sub.j +V.sub.BIAS and V.sub.BIAS voltages is used as a differential input signal to the analog multiplier AM.sub.j, which exhibits common-mode rejection for the V.sub.BIAS term.

During training, the .phi..sub.P signal applied to the mode control line MCL is a ONE, as previously noted, and this causes the input line multiplexers ILM.sub.i and ILM.sub.(i+M) to disconnect the input lines IL.sub.i and IL.sub.(i+M) from the input driver amplifier ID.sub.i output terminals and connect them instead to the non-inverting and inverting input terminals of a differential charge-sensing amplifier BDQSi. The voltage .delta..sub.j induces a differential change in charge between input lines IL.sub.j and IL.sub.(i+M) proportional to .delta..sub.j (C.sub.i,j -C.sub.(i+M),j), which differential change in charge is sensed using the differential charge sensing amplifier BDQSi.

Differential charge-sensing amplifier BDQSi includes a fully differential amplifier provided with integrating capacitors IC.sub.i and IC.sub.(i+M) in respective degenerative feedback connections from each of its output terminals to each of its input terminals. Resetting of differential charge-sensing amplifier BDQSi is similar to the resetting of a single-ended amplifier such as QS.sub.j, except for involving two integrating capacitors IC.sub.i and IC.sub.(i+M), rather than just the one integrating capacitor C.sub.ij. Resetting of differential charge-sensing amplifier BDQSi is done responsive to a pulse .phi..sub.U, which occurs during the time when mode control line MCL has a ONE thereon conditioning input line multiplexers ILM.sub.i and IM.sub.(i+M) to connect input lines IL.sub.i and IL.sub.(i+M) to the differential charge-sensing amplifier BDQSi. Resetting is normally done shortly after a ZERO to ONE transition appears in the .phi..sub.P signal applied to mode control line MCL and may also be done at other times. This procedure corrects for capacitive unbalances on the input lines IL.sub.i and IL.sub.(i+M) during back-propagation computations that follow the resetting procedure. In these computations voltages +.DELTA..sub.i +V.sub.BIAS and -.DELTA..sub.i V.sub.BIAS are developed at the (+) and (-) output terminals of the fully differential amplifier included in differential charge-sensing amplifier BDQSi. The voltage +.DELTA..sub.i +V.sub.BIAS is used by the preceding neural net layer during the back-propagation training procedure, if such a preceding neural net layer exists. The use of single-ended +.DELTA..sub.i and +.DELTA..sub.j drive is shown in FIG. 4A, presuming the neural net layers to be integrated within separate monolithic integrated circuits, and presuming the limitation on number of pin-outs is restrictive. Where a plurality of neural net layers are integrated within the same monolithic integrated circuitry, or where maximum pin-out count is not a restrictive design factor, balanced .DELTA. signals may be applied from one neural net layer to the preceding one. So, too, if the non-linear voltage amplifier NL.sub.j is of a correct type (for example, a long-tailed pair connection of transistors) y.sub.j +V.sub.BIAS and -y.sub.j +V.sub.BIAS balanced output signals may be supplied to the next neural net layer. Indeed, the y'.sub.j signals applied to the analog multiplier AM.sub.j may be generated in balanced form, replacing differentiator DF.sub.j and sample-and-hold circuit SHj with balanced circuitry.

FIG. 5 shows how in either FIG. 2 or 4 each output line OL.sub.j for j=1, . . . N may be pulsed during calculation of y'.sub.j terms. Each output line OL.sub.j is connected by a respective capacitor CO.sub.j to the output terminal of a pulse generator PG, which generates the doublet pulse. FIG. 5 shows the doublet pulse applied to the end of each output line QL.sub.j remote from the--terminal of the associated differential amplifier DA.sub.j in the charge-sensing amplifier QS.sub.j sensing the charge on that line. It is also possible to apply the doublet pulses more directly to those--terminals by connecting to these terminals respective ones of the plates of capacitors CO.sub.j that are remote from the plates connecting to pulse generator PG.

Each output line OL.sub.j has a respective capacitor CO.sub.j connected between it and a point of reference potential, and each output line OL.sub.(j+N) has a respective capacitor CO.sub.(j+N) connected between it and the point of reference potential, which capacitors are not shown in the drawing. The respective capacitances of the capacitors CO.sub.j and CO.sub.(j+N) are all of the same value, so that the back-propagation algorithm is not affected by the presence of these capacitors. Arrangements for adding the doublet pulse to v.sub.j before its application to the non-linear amplifier NL.sub.j can be used, rather than using the FIG. 5 arrangement.

FIG. 6 comprising component FIGS. 6A and 6B shows further modification that can be made to the FIG. 4 modification for the FIG. 3 neural net. This modification, as shown in FIG. 6A provides for a pair of input lines IL.sub.i and IL.sub.(i+M) for driving each quad configuration of digital capacitors DC.sub.i,j, DC.sub.i,(j+N), DC.sub.(i+M),j and DC.sub.(i+M),(j+N) push-pull rather than single-ended. Push-pull, rather than single-ended drive is provided to the differential charge sensing amplifier DQSj, doubling its output response voltage. Push-pull drive also permits differential charge sensing amplifier DQSj to be realized with differential-input amplifiers that do not provide for common-mode suppression of their output signals, if one so desires.

FIG. 6B differs from FIG. 4B in that the single-ended charge-sensing amplifier. QS.sub.j does not appear, being inappropriate for sensing differences in charge appearing on a pair of input lines. Instead, .DELTA..sub.j +B.sub.BIAS is developed in the following neural net layer and is fed back to analog multiplier AM.sub.j via the output multiplexer OM.sub.j when the .phi..sub.P signal on mode control line MCL is a

FIG. 4A shows circuitry that may be used in each neural net layer to provide balanced input signal drive to a pair of input lines IL.sub.i and IL.sub.(i+M) during normal operation and to differentially sense the charge on those input lines during back-propagation calculations. Both function may be implemented with separate apparatus.

During normal operation the .phi..sub.P Signal appearing on mode control line MCL is a ZERO, conditioning an input multiplexer IM.sub.i (designated output multiplexer OM.sub.j of the previous layer) to apply x.sub.i signal to the non-inverting (+) input terminal of differential amplifier ID.sub.i and conditioning input line multiplexers ILM.sub.i and ILM.sub.(i+M) to connect the non-inverting (+) and inverting (-) output terminals of differential amplifier ID.sub.i to input lines IL.sub.i and IL.sub.(i+M) respectively. A signal .phi..sub.P is a ONE during normal operation and appears in the .phi..sub.U +.phi..sub.P control signal applied to a transmission gate between the non-inverting (+) output terminal of differential amplifier ID.sub.i and its inverting (-) input terminal, rendering that transmission gate conductive to provide direct-coupled feedback between those terminals. This d-c feedback conditions differential amplifier ID.sub.i to provide x.sub.i and -x.sub.i responses at its (+) and (-) output terminals to the x.sub.i signal applied to its (-) input terminal. Other transmission gates within the duplex circuitry DPX.sub.i are conditioned to be non-conductive during normal operation.

During back-propagation calculations, the .phi..sub.P signal appearing on mode control line MCL is a ONE, conditioning multiplexer OM.sub.j (shown in FIG. 6B) to apply .DELTA..sub.i signal from the non-inverting (+) output terminal of differential amplifier ID.sub.i to the preceding neural net layer, if any, and conditioning input line multiplexers ILM.sub.i and ILM.sub.(i+M) to connect the input lines IL.sub.i and IL.sub.(i+M) to respective ones of the non-inverting (+) and inverting (-) input terminals of differential amplifier BDQS.sub.j. Integrating capacitors IC.sub.i and IC.sub.(i+M) connect from the (+) and (-) output terminals of differential amplifier BDQS.sub.j to its (-) and (+) input terminals when transmission gates in duplex circuitry DPX.sub.i that are controlled by .phi..sub.U .multidot..phi..sub.P signal receive a ZERO during back-propagation calculations. The charge conditions on integrating capacitors IC.sub.i and IC.sub.(i+M) are reset when .phi..sub.U occasionally pulses to ONE during back-propagation calculations. This happens in response to transmission gates in duplexer circuitry DPX.sub.i receptive of .phi..sub.U and .phi..sub.U +.phi..sub.P control signals being rendered conductive responsive to .phi..sub.U being momentarily a ONE, while transmission gates in duplexer circuitry DPX.sub.i receptive of .phi..sub.U control signal being rendered non-conductive.

A column sign detector (not shown) receives output signal from differential amplifier ID.sub.i directly as its input signal and can simply be a voltage comparator and sample and hold for the x.sub.i and -x.sub.i output signals from the differential amplifier ID.sub.i. This signal is applied to line CSL.sub.i shown in FIG. 6A.

FIG. 7 shows apparatuses for completing the back-propagation computations, as may be used with the FIG. 1 neural net manifoldly modified per FIG. 2, with the FIG. 3 neural net manifoldly modified per FIG. 4, or with the FIG. 1 neural net manifoldly modified per FIG. 6. The weights at each word storage element WSE.sub.i,j in the interstitial memory array IMA are to be adjusted as the column addresses and j row addresses are scanned row by row, one column at a time. An address scanning generator ASG generates this scan of i and j addresses shown applied to interstitial memory array IMA, assuming it to be a random access memory. The row address j is applied to a row multiplexer RM that selects .delta..sub.j to one input of a multiplier MULT, and the column address i is applied to a column multiplexer CM that selects x.sub.i to another input of the multiplier MULT.

Multiplier MULT is of a type providing a digital output responsive to the product of its analog input signals. Multiplier MULT may be a multiplying analog-to-digital converter, or it may comprise an analog multiplier followed by an analog-to-digital converter, or it may comprise an analog-to-digital converter for each of its input signals and a digital multiplier for multiplying together the converted signals. Multiplier MULT generates the product x.sub.i .delta..sub.j as reduced by a scaling factor .eta., which is the increment or decrement to the weight stored in the currently addressed word storage element WSE.sub.ij in the memory array IMA. The former value of weight stored in word storage element WSE.sub.ij is read from memory array IMA to a temporary storage element, or latch, TS. This former weight value is supplied as minuend to a digital subtractor SUB, which receives as subtrahend .eta. x.sub.i .delta..sub.j from multiplier MULT. The resulting difference is the updated weight value which is written into word storage element WSE.sub.i,j in memory array IMA to replace the former weight value.

FIG. 8 shows how trained neural net layers L.sub.0, L.sub.1 and L.sub.2 are connected together in a system that can be trained. L.sub.0 is the output neural net layer that generates y.sub.j output signals; is similar to that described in connection with FIGS. 1 and 2, in connection with FIGS. 3 and 4, or in connection with FIG. 6; and is provided with a back-propagation processor BPP.sub.0 with elements similar to those shown in FIG. 2, 4 or 6 for updating the weights stored in the interstitial memory array of L.sub.0. L.sub.1 is the first hidden neural net layer which generates y.sub.i output signals supplied to the output neural net layer as its x.sub.i input signals. These y.sub.i output signals are generated by layer L.sub.1 as its non-linear response to the weighted sum of its x.sub.h input signals. This first hidden neural net layer L.sub.1 is provided with a back-propagation processor BPP.sub.1 similar to BPP.sub.0. L.sub.2 is the second hidden neural net layer, which generates y.sub.h output signals supplied to the first hidden neural net layer as its x.sub.h input signals. These y.sub.h output signals are generated by layer L.sub.2 as its non-linear response to a weighted summation of its x.sub.g input signals. This second hidden layer is provided with a back-propagation processor similar to BPP.sub.0 and to BPP.sub.1.

FIG. 8 presumes that the respective interstitial memory array IMA of each neural net layer L.sub.0, L.sub.1, L.sub.2 has a combined read/write bus instead of separate read input and write output busses as shown in FIG. 2, 4 or 6. FIG. 8 shows the .delta..sub.j, .DELTA..sub.i and .delta..sub.h signals being fed back over paths separate from the feed forward paths for y.sub.j, y.sub.i and y.sub.h signals, which separate paths are shown to simplify conceptualization of the neural net by the reader. In actuality, as shown in FIG. 2, 4 or 6, a single path may be used to transmit y.sub.j in the forward direction and Aj in the reverse direction, etc. Back-propagation processor BPPo modifies the weights read from word storage elements in neural net layer L.sub.0 interstitial memory array by .eta. x.sub.i .delta..sub.j amounts and writes them back to the word storage elements in a sequence of read-modify-write cycles during the training procedure. Back-propagation processor BPP1 modifies the weights read from word storage elements in neural net layer L.sub.1 interstitial memory array by .eta. x.sub.h .delta..sub.i amounts and writes them back to the word storage elements in a sequence of read-modify-write cycles, during the training procedure. Back-propagation processor BPP.sub.2 modifies the weights read and storage elements in neural net layer L.sub.2 interstitial memory array by .eta. x.sub.g .delta..sub.h amounts and writes them back to the word storage element in a sequence of read-modify-write cycles during the training procedure.

The neural nets thusfar described make extensive use of pairs of capacitors wherein the capacitances of each pair of capacitances are determined in response to a digital word, sum to a prescribed constant value, and differ so as to determine the weighting to be applied to a synapse input signal. The foregoing specification also describes the usefulness in a neural net of two pairs of capacitors, wherein the capacitances of each pair of capacitances are determined in response to the same digital word. While the capacitors C.sub.i,i and C.sub.i,(j+N) for all i and j could in some instances be fixed-value capacitors, so there would never be any alteration in the weighting of input voltages x.sub.i where i=1, . . . M, such neural nets lack the capacity to adapt to changing criteria for neural responses. Such adaptation is necessary, for example, in a neural network that is to be connected for self-learning.

So it is desirable to provide for altering the capacitances of each pair of capacitors C.sub.i,j and C.sub.(i+M),j associated in FIG. 1 or 2 neural net with a respective pair of values of i and j. This alteration is to be carried out in a complementary way, so the sum of the capacitances of C.sub.i,j and of C.sub.(i+M),j remains equal to C.sub.k. In FIG. 2 neural net it is also desirable to provide for altering the capacitances of each pair of capacitors C.sub.i,(j+N) and C.sub.(i+M),(j+N) in a complementary way, so the sum of their capacitances remains equal to C.sub.k.

Similarly, it is desirable to provide for altering the capacitances of each pair of capacitors C.sub.i,j and C.sub.i,(j+N) associated in FIG. 3 or 4 neural net with a respective pair of values of i and j. This alteration is to be carried out in a complementary way, so the sum of the capacitances of C.sub.i,j and of C.sub.i,(j+N) remains equal to C.sub.k. In FIG. 4 neural net it is also desirable to provide for altering the capacitances of each pair of capacitors C.sub.(i+M),j and C.sub.(i+M),(j+N) in a complementary way, so the sum of their capacitances of remains equal to C.sub.k.

Altering the capacitances of each pair of capacitors can be implemented along the lines of W. E. Engeler's teachings in regard to "digital" capacitors having capacitances controlled in proportion to binary-numbers used as control signals, as particularly disclosed in connection with FIG. 11 of his U.S. Pat. No. 3,890,635 issued 17 June 1975, entitled "VARIABLE CAPACITANCE SEMICONDUCTOR DEVICES" and assigned to General Electric Company. Each pair of capacitors is then two similar ones of these capacitors and their capacitances are controlled by respective control signals, one of which is the one's complement of the other. Another way of realizing the pair of capacitors is to control the inverted surface potentials of a pair of similar-size metal-oxide-semiconductor (MOS) capacitors with respective analog signals developed by digital-to-analog conversion.

Such methods of constructing a pair of capacitors use separate capacitive element structures for each capacitor, portions of which capacitive element structures are unused when weighting values are chosen to be low. This undesirably tends to make the capacitive element structures take up nearly twice as much area on an integrated circuit die than is necessary, it is here pointed out.

In the method of constructing a pair of capacitors in accordance with the invention, each of a set of component capacitors with capacitances related in accordance with powers of two is selected to be a component of one or the other of the pair of capacitors, the selecting being done by field effect transistors (FETs) operated as transmission gates. This method, which requires a minimum capacitor size providing capacitance half as large as the capacitance associated with the minimum weight, easily provides for a 2.sup.4 :1 range of capacitive weights without requiring much concern about unbalanced stray capacitances on the balanced input lines or balanced output lines affecting the accuracy of the scaling of the differential capacitance between those lines. With present day design rules a minimum-area capacitor of three square microns is feasible, which makes a capacitor eight times as large have an area of twenty-four square microns. There is no unused portion of the capacitive element structures in this method.

FIG. 9 is a conceptual schematic diagram of a pair of capacitors with digitally programmed capacitances, designed in accordance with the invention to be operated as a half bridge. The two capacitors are provided by selective connection of component capacitive elements C0, C1, C2, C3 and C4, which have a shared first plate labelled COMMON LINE and have respective second plates POSITIVE LINE and NEGATIVE LINE. As related to FIG. 1, to FIG. A or to FIG. 6A, COMMON LINE corresponds to the single-ended input line IL.sub.i ; and POSITIVE LINE and NEGATIVE LINE correspond to the balanced output lines OL.sub.j and OL.sub.(j+N), respectively. As related to FIG. 3 or to FIG. 4A, COMMON LINE corresponds to the single-ended output line OL.sub.j ; and POSITIVE LINE and NEGATIVE LINE correspond to the balanced input lines IL.sub.i and IL.sub.(i+M), respectively.

Component capacitive elements C0, C1, C2, C3 and C4 of the FIG. 9 capacitor pair have respective capacitances weighted in 2.sup.0 :2.sup.0 :2.sup.1 :2.sup.2 :2.sup.3 ratio; have respective first plates each connected to COMMON LINE; and have respective second plates connected by respective ones of single-pole-double-throw electronic switches SW1, SW2, SW3 and SW4 each to POSITIVE LINE or to NEGATIVE LINE. Single-pole-double-throw electronic switches SW1, SW2, SW3 and SW4 each provide for connection to POSITIVE LINE or to NEGATIVE LINE, as determined by a respective bit of a weighting word, which word is stored in a respective word storage element WSE.sub.i,j. The least significant bit of the weighting word is stored in a bit store BS1 and successively more significant bits are stored in bit stores BS2, BS3 and BS4 FIG. 9 shows bit stores BS1, BS2, BS3 and BS4 as respective square boxes within the rectangular box representing word storage element WSE.sub.i,j.

A tabulation of the various connections that can be made responsive to a four-bit weighting word, as shown in FIG. 10, suffices to indicate that a continuous set of incremental weights extending over a range with both positive and negative values is made possible by the method of constructing a pair of capacitors in accordance with the invention. The most significant bit of the four-bit weighting word governs connection of capacitive element C4 to POSITIVE LINE or to NEGATIVE LINE by electronic switch SW4 in the reverse sense that the less significant bits of the weighting word govern connections of capacitive elements C1, C2 and C3 to POSITIVE LINE or to NEGATIVE LINE by electronic switches, or multiplexers SW1, SW2 and SW3. A feeling of why this is done (which is to accommodate the use of two's complement numbers) and a feeling of why the NEGATIVE LINE is provided with the bias capacitive element C0 can be gotten from study of the FIG. 10 table.

The first, second, third and fourth component terms in the NEGATIVE LINE and POSITIVE LINE capacitances (which are to the COMMON LINE) are determined by which of the NEGATIVE LINE and POSITIVE LINE capacitive elements with weights of four, two, one, and one-half are switched to, responsive to the most significant bit, the secondmost significant bit, the thirdmost significant bit and the least significant bit of the four-bit weighting word, respectively, as read from left to right. The fifth component term in the NEGATIVE LINE capacitance is the capacitance of the bias capacitive element corresponding to the bias capacitive element CO of FIG. 9. This fifth component term is constant in its application and is not switched responsive to the weighting word or any bit thereof.

FIG. 11 shows another pair of capacitors with digitally programmed capacitances, designed in accordance with the invention for half bridge operation, which pair is composed of component capacitive elements C20, C21, C22, C23 and C24. Capacitive elements C20, C21, C22, C23 and C24 have respective capacitances weighted in 2.sup.0 :2.sup.0 :2.sup.1 :2.sup.2 :2.sup.3 ratio; have respective first plates each connected to COMMON LINE; and have respective second plates connected by respective ones of single-pole-double-throw electronic switches SW20, SW21, SW22, SW23 and SW24 each to POSITIVE LINE or to NEGATIVE LINE. Single-pole-double-throw electronic switches SW20, SW21, SW22, SW23 and SW24 each provide for connection to POSITIVE LINE or to NEGATIVE LINE, as determined by a respective bit of a weighting word, which word is stored in a respective word storage element WSE.sub.i,j '. FIG. 11 shows bit stores BS14, BS10, BS13, BS12 and BS11 as respective square boxes arranged from left to right within the rectangular box representing word storage element WSE.sub.i,j '. This arrangement from left to right is in accordance with the bit order of a binary diminished radix number system known as a one's complement system because the sign of the number can be reversed simply by replacing each bit in the number with the bit complement. The one's complement system codes arithmetic zero in two ways. The arrangement of bits shown in FIG. 11 allows conventional digital adders to be used for signed addition in one's complement arithmetic through the expedient of end-around carry, wherein the carry out from the single-bit adder generating the leftmost bit-place of the plural-bit sum is returned as carry in to the single-bit adder generating the rightmost bit-place of the plural-bit sum. It is this arrangement of bits that will be presumed to be used for one's complement arithmetic in the remainder of this specification. The bits in bit stores BS14, BS10, BS13, BS12 and BS11 control electronic switches SW24. SW20, SW23, SW22 and SW21, respectively. The leftmost bit of the five-bit weighting word governs connection of capacitive element C24 to POSITIVE LINE or to NEGATIVE LINE by electronic switch SW24 in the reverse sense that the bits of the weighting word to the right govern connections of capacitive elements C20, C23, C22 and C21 to POSITIVE LINE or to NEGATIVE LINE by electronic switches, or multiplexers SW20, SW23, SW22 and SW21.

Study of the FIG. 12 table, which tabulates the capacitances between COMMON LINE and each of the POSITIVE LINE and NEGATIVE LINE connections for the various one's complement numbers used as weighting words in the FIG. 11 pair of digitally programmable capacitors, helps provide a feeling of how this accommodates the use of one's complement numbers. The first, second, third, fourth and fifth component terms in the NEGATIVE LINE and POSITIVE LINE capacitances (which are to the COMMON LINE) are determined by which of the NEGATIVE LINE and POSITIVE LINE capacitive elements with weights of four, one-half two, one, and one-half are switched to, responsive to the respective bits of the five-bit weighting word as read from left to right.

The arithmetic of the one's complement number system for the set of weighting words used in the FIG. 11 capacitor pair has two zeroes, a "positive" zero of 00000 and a "negative" zero of 11111. When the interstitial memory array IMA has its contents modified by the increment .eta.x.sub.i .delta..sub.j supplied in two's complement form from multiplier MULT as shown in FIG. 7, the effects on the neural net training of this double zero arithmetic can be suppressed by the subtractor SUB being of the type using end-around carry. Alternatively, the effects of the double zero arithmetic may be left unsuppressed by subtractor SUB to make change between excitory and inhibitory weights less inclined to happen.

The FIG. 11 capacitor pair has the advantage for half bridge operation that the range of excitory weights is as wide as the range of inhibitory weights, which may overweigh in design considerations the disadvantages of the extra bit storage and somewhat more complicated arithmetic used in training. Symmetry in the ranges of available excitory and inhibitory weights is an especially important design consideration where the range of available weights is very restricted--such as just -1, 0 and +1 in a modification of the FIG. 11 capacitor pair wherein elements C22, SW22, C23, SW23, C24, SW24, BS12, BS13 and BS14 are dispensed with. Also, though there is extra bit storage in the FIG. 11 capacitor pair than in the FIG. 9 capacitor pair, there is greater regularity in the weighting capacitor pair structure with all component MOS capacitors being switched.

AMOS capacitor in a monolithic integrated circuit inherently has a substantial stray capacitance to substrate ground from the one of its two plates adjacent to the substrate. To prevent such stray capacitance appearing in unbalanced form, more on one of the POSITIVE LINE and NEGATIVE LINE than on the other, the component capacitive elements C0, C1, C2, C3 and C4 of FIG. 9 are poled so as to place their stray capacitances to substrate on the COMMON LINE. Similarly, in FIG. 11 the component capacitive elements C20, C21, C22, C23 and C24 are poled so as to place their stray capacitances to substrate on the COMMON LINE. This balancing out of stray capacitance is also done in the capacitor quads of FIGS. 24 and 25.

The multiplexers employed in various portions of the circuits described in this specification are customarily constructed of single-pole switch elements, similar to the SW1, SW2, SW3 and SW4 switch elements of FIG. 9 and to the SW21, SW22, SW23 and SW24 switch elements of FIG. 11. Each of these single-pole switch elements is conventionally a pair of so-called "transmission gate" connections of one or more field effect transistors in CMOS design. A suitable transmission gate is provided by the paralleled channels of a p-channel FET and an n-channel FET having oppositely swinging control voltages applied to their respective gate electrodes to control the selective conduction of those paralleled channels.

FIG. 9 shows the SW4 switch element connected to POSITIVE LINE and NEGATIVE line oppositely from the SW1, SW2 and SW3 switch elements; and FIG. 11 shows the SW24 switch element connected to POSITIVE LINE and NEGATIVE line oppositely from the SW1, SW2 and SW3 switch elements. In an actual integrated circuit layout of the FIG. 9 capacitor pair, rather than this being done for the various word storage elements, the bit lines can write to the opposite halves of the flip-flops providing bit stores BS4 that the bit lines write to the flip-flops providing bit stores BS1, BS2 and BS3. Similarly, in an actual integrated circuit layout of the FIG. 11 capacitor pair, rather than this being done for the various word storage elements, the bit lines can write to the opposite halves of the flip-flops providing bit stores BS24 that the bit lines write to the flip-flops providing bit stores BS21, BS22 and BS23.

FIG. 13 shows in detail the electrical connections to a corresponding memory element MEX of capacitive element CX that is one of the FIG. 9 switched capacitive elements C1, C2, C3 and C4. The memory element MEX is a flip-flop connection of enhancement-mode field effect transistors (FETs) Q1, Q2, Q3 and Q4 and is supplied a relatively positive operating supply voltage V.sub.DD and a relatively negative operating supply voltage VSS. One plate of capacitive element CX is shown with fixed connection to COMMON LINE. The switch element SWX, used to connect the other plate of capacitive element CX selectively either to POSITIVE LINE or to NEGATIVE LINE, comprises enhancement-mode FETs Q5, Q6, Q7 and Q8. Enhancement-mode n-channel FETs Q10 and Q11 are selectively rendered conductive by a WRITE command applied to their gate electrodes to impose the bit conditions D and DBAR on the complementary output connections Q and QBAR of memory element MEX to switch element SWX.

If D is high (e.g., V.sub.DD) and DBAR is low (e.g., VSS), when the WRITE command is no longer applied to the gate electrodes of Q10 and Q12, DBAR being low conditions Q1 to be conductive and Q2 to be nonconductive to maintain Q high at V.sub.DD potential; and D being high conditions Q3 to be nonconductive and Q4 to be conductive to maintain QBAR low at V.sub.SS potential. In switch element SWX, Q being high conditions p-channel FET Q5 for non-conduction and n-channel FET Q6 for conduction, and QBAR being low conditions p-channel FET Q7 for conduction and n-channel FET Q8 for non-conduction. The NEGATIVE LINE is connected to capacitive element CX by FETs Q6 and Q7 being conditioned for conduction; and the POSITIVE LINE is disconnected from capacitive element CX by FETs Q5 and Q8 being conditioned for nonconduction.

If D is low (e.g., V.sub.SS) and DBAR is high (e.g., V.sub.DD), when the WRITE command is no longer applied to the gate electrodes of Q10 and Q12, DBAR being high conditions Q1 to be nonconductive and Q2 to be conductive to maintain Q low at Vss potential; and D being low conditions Q3 to be conductive and Q4 to be nonconductive to maintain QBAR high at V.sub.DD potential. In switch element SWX, Q being low conditions p-channel FET Q5 for conduction and n-channel FET Q6 for non-conduction, and QBAR being high conditions p-channel FET Q7 for non-conduction and n-channel FET Q8 for conduction. The POSITIVE LINE is connected to capacitive element CX by FETs Q5 and Q8 being conditioned for conduction; and the NEGATIVE LINE is disconnected from capacitive element CX by FETs Q6 and Q7 being conditioned for nonconduction.

FIGS. 14-23 depict two basic monolithic structure cells, the one at the left of these figures and the one at the left of these figures being laid out as mirror duplicates of each other. The basic monolithic structure cell can be replicated a small number of times to form a pair of weighting capacitors composed of switched capacitive elements. A network of such pairs of weighting capacitors formed by row and column array of the basic monolithic structure cells, with adjacent column structures mirroring each other, can be used in constructing weighted summation circuitry in accordance with the invention. The smallest switched capacitive element(s) can each consist of one basic monolithic structure cell, and the larger switched capacitive elements can each be a combination of a plurality of the basic monolithic structure cells having electrical connections in parallel. The bit storage element need not be duplicated if a single bit storage element is used to control both sets of electronic switches in the combination, however, and modifications of the monolithic circuit layout that eliminate the redundant bit storage element are possible. A quad of weighting capacitors composed of switched capacitive elements can simply comprise two pairs of weighting capacitors, each pair composed of switched capacitive elements as just described, but again there are redundant bit storage elements. Layout modifications are possible that can remove these redundant bit storage elements, so there is the possibility of layout reconfiguration to save area on the monolithic die. Reconfigurations where the first and second switched capacitive elements controlled by the same bit storage element adjoin each other can give. rise to T-shaped or L-shaped basic monolithic structure cells. A row of T-shaped or L-shaped basic monolithic structure cells exhibits teeth along one edge of the row. Where there are rows of T-shaped or L-shaped basic monolithic structure cells, a technique to save area on the monolithic die is to flip alternate rows of cells about their row axes so their teeth can be interleaved.

Monolithic structure cells as shown in FIGS. 14--23 have been used in implementing a network of pairs of weighting capacitors where two-bit digital words encode weighting factors in one's complement arithmetic. Each weighting capacitor in the pair has an effective weight of 0.5.

In neural net layers of the type using capacitive structures connecting input lines and differentially sensed output line pairs, the bits of the digital word stored in the two bit storage elements of a basic monolithic structure cell both being ZEROs or both being ONEs causes the pair of weighting capacitors to be connected to opposite ones of a pair of differentially sensed output lines, so an effective weight of zero is provided by the capacitor pair. The digital word stored in the two bit storage elements being 01 causes both weighting capacitors to be connected to the positive output line, so the capacitor pair provides an effective weight of +1. The digital word being 10 causes, both weighting capacitors to be connected to the negative output line, so the capacitor pair provides an effective weight of -1.

In neural net layers of the type using capacitive structures connecting output lines and differentially driven input line pairs, the digital word stored in the two bit storage elements being either 00 and 11 causes the two weighting capacitors to be connected to opposite ones of a pair of differentially driven input lines, so an effective weight of zero is provided by the capacitor pair. The digital word stored in the two bit storage elements being 01 causes both weighting capacitors to be connected to the positive input line, so the capacitor pair provides an effective weight of +1. The digital word stored in the two bit storage elements being 10 causes both weighting capacitors to be connected to the negative input line, so the capacitor pair provides an effective weight of -1.

FIG. 24 is a conceptual schematic diagram of how a quad connection of four capacitors having their capacitances digitally controlled by two's complement weighting factor words can be constructed in accordance with the invention. Lines LINE1, LINE2, LINE3 and LINE4 of FIG. 24 may be considered to correspond to POSITIVE LINE, to NEGATIVE LINE, to COMMON LINE and to ac ground, respectively, of FIG. 9. One pair of the four capacitors in the quad is provided by selective connection of component capacitive elements C0, C1, C2, C3 and C4 to lines LINE1 and LINE2. Capacitive elements C0, C1, C2, C3 and C4 have respective capacitances weighted in 2.sup.0 :2.sup.0 :2.sup.1 :2.sup.2 :2.sup.3 ratio; have respective first plates each connected to LINE3; and have respective second plates connected by respective ones of single-pole-double-throw electronic switches SW1, SW2, SW3 and SW4 each to LINE1 or to LINE2. Single-pole-double-throw electronic switches SW1, SW2, SW3 and SW4 each provide for connection to LINE1 or to LINE2, as determined by a respective bit of a weighting word, which word is stored in a respective word storage element WSE.sub.i,j. The other pair of the four capacitors in the quad is provided by selective connection of component capacitive elements C10, C11, C12, C13 and C14. Capacitive elements C10, C11, C12, C13 and C14 have respective capacitances weighted in 2.sup.0 :2.sup.0 :2.sup.1 :2.sup.2 :2.sup.3 ratio; have respective first plates each connected to LINE4; and have respective second plates connected by respective ones of single-pole-double-throw electronic switches SW11, SW12, SW13 and SW14 each to LINE 1 or to LINE2. Component capacitive elements C10, C11, C12, C13 and C14 have respective capacitances which correspond to the respective capacitances of component capacitive elements C0, C1, C2, C3 and C4, respectively. Single-pole-double-throw electronic switches SW11, SW12, SW13 and SW14 each provide for connection to LINE1 or to LINE2 in a way complementary to the way the electronic switches SW1, SW2, SW3 and SW4 do, as determined by a respective bit of the weighting word stored in a respective word storage element WSE.sub.i,j.

Lines LINE1, LINE2, LINE3 and LINE4 of FIG. 24 have thusfar been considered to correspond to POSITIVE LINE, to NEGATIVE LINE, to COMMON LINE and to ac ground, respectively. In such case, to prevent stray capacitance to substrate ground appearing in unbalanced form on the POSITIVE LINE and the NEGATIVE LINE--i.e., more on one of the lines LINE1 and LINE2 than on the other--the component capacitive elements C0, C1, C2, C3 and C4 are poled so as to place their stray capacitances to substrate on LINE3, the COMMON LINE. The component capacitive elements C10, C11, C12, C13 and C14 are poled so as to connect to LINE4, and thence to substrate ground, their stray capacitances to substrate, which avoids their having to be charged and discharged at all.

Alternatively, lines LINE1, LINE2, LINE3 and LINE4 of FIG. 24 can be considered to correspond to COMMON LINE, to ac ground, to POSITIVE LINE and to NEGATIVE LINE, respectively. In such case, to prevent stray capacitance to substrate ground appearing in unbalanced form, more on one of the lines POSITIVE LINE and NEGATIVE LINE than on the other, the component capacitive elements C0, C1, C2, C3 and C4 are poled so as to place their stray capacitances to substrate on the POSITIVE LINE, LINE3; and the component capacitive elements C10, C11, C12, C13 and C14 are poled so as to place their stray capacitances to substrate on the NEGATIVE LINE, LINE4. However, this method of avoiding unbalanced stray capacitance to substrate appearing on the POSITIVE LINE and NEGATIVE LINE relies on matching between the stray capacitances of component capacitive elements C0, C1, C2, C3 and C4 and the stray capacitances of component capacitive elements C10, C11, C12, C13 and C14, unless the stray capacitances to substrate of the POSITIVE LINE and of the NEGATIVE LINE is shunted by relatively low source impedance balanced drive from differential amplifier ID.sub.i. So, the former method of avoiding unbalanced stray capacitance to substrate appearing on the POSITIVE LINE and NEGATIVE LINE is preferred over the latter method, at least where these lines are to differentially sensed by an output driver amplifier OD.sub.j.

The FIG. 6 type of neural net uses pairs of input lines driven by balanced input signals for connection to the pairs of differentially sensed output lines by weighting capacitors connected in quad configurations and operated as full bridges. In the FIG. 6 type of neural net, lines LINE3 and LINE4 of FIG. 24 can be chosen to correspond to one of the pairs of input lines driven by balanced input signals, and LINE1 and LINE2 can be considered to correspond to one of the the pairs of differentially sensed output lines. This choice arranges for the unbalanced stray capacitance to substrate appearing on LINE3 and LINE4 to be shunted by relatively low source impedance balanced drive from differential amplifier ID.sub.i.

FIG. 25 is a conceptual schematic diagram of how a quad connection of four capacitors having their capacitances digitally controlled by one's complement weighting factor words can be constructed in accordance with the invention. One pair of the four capacitors is provided by selective connection of component capacitive elements C20, C21, C22, C23 and C24. Capacitive elements C20, C21, C22, C23 and C24 have respective capacitances weighted in 2.sup.0 :2.sup.0 :2.sup.1 :2.sup.2 :2.sup.3 ratio; have respective first plates each connected to LINE3; and have respective second plates connected by respective ones of single-pole-double-throw electronic switches SW20, SW21, SW22, SW23 and SW24 each to LINE1 or to 2LINE. Single-pole-double-throw electronic switches SW24, SW20, SW23, SW22 and SW21 each provide for connection to LINE1 or to LINE2, as determined by successive bits of a weighting word stored in bit stores BS14, BS10, BS13, BS12 and BS11 of a respective word storage element WSE.sub.i,j '. The other pair of the four capacitors is provided by selective connection of component capacitive elements C30, C31, C32, C33 and C34. Capacitive elements C30, C31, C32, C33 and C34 have respective capacitances weighted in 2.sup.0 :2.sup.0 :2.sup.1 :2.sup.2 :2.sup.3 ratio; have respective first plates each connected to LINE 4; and have respective second plates connected by respective ones of single-pole-double-throw electronic switches SW30, SW31, SW32, SW33 and SW34 each to LINE1 or to LINE2. Component capacitive elements C30, C31, C32, C33 and C34 have respective capacitances which correspond to the respective capacitances of component capacitive elements C20, C21, C22, C23 and C24, respectively. Single-pole-double-throw electronic switches SW34, SW30, SW33, SW32 and SW31 each provide for connection to LINE1 or to LINE2 in a way complementary to the way the electronic switches SW24, SW20, SW23, SW22 and SW21 do,, as determined by successive bits of a weighting word stored in bit stores BS14, BS10, BS13, BS12 and BS11 of a respective word storage element WSE.sub.i,j '.

In the FIG. 25 quad connection of capacitors, as in the FIG. 24 quad connection of capacitors, lines LINE1, LINE2, LINE3 and LINE4 can be considered to correspond to POSITIVE LINE, to NEGATIVE LINE, to COMMON LINE and to ac ground, respectively, or alternatively, can be considered to correspond to COMMON LINE, to ac ground, to POSITIVE LINE and to NEGATIVE LINE, respectively. In the FIG. 6 type of neural net lines LINE1 and LINE2 of FIG. 25 can be considered to correspond to one of the pairs of input lines driven by balanced input signals, and LINE3 and LINE4 can be considered to correspond to one of the the pairs of differentially sensed output lines. In any case, the component capacitive elements C20, C21, C22, C23 and C24 are poled so as to place their stray capacitances to substrate on LINE3; and the component capacitive elements C30, C31, C32, C33 and C34 are poled so as to place their stray capacitances to substrate on LINE4. The reasons for this are the same as offered in regard to the FIG. 24 quad connection of capacitors.

FIG. 26 illustrates how, in order to provide greater resolution in weighting, at every i,j crosspoint in the neural net layer a plurality of similar weighting capacitor quads are employed rather than just a single weighting capacitor quad comprising digital capacitors DC.sub.i,j, DC.sub.(i+m),j, DC.sub.i,(j+n) and DC.sub.(i+m),(j+n). In FIG. 26, one weighting capacitor quad MSWC.sub.i,j is used to provide weighting responsive to the leftmost (generally more significant) bits of the eight-bit weighting word stored in the word storage element WSE.sub.i,j of the interstitial memory array, and another weighting capacitor quad LSWC.sub.i,j is used to provide weighting responsive to the rightmost (generally less significant) bits of the weighting word stored in the word storage element WSE.sub.ij of the interstitial memory array. The relative significances of the weighting provided by the weighting capacitor quad MSWC.sub.i,j and by the other weighting capacitor quad LSWC.sub.i,j are in a prescribed ratio.

Much as previously, weighting capacitor quad MSWC.sub.i,j supplies to a charge-sensing amplifier FDQS.sub.j weighted response to x.sub.i input signal. Another charge-sensing amplifier FDQS.sub.(j+n) besides charge sensing amplifier FDQS.sub.j is used for the forward propagation of signals at the i,j crosspoint of the neural net layer and is supplied via weighting capacitor quad LSWC.sub.i,j with another weighted response to x.sub.i input signal. The output port of analog multiplier AM.sub.j supplies analog signal to the input port of an analog scaling amplifier ASA.sub.j, and the analog scaling amplifier ASA.sub.j responds with an analog signal at its output port that is scaled down from the analog signal at its input port by a factor equal to the ratio of the relative significances of the weighting provided by the weighting capacitor quad MSWC.sub.i,j and by the other weighting capacitor quad LSWC.sub.i,j. The input port of charge-sensing amplifier FDQS.sub.(j+N) and the output port of the analog scaling amplifier ASA.sub.j are multiplexed to the weighting capacitor quad LSWC.sub.i,j via output lines OL.sub.(j+ 2N) and OL.sub.(j+3N) by multiplexers OLM.sub.(j+2N) and OLM.sub.(j+3N). Differential-input charge-sensing amplifiers FDQS.sub.j and FDQS.sub.(j+N) are both linear charge-sensing amplifiers; and in a weight and sum circuit FW&S.sub.j their output responses are added after being scaled in accordance with the ratio of the relative significances of the weighting provided by the weighting capacitor quad MSWC.sub.i,j and by the other weighting capacitor quad LSWC.sub.i,j. Weight and sum circuit FW&S.sub.j thereby generates the input signal for non-linear amplifier NL.sub.i,j. For example, suppose the word storage element WSE.sub.i,j stores an eight-bit word describing a weight in two's complement arithmetic, four bits of which control the weighting afforded by the weighting capacitor quad MSWC.sub.i,j and four bits of which control the weighting afforded by the weighting capacitor quad LSWC.sub.i,j. Then, the output signal from charge-sensing amplifiers FDQS.sub.j and FDQS.sub.(j+N) are weighted in 16:1 ratio before being summed in the weight and sum circuit FW&Sj to generate the input signal for the non-linear amplifier NL.sub.j.

While FIG. 26 shows portions of a neural net layer in which neither the input lines to or the output lines from the weighting capacitance network are single-ended, modifications to provide for single-ended operation per FIGS. 2A and 2B or per FIGS. 4A and 4B are possible, of course, without affecting the way in which different magnitudes of weighting are achieved with groups of switched capacitor elements which groups are alike. While FIG. 26 shows two ranks of switched capacitor elements, one for responding to the more significant bits of the weighting words stored in the interstitial memory array and the other for responding to the less significant bits of the weighting words stored in the interstitial memory array, there may instead be three or more such ranks of switched capacitor elements in accordance with the invention. Indeed, bit-slicing the weighting factors may be a preferred structure for neural net layers, owing to the facts that each portion of a neural net layer associated with a single bit slice of the weighting factors is like the other portions of the neural net layer respectively associated with the other bit slices of the weighting factors; and that that single design can use the minimum-size digitally controlled capacitors throughout.

FIG. 27 shows more particularly the nature of two corresponding pairs of capacitors in the weighting capacitor quads MSWC.sub.i,j and LSWC.sub.i,j when the word storage element WSE.sub.i,j stores an eight-bit weight in two's complement arithmetic. One pair of capacitors in the quad is formed from the capacitive elements C0, C1, C2, C3, and C4, which determine the differential capacitance to COMMON LINE from lines POSITIVE LS LINE and NEGATIVE LS LINE in the weighting capacitor quad LSWC.sub.i,j, and from the capacitive elements C5, C6, C7, and C8, which determine the differential capacitance to COMMON LINE from lines POSITIVE MS LINE and NEGATIVE MS LINE in the weighting capacitor quad MSWC.sub.i,j. The capacitive elements C5, C6, C7 and C8 in the weighting capacitor quad MSWC.sub.i,j have capacitances respectively similar to those of capacitive elements C1, C2, C3, and C4 in the weighting capacitor quad LSWC.sub.i,j. The other pair of capacitors in the quad is formed from the capacitive elements C10, C11, C12, C13, and C14, which determine the differential capacitance to signal ground from lines POSITIVE LS LINE and NEGATIVE LS LINE in the weighting capacitor quad LSWC.sub.i,j, and from the capacitive elements C15, C16, C17, and C18, which determine the differential capacitance to signal ground from lines POSITIVE MS LINE and NEGATIVE MS LINE in the weighting capacitor quad MSWC.sub.i,j. The capacitive elements C15, C16, C17 and C18 in the weighting capacitor quad MSWC.sub.i,j have capacitances respectively similar to those of capacitive elements C11, C12, C13, and C14 in the weighting capacitor quad LSWC.sub.i,j. Capacitive elements in the capacitor quad that have the same last digit in their call-out have capacitances that are similar to each other.

FIG. 27 shows bit stores BS8, BS7, BS6, BS5, BS4, BS3, BS2 and BS1, which respectively store progressively less significant bits of a weighting word, as respective square boxes arranged from left to right within the rectangular box representing the word storage element WSE.sub.i,j. The most significant bit of the eight-bit weighting word governs connection of capacitive element C8 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switch SW8 in the reverse sense that the three next most significant bits of the eight-bit weighting word govern connection of capacitive elements C7, C6 and C5 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switches SW7, SW6, and SW5 and in the reverse sense that the four least significant bits of the weighting word govern connections of capacitive elements C1, C2, C3 and C4 to POSITIVE LS LINE or to NEGATIVE LS LINE by electronic switches SW1, SW2, SW3 and SW4. The four most significant bits of the eight-bit weighting word govern connection of capacitive elements C18, C17, C16 and C15 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switches SW18, SW17, SW16, and SW15 in the reverse sense that it governs connection of capacitive elements C8, C7, C6 and C5 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switches SW8, SW7, SW6 and SW5. The four least significant bits of the eight-bit weighting word govern connection of capacitive elements C14, C13, C12 and C11 to POSITIVE LS LINE or to NEGATIVE LS LINE by electronic switches SW14, SW13, SW12, and SW11 in the reverse sense that it governs connection of capacitive elements C4, C3, C2 and C1 to POSITIVE LS LINE or to NEGATIVE LS LINE by electronic switches SW4, SW3, SW2 and SW1.

In the FIG. 26 neural net layer using capacitor quads per FIG. 27, then, when respective capacitive elements C8 and C18 are included in each quad of capacitors in the weighting capacitor quad MSWC.sub.i,j ; these respective capacitive elements C8 and C18 are switched in reverse sense from all other capacitive elements in that quad of capacitors to implement two's complement arithmetic in regard to weighting value. Respective non-switched C0 and C10 minimum-weight capacitive elements are included in each quad of capacitors in the weighting capacitor quad LSWC.sub.i,j, to bias the zero-capacitance condition to correspond to the all-ZERO condition of the two's complement numbers that can be stored in the word storage element WSE.sub.i,j of the interstitial memory array. However, the weighting capacitor quad MSWC.sub.i,j includes no non-switched minimum-weight capacitive elements.

FIG. 28 shows more particularly the nature of a quad of capacitors in the weighting capacitance networks MSWC.sub.i,j and LSWC.sub.i,j when the word storage element WSE.sub.i,j stores nine-bit weights in one's complement form in bit stores BS10, BS11, BS12, BS13, BS14, BS15, BS16, BS17 and BS18. One pair of capacitors in the quad is formed from the capacitive elements C20, C21, C22, C23, and C24, which determine the differential capacitance to COMMON LINE from lines POSITIVE LS LINE and NEGATIVE LS LINE in the weighting capacitor quad LSWC.sub.i,j, and from the capacitive elements C25, C26, C27, and C28, which determine the differential capacitance to COMMON LINE from lines POSITIVE MS LINE and NEGATIVE MS LINE in the weighting capacitor quad MSWC.sub.i,j. The capacitive elements C25, C26, C27 and C28 in the weighting capacitor quad MSWC.sub.i,i have capacitances respectively similar to those of capacitive elements C21, C22, C23, and C24 in the weighting capacitor quad LSWC.sub.i,j. The other pair of capacitors in the quad is formed from the capacitive elements C30, C31, C32, C33, and C34, which determine the differential capacitance to signal ground from lines POSITIVE LS LINE and NEGATIVE LS LINE in the weighting capacitor quad LSWC.sub.i,j, and from the capacitive elements C35, C36, C37, and C38, which determine the differential capacitance to signal ground from lines POSITIVE MS LINE and NEGATIVE MS LINE in the weighting capacitor quad MSWC.sub.i,j. The capacitive elements C35, C36, C37 and C38 in the weighting capacitor quad MSWC.sub.i,j have capacitances respectively similar to those of capacitive elements C31, C32, C33, and C34 in the weighting capacitor quad LSWC.sub.i,j. Capacitive elements in the capacitor quad that have the same last digit in their call-out have capacitances that are similar to each other. The weighting capacitor quad MSWC.sub.i,j has no capacitive element corresponding to capacitive element C20 or C30 in the weighting capacitor quad LSWC.sub.i,j.

FIG. 28 shows bit stores BS18, BS10, BS17, BS16, BS15, BS14, BS13, BS12 and BS11 which respectively store successive bits of a nine-bit weight in one's complement arithmetic, as respective square boxes arranged from left to right within the rectangular box representing the word storage element WSE.sub.i,j '. The leftmost bit of the nine-bit weighting word stored in bit store BS18 governs connection of capacitive element C28 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switch SW28, depending whether that bit is a ONE or a ZERO; and that leftmost bit also governs connection of capacitive element C38 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switch SW38, depending whether that bit is a ZERO or a ONE. The next to leftmost bit of the nine-bit weighting word stored in bit store BS10 governs connection of capacitive element C20 to POSITIVE LS LINE or to NEGATIVE LS LINE by electronic switch SW20, depending whether that bit is a ONE or a ZERO; and that next to leftmost bit also governs connection of capacitive element C30 to POSITIVE LS LINE or to NEGATIVE LS LINE by electronic switch SW30, depending whether that bit is a ZERO or a ONE. The three next most left bits of the nine-bit weighting word stored in bit stores BS17, BS16 and BS15 govern connection of capacitive elements C27, C26 and C25 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switches SW27, SW26 and SW25, respectively, depending whether those bits are respectively each a ONE or a ZERO. The bits stored in bit stores BS17, BS16 and BS15 also govern connection of capacitive elements C37, C36 and C35 to POSITIVE MS LINE or to NEGATIVE MS LINE by electronic switches SW37, SW36 and SW35, respectively, depending whether those bits are respectively each a ZERO or a ONE. The four rightmost bits of the weighting word stored in bit stores BS11, BS12, BS13 and BS14 govern connections of capacitive elements C21, C22, C23 and C24 to POSITIVE LS LINE or to NEGATIVE LS LINE by electronic switches SW21, SW22, SW23 and SW24, respectively, depending whether those bits are respectively each a ONE or a ZERO. The four bits stored in bit stores BS11, BS12, BS13 and BS14 also govern connections of capacitive elements C31, C32, C33 and C34 to POSITIVE LS LINE or to NEGATIVE LS LINE by electronic switches SW31, SW32, SW33 and SW34, respectively, depending whether those bits are respectively each a ZERO or a ONE.

One skilled in the art and acquainted with the foregoing specification will be able to design numerous variants of the preferred embodiments of the invention described therein, and this should be borne in mind when construing the following claims. For example, the fixed-capacitance weighting capacitors in the matrix multipliers disclosed by W. E. Engeler in his U.S. Pat. No. 4,156,284 issued 22 May 1979 and entitled "SIGNAL PROCESSING APPARATUS" can be replaced in accordance with the invention with adjustable-capacitance weighting capacitors having their capacitances adjusted in accordance with digital signals stored in respective word storage elements of an interstitial memory array. As a further example, the fixed-capacitance switched capacitors described by Y. P. Tsividis and D. Anastassion in their letter "Switched-Capacitor Neural Networks" appearing in ELECTRONICS LETTERS, 27th Aug. 1987, Vol. 23, No. 18, pages 958,959 (IEE) can be replaced in accordance with the invention with adjustable-capacitance switched capacitors having their capacitances adjusted in accordance with digital signals stored in respective word storage elements of an interstitial memory array.

Claims

1. In an integrated-circuit having within its confines an array of weighted summation circuits, N in number, each weighted summation circuit capable of generating a weighted sum response to a common plurality of input signals, M in number, and including at least one corresponding capacitive element for determining weighting of each of said input signals within that said weighted summation circuit,

an improvement wherein said at least one corresponding capacitive element is of a type having its capacitance value determined incrementally in accordance with bits of a digital word which is received at a control input port.

2. The improvement as set forth in claim 1 wherein some of said corresponding capacitive elements are each of a type having its respective capacitance value determined incrementally in accordance with bits of a digital word received at a respective control input port thereof.

3. The improvement as set forth in claim 1 wherein each said corresponding capacitive element is of a type having its capacitance value determined incrementally in accordance with bits of a digital word received at a control input port therof.

4. An integrated-circuit having therewithin:

an array of weighted summation circuits, N in number, each weighted summation circuit capable of generating a respective weighted sum response to a common plurality of input signals, M in number, said input signals corresponding to respective ones of consecutive ordinal numbers first through M.sup.th, said weighted summation circuits corresponding to respective ones of consecutive ordinal numbers first through N.sup.th, the respective weighted sum response generated by each of said weighted summation circuits being identified by an ordinal number identifying said weighted summation circuit respectively generating said ordinal number;

an array of word storage elements, M times N in number, M being an integer more than one and N being an integer more than one, each word storage element for temporarily storing a respective digital word descriptive of a weighting to be afforded by said array of weighted summation circuits to one of said first through M.sup.th input signals in generating one of said first through N.sup.th weighted sum responses; and

within each of first through N.sup.th said weighted summation circuits, at least one corresponding capacitive element for determining the weighting of each of said first through M.sup.th input signals of a type having its capacitance value determined incrementally in accordance with bits of the respective digital word descriptive of that weighting, as temporarily stored in said array of word storage elements.

5. An improvement as set forth in claim 4 wherein each said word storage element is at a location within said integrated-circuit proximate to each said capacitive element having its capacitance value determined incrementally in accordance with bits of a digital word temporarily stored therein.

6. In a neural net layer having a plurality of weighted summation circuits, N in number, arrayed within the confines of an integrated-circuit; each weighted summation circuit capable of generating a weighted sum response to a common plurality of input signals, M in number, and including at least one corresponding capacitive element for determining weighting of each of said input signals within that said weighted summation circuit: an improvement wherein each corresponding capacitive element for determining the weighting of a respective one of said input signals within a respective one of said weighted summation circuits is a capacitive element having its capacitance value determined incrementally in accordance with bits of a digital word as temporarily stored in a respective word storage element included in an array of word storage elements within the confines of said integrated-circuit.

7. The improvement as set forth in claim 6 wherein each said word storage element is at a location within the confines of said integrated-circuit proximate to each said capacitive element having its capacitance value determined incrementally in accordance with the bits of a digital word temporarily stored therein.