COMPRISING RESISTIVE ELEMENTS"

~k ~k ~k ~k ~k

D E S C R I P T I O N TECHNICAL FIELD

The present invention refers to the field of mathematical calculations performed by electronic circuits employing resistive elements.

STATE OF THE ART

Resistive memories frequently find an application in mathematical calculation electronic circuits. With reference to this matter, document US-A-9152827 de ^¬ scribes a circuit provided with resistive memories organized in a crosspoint matrix and structured for performing a matrix-vector product operation of the type x = Ab, wherein x is a current vector, A is a conductance matrix, and b is a vector of voltages applied to each row of the crosspoint matrix.

Document US-A-2017/0040054 describes matrixes of resistive memories for computational accelerators, particularly for solving systems of equations and by resorting to iterative numeral techniques. It is ob ^¬ served that such solution requires several iteration to obtain the convergence.

An approach analogous to the one disclosed in document US-A-2017-0040054 is found in "Mixed-Preci ^¬ sion Memcomputing", M. Le Gallo et al . , arXiv: 1701.04279 [cs.ET] which faces the problem of solving linear systems by an iterative approach (Krylov subspaces) by combining matrixes of phase changing memories for a mixed precision solution.

SUMMARY OF THE INVENTION

The Applicant has observed that the computational methods of the prior art exhibit limitations both in relation to the type of executable computational op ^¬ erations (i.e., simple multiplications) and in terms of computational load, with reference to the possi ^¬ bility of solving algebraical problems.

A first object of the invention is a mathematical problem solving circuit as described in independent claim 1. Particular embodiments are described in de ^¬ pendent claims from 2 to 15.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will be described in the following with reference to non-limiting examples given only in an illustrative way in the following drawings. These drawings show different aspects and embodiments of the present invention and, when appropriate, similar structures, components, materials and/or elements in the different figures are indicated by the same ref ^¬ erence numbers .

- Figure 1 shows an example of a mathematical problem solving circuit;

- Figure 2 shows an example of a first circuit for solving square systems of equations;

- Figure 3 shows, by an example, a second circuit for solving square systems of equations;

- Figure 4 shows an example of a first circuit for calculating eigenvectors;

- Figure 5 shows an example of a second circuit for calculating eigenvectors;

- Figure 6 shows the first circuit for solving square systems of equations in an embodiment using three-terminal resistive elements;

- Figure 7 shows the first circuit for solving square systems of equations in an embodiment using three-terminal resistive elements combined with re ^¬ sistive memories;

- Figure 8 shows the first circuit for calculat ^¬ ing eigenvectors in an embodiment comprising an open- loop operational amplifier;

- Figure 9 shows an example of a linear regression accelerator circuit;

- Figure 10 shows an example of a neural network which is trained by using a linear regression accelerator circuit.

DE TAILED DE SCRIPTION OF THE INVENTION

While the invention is susceptible to different modifications and alternative constructions, some re ^¬ spective illustrated embodiments are shown in the drawings and will be described specifically in the following. Anyway, it is stated that there is no in ^¬ tention to limit the invention to the particular illustrated embodiment, on the contrary, the invention intends to cover all the modifications, alternative constructions, and equivalents falling in the scope of the invention as defined in the claims.

Figure 1 refers to an example of a circuit for solving mathematical problems 100. The solving cir ^¬ cuit 100 comprises a crosspoint matrix MG including a plurality of row conductors Li, a plurality of col ^¬ umn conductors C , and a plurality of analog resis ^¬ tive memories G±j each connected between a respective row conductor Li and a respective column conductor Cj .

For the purpose of the present invention, a re ^¬ sistive memory (also known as memristor) is a circuit element with two terminals (in other words a dipole) having a conductance which can be configured for tak ^¬ ing a value which is maintained until a possible new configuration is attained. For example, the following devices are resistive memories: the Resistive Random Access Memories (RRAM) , the Conductive Bridging Ran ^¬ dom Access Memories (CBRAM) , the Phase Change Memo ^¬ ries (PCM) , the Magnetoresistive Random Access Memo ^¬ ries (MRAM) of different types, the Ferroelectric Random Access Memories (FeRAM) of different types, the organic material memories, or other devices which can change the conductance thereof due to electric, magnetic fields, thermal, optical, mechanical opera ^¬ tions or any other type of operations, or a combina ^¬ tion thereof.

The above described resistive memories G±j are preferably of an analog type which means they can take continuous conductance values in an operative range; however, it is not excluded the possibility of resistive memories G±j of the digital type, this means they can take conductance values falling in a finite set of values.

The particular shown example illustrates a number N = 3 of row conductors Li, an equal number N = 3 of column conductors Cj and a number N x N = 3 x 3 of analog resistive memories represented by their conductance G±j .

Moreover, the circuit 100 comprises a plurality of operational amplifiers AOi having a closed-loop configuration, and each having: a first input terminal INli connected to a respective row conductor Li, a second input terminal IN2i connected to a ground terminal GR and an output terminal OUi. Particularly, the first input terminal INli is the inverting ter ^¬ minal of the respective operational amplifier OAi, while the second input terminal IN2i is the non-in ^¬ verting terminal of the respective operational am ^¬ plifier OAi .

The output terminal OUi of each operational am ^¬ plifier OAi is connected to a respective column C . Specifically, according to Figure 1, each output ter ^¬ minal OUi is connected to an end of a column Cj wherein j = i. Figure 1, according to the described example, shows three operational amplifiers OAi (OAI, OA2, OA3) and, therefore, three first input terminals INli (INli, IN12, IN13), three second input terminals IN2i (IN21, IN22, IN23), and three output terminals OUi (OUI, OU2, OU3) are shown.

Due to the negative feedback, it is observed that each operational amplifier AOi can operate in order to take the respective first input terminal INli to a virtual ground, in other words to take a voltage value near (theoretically equal to) the one of the ground GR assumed by the second input terminal IN2i. It is observed that the virtual ground is theoreti ^¬ cally assumed by operational amplifiers having an infinite gain. Particularly, each operational amplifier AOi has an inverting configuration.

With reference to the mathematical problem to be solved, the plurality of resistive memories Gij can be configured for representing, by respective con ^¬ ductance values (again indicated by Gi ), a first plurality of known values of the mathematical prob ^¬ lem.

Moreover, the circuit 100 enables to detect a plurality of electric magnitudes of the circuit it ^¬ self, adapted to be configurable to represent a sec ^¬ ond known value or a second plurality of known values of the mathematical problem to be solved. This second plurality of known values comprise, for example, val ^¬ ues of currents injected in each row conductor or the known value can be a conductance (or another parame ^¬ ter) of an additional electric component connectable to the plurality of operational amplifiers AOi.

The plurality of operational amplifiers AOi de ^¬ fine a plurality of output voltages Vi (measurable at each output terminal OUi) representative of a plu ^¬ rality of values solving the mathematical problem.

According to particular embodiments, the solving circuit 100 is useable for the approximate solution of problems of linear algebra, among which, are ex- emplifyingly listed the following:

- the solution of a square system of equations, expressable in a matrix form;

- the inversion of real square matrixes;

- the calculation of eigenvectors.

The solving circuit 100 can further comprise at least one apparatus for measuring voltage values Vi (not shown) . Such measuring apparatus can be analog (i.e., a potentiometer) or digital so that a conversion of the voltage values Vi from analog to digital is required.

It is observed that in the present description, identical or analogous circuit components will be indicated by the same identifying symbols in the fig ^¬ ures .

Solution of a square system of equations

Figure 2 refers to a first embodiment of the above described circuit 100 and shows, by an example, a first circuit for solving square systems of equations 200. The first system solving circuit 200 is capable of solving a square system of equations expressable by the following matrix form:

Ax = b (1)

wherein :

A: is a matrix of real coefficients having a di ^¬ mension N x N;

x: is an unknown vector, the length thereof being

N;

b: is a vector of real elements, the length thereof being N.

It is observed that the resistive memories Gij can be configured to take a predetermined value (ex ^¬ cept for an uncertainty margin) inside a range be ^¬ tween a minimum value Gmin and a maximum value Gmax .

Each conductance value Gij of a resistive memory of the matrix 1 is equal or proportional to an element Aij of the matrix A. Particularly, the first circuit for solving square systems of equations 200, shown in Figure 2, refers to elements Aij of the matrix A, which are all positive.

The first system solving circuit 200 is also pro ^¬ vided with a plurality of current generators Ii, each connected, for example, to a first end of the row conductors Li. It is observed that each operational amplifier AOi is connected to the respective row con ^¬ ductor Li at, for example, a second end, opposite to the first end connected to the current generator Ii. According to the example, the resistive memories G±j exhibit a respective terminal connected to a corre ^¬ sponding row conductor at a node n±j, comprised between the first and second ends of the specific row conductor L± .

The current generators I± are configured to inject in the respective row conductors, a current I± (Ii, I2, I3) , the values thereof are selected so that they are equal or proportional to the elements of the known vector b of the matrix system (1) . The currents Ii are constant or pulsed currents.

In the operation, the currents I± supplying the row conductors L± flow to the first input terminal INi (in other words the virtual ground terminal INli) of each operational amplifier AOi. The electric volt ^¬ ages Vi are evaluated between the respective output terminal OUi and ground GR.

From Figure 2, it is possible to write the fol ^¬ lowing equations based on the Ohm and Kirchhoff laws:

Vi G11+V2 G12+V3 Gi3 Ii i G21+V2 G22+V3 G23 = - I2 (2)

i G31+V2 G32+V3 G33 = - I3

The equations (2) can be rewritten in a compact way as in the expression (3) :

ZjVjGij = -Ii (3)

and also, by an algebraic notation, by the fol ^¬ lowing matrix equation (4) :

AV = -I (4)

wherein A is the matrix of the conductances G±j, I is the known vector -b and the vector of the volt ^¬ ages V solves the linear system Ax = b, as already expressed by ( 1 ) .

In the operation, as hereinbefore stated, the values of the conductances G±j are configured and the first circuit 200 is supplied by predetermined cur ^¬ rent values I±.

The operational amplifiers AO± take the respec ^¬ tive first input terminal INi (the inverting termi ^¬ nal) and therefore all the row conductors Li to the ground and consequently the electric voltages Vi, at the output terminals OUi, assume the final values, as expressed by the relationship (4) .

It is observed that selecting high gain opera ^¬ tional amplifiers OAi, enables to very quickly reach the virtual ground, and the voltages also quickly converge to values Vi to be determined. Particularly, for this embodiment and for other embodiments, oper ^¬ ational amplifiers having a nominal gain greater than 105 can be selected.

Measuring the voltage values Vi (specifically Vi, V2, and V3) , reached in the first system solving cir ^¬ cuit 200, enables to obtain elements of the vector x, and consequently to solve the equation system (1) .

The first system solving circuit 200 of Figure 2 refers to a matrix of coefficients A containing only positive elements. By considering a matrix of the coefficients Ai containing both positive and negative elements, the solution of the equation systems

Ai x = b (la)

can be obtained by the embodiment in Figure 3 which shows an example of a second circuit for solving square systems of equations 300, in a specific exam ^¬ ple wherein the matrix of coefficients Ai has a di ^¬ mension N = 3.

The second system solving circuit 300 is configured by observing that the matrix Ai can be expressed as the difference between a first matrix A+ and a second matrix A-, both containing only positive elements : Ai = A+ - A- (5)

Particularly, the second matrix A- contains the absolute values of the negative elements of the ma ^¬ trix Ai .

According to the expression (5), the second system solving circuit 300 comprises a first crosspoint matrix MG+, analogous to the matrix MG of Figure 2, the resistive memories thereof have conductances Gij+ corresponding to the elements of the first matrix A+ and a second crosspoint matrix MG- (structurally anal ^¬ ogous to the first one) , the resistive memories thereof have conductances Gij- corresponding to the elements of the second matrix A- .

The two crosspoint matrixes have the same column conductors C , along each of them it is interposed (between a resistive memory Gij+ and a resistive memory Gij-) a respective inverting device Invl (in the shown example: Invl, Inv2, Inv3) configured to invert the sign of the electric voltage from its inlet to its output and corresponding to the difference indicated by relationship (5) .

The first crosspoint matrix MG+ comprises N = 3 row conductors L± ₊ (Li ₊ , L2 ₊ , L3+) , each of them has an end connected to a respective supply node N± (Ni, N2, N3) connected to a corresponding current generator li.

The second crosspoint matrix MG- of the example comprises N = 3 row conductors L±- (Li-, L2-, L3-) , each of them has an end connected to one of the supply nodes N± (Ni, N ₂ , N ₃ ) .

Particularly, the operational amplifiers OA± of the second system solving circuit 300 operate as hereinbefore described with reference to the virtual ground and have the respective inverting input ter ^¬ minal INli connected to each row conductor L±+ and L±- , particularly, at the supply nodes Ni, N2 and N3, which therefore operate as virtual ground nodes.

Each of the known currents (li, I2 and I3) is split at one of the supply nodes Ni, N2 and N3 into two terms, that is I±+ and I±-:

With reference to the first crosspoint matrix MG+ , it is valid the equation:

∑jVjGij+ = -Ii+ (7)

wherein Vj indicates the voltage vector Vi, V ₂ , V ₃ .

With reference to the second crosspoint matrix MG~, it is valid the following equation:

-ZjVjGij- = -li- (8)

By summing to each other the equations (7) and (8), it is obtained:

∑jVj ( Gij+ - Gij- ) = - Ii+ - Ii- = - I (9) Based on the relationship (5), it is obtained:

( Gij+ - Gij- ) = (A+ - A) = -Ai (10) Therefore, it is possible to rewrite (8) as follows :

(A+ - A- ) x = Ai x = -I = b (11) By changing the currents I± to the values of the known vector b, measuring the voltage Vj enables to solve the system of equations (la) also for a matrix of coefficients Ai , comprising negative and positive elements .

Inverting square matrixes

Moreover, it is observed that the problem solving circuit 100, also in the embodiments in Figure 2 and Figure 3, can be used for inverting square matrixes, in other words for calculating an inverse matrix A ^{- 1} of the matrix A .

The inverse matrix A ^{- 1} meets the relationship:

AA- ¹ = U (12)

wherein U is the identity matrix, whose elements are all null, except for the elements of the diagonal which are equal to 1.

It is observed that for determining the inverse matrix A ^{- 1} , it is required to solve the following systems of equations:

AAi- ¹ = Ui (13)

wherein :

- Ai ^{_ 1} is the i-th column of the inverse matrix A ^~ i

- Ui is the i-th column of the identity matrix U. If the dimension of matrix A is N x N, it is required to solve N systems according to (13) . In this case, the first circuit for solving systems of equations 200 can be configured, with reference to one of the systems of equations (13), so that:

- the crosspoint matrix MG has conductances Gij equal to the values of the matrix A to be inverted;

- the currents I± take the values U± of the i-th column of the identity matrix U.

Under such conditions, in the same way as indi ^¬ cated with reference to Figure 2, the values of volt ^¬ ages Vj represent the column A± ^_1 of the inverse matrix A- ¹ .

By using the mode described with reference to Figure 3, it is possible to invert matrixes contain ^¬ ing also negative and positive elements.

Calculating eigenvectors

As hereinbefore discussed, the circuit 100 can be used for calculating eigenvectors as it is for exam ^¬ ple shown in Figure 4 illustrating a first eigenvec ^¬ tor calculating circuit 400.

The problem to be solved can be expressed by the relationship :

wherein :

- A is a N x N matrix of known elements;

- x is the eigenvector of the matrix A, which is unknown;

- λ is the (scalar) eigenvalue of the matrix A, which is known.

In the first eigenvector calculating circuit 400, the operational amplifiers AOi have a trans-impedance configuration and, particularly, the first input terminal INli (the inverting one "-") is connected to the output terminal OUi of each operational amplifier OAi by a respective feedback resistor having conductance equal to λΘο, wherein Go is a known value and is a reference conductance.

Particularly, each feedback resistor Go (imple- mentable by a resistive memory) is connected between an output terminal OUi and an input node INi (in the example INi, IN2, IN3) respectively connected to the inverting terminal of each operational amplifier OAi. The output terminals OUi of the operational am ^¬ plifiers AOi are connected to the row conductors L± by respective inverters INV1I, INV2 and INV3 for ob ^¬ taining the closed-loop configuration.

The example of Figure 4 refers to the case of N = 3 for a matrix A containing only positive elements.

Currents Ii, I2 and I3 which flow to the input nodes INi of the respective operational amplifier, can be expressed by:

11 = V1G11 + V2G12 + V3G13

12 = V1G21 + V2G22 + V3G23 (15)

13 = V1G31 + V2G32 + V3G33

For a generic value N, the system (15) can be rewritten in the following way:

Ii = ZjVjGij (16) or, by a matrix notation, in the following way:

I = GV (17)

The vector of the current I is transformed by the operational amplifiers OAi having the trans-impedance configuration into a voltage vector V:

The transformation expressed by the relationship (18) is enabled by each operational amplifier OAi operating as a trans-impedance amplifier, and by the respective inverter INVi . Wherein GTIA is the conductance of the feedback resistance :

wherein λ is the eigenvalue, and Go is the refer ^¬ ence conductance.

Therefore, it is possible to write:

GV=Go V (20)

The comparison of the latter with the relationship Αχ=λχ, enables to obtain:

A = G/Go, x = V (21)

Therefore, in the circuit 400 of Figure 4, the conductances Gij of the resistive memories of the crosspoint matrix MG have values equal to the ones of the elements of the matrix A multiplied by the value of the reference conductance Go.

The feedback resistance GTIA is given by the prod ^¬ uct of the eigenvalue λ and of the reference conduct ^¬ ance Go. The measured values of the voltages V± (Vi, V2, and V3) correspond to the requested eigenvector x. It is observed that a complete set of eigenvectors x is obtained by modifying the eigenvalue, in other words the value of the feedback resistance GTIA .

It is observed that a possible application of the first eigenvector calculating circuit 400 is found in the link matrixes of the ranking algorithms (for ex ^¬ ample, for Google) , wherein the eigenvalue expresses the importance score of each page. Therefore, the described solution seems extremely advantageous for accelerating the ranking of Internet pages and generally for analyzing "big data".

The first eigenvector calculating circuit 400 can be also applied for an approximate numeral solution of differential equations. When a differential equa ^¬ tion is transformed in a finite difference equation, it has again the matrix form Ax = λχ . For example, the Schrodinger equation takes such form, wherein A is a semi-diagonal matrix, λ is the eigenvalue of the energy, and x is the solution eigenfunction of the problem.

If the matrix A, of which the eigenvectors are sought, contains both positive and negative elements, the problem can be solved by the second eigenvector calculating circuit 500 shown, by an example, in Figure 5 (still for N = 3) .

The second eigenvector calculating circuit 500 comprises the first crosspoint matrix MG+ and second crosspoint matrix MG- which respectively correspond to the matrixes G+ and G-, containing positive ele ^¬ ments according to the notation: A = (G+ - G-) /Go (22)

The two crosspoint matrixes have the same column conductors C , along each of them it is interposed (between a resistive memory Gij+ and a resistive memory Gij-) a respective inverting device Invj (Inv4, Inv5, Inv6) configured to invert the sign of the electric voltage from the input thereof to the output thereof.

The first crosspoint matrix MG+ comprises N = 3 row conductors L±+ (Li+, L2 ₊ , L3+) , each of them has an end connected to a respective connecting node N± _A (NIA, N2A, N3A) . The second crosspoint matrix M _G - of the example, comprises N = 3 row conductors L±- (Li-, L2-, L3-) , each of them has an end connected to one of the connecting nodes N± _A (NIA, N2A, N3A ) .

Particularly, the operational amplifiers OAi of the second system solving circuit 300 operate as hereinbefore indicated with reference to the virtual ground and have the respective input node INi con ^¬ nected to each row conductor L± ₊ and Li-, particularly, at the contact nodes NIA, N2A, and N3A, and therefore operate as virtual ground nodes.

For the circuit in Figure 5, with reference to the currents I±+ (Ii+, I2+, I3+) , which are in the row conductors of the first crosspoint matrix MG+, the following equations in a compact form are valid:

Ii ₊ = ∑ _D V ₃ G ₁₃₊ (23)

With reference to the currents I±+ (Ii+, I2+, I3+) , flowing in the row conductors of the second crosspoint matrix MG ^~ , the following equations in a compact form are valid:

Ii- = -∑ _D V ₃ G ₁₃ - (24)

The currents I± ₊ and I±- are summed to each other in the respective connecting nodes NIA, 2A, and N3A, according to the Kirchhoff law, for obtaining:

Ii = Ii+ + Ii (25)

Consequently, the current vector I can be ex ^¬ pressed by an algebraic notation as:

I = GV (26)

wherein :

Gij = Gij+ - Gij- (27)

The current vector I is transformed in a voltage vector V by the trans-impedance operational amplifi ^¬ ers OAi and by the respective inverters Invi :

wherein, as for the hereinbefore described cir ^¬ cuit of Figure 4, GTIA is the conductance of the feed ^¬ back resistor:

GTIA = GO (29) By considering the relationship Ax = λχ and comparing it with the relationship (29) , it is obtained:

A = G/Go, x = V (30)

analogously to the equation (21) of the circuit in Figure 4.

It is observed that each operational amplifier AOi of the first eigenvector calculating circuit 400 is an active circuit with its own positive and nega ^¬ tive supplies. In the row and column conductors of the first eigenvector calculating circuit 400, currents/voltages are generated in order to meet the Ohm and Kirchhoff laws, which determine the solution of the problem. An analogous consideration is valid for the circuits of the other shown figures.

Further embodiments

It is observed that the conductance values Gij of all the above described circuits, can be obtained not only by using resistive memories (in other words memristors) but by using, for each value of the re ^¬ quired conductance Gi , a resistive element (prefer ^¬ ably a reconfigurable one) having three or more ter ^¬ minals such as, for example: a field-effect transis ^¬ tor, a floating-gate transistor, a flash memory, a charge-trapping memory. For example, as charge-trapping memory, a device having a metal-oxide-nitride- oxide-semiconductor (MONOS) structure can be used.

For this purpose, Figure 6 shows a first embodi ^¬ ment (200A) of the first circuit for solving square systems of equations 200, hereinbefore described with reference to Figure 2, wherein the conductances Gij are obtained by respective MONOS-type devices Di .

The gate voltages Vg, ij of each device Dij are controlled by respective tension generators (not shown) , and can be partially short-circuited to each other for limiting the number of connections, for example by connecting all the gate terminals of a column, or of a row, to the voltage generator itself. The same circuit in Figure 6 can be used with float ^¬ ing-gate transistors, or with simple field-effect transistors, or with other types of three terminal elements .

Moreover, it is observed that the three terminal resistive element can also comprise a suitable com ^¬ bination of a two terminal element (for example, a memristor) and of a three terminal element. Figure 7 exemplifyingly shows a second embodiment (200B) of the first circuit for solving square systems of equa ^¬ tions 200, wherein each conductance Gij is obtained by suitably configuring a three terminal element com- prising a transistor device TRij and a memristor device MRij .

It is useful to observe that what is described with reference to possible resistive elements useable for obtaining conductance values Gij is still valid for the conductance value GO described with reference to the first eigenvector calculating circuit 400.

Further, it is observed that all the described embodiments can operate by using constant conductance resistive elements, in other words non-reconfigurable elements .

Now it is made reference to the closed-loop con ^¬ figuration of the operational amplifiers OAi in the above described different embodiments. According to a further embodiment, it is also provided that one of the operational amplifiers OAi is in an open-loop configuration wherein the output terminal OUi is not connected to the column conductors Ci .

For example, Figure 8 shows a different embodi ^¬ ment (400B) of the first eigenvector calculating circuit 400, wherein the output terminal OUI of the op ^¬ erational amplifier OAI is not connected to the first column conductor CI and therefore the voltage to be measured VI' (as an element of the sought eigenvector V) is available at a terminal TOU1 driven by the operational amplifier OA1 (by the respective invert ^¬ ing device Invl) . According to the example of Figure 8, the first column CI is supplied by a voltage VI generated by a voltage generator GEN. There is also the possibility the first column conductor CI is open, as an alternative or in addition to what shown in Figure 8, in other points inside the crosspoint matrix MG .

It is observed, as already said, that an appli ^¬ cation of the mathematical problem solving circuit 100 and of the above described embodiments thereof consists of analyzing big data, for example calcu ^¬ lating the rank page of web pages. Other applications comprise the approximate solution of differential equations, such as for example the Schrodinger equa ^¬ tion, and other meteorology, financial, biology prob ^¬ lems, etcetera.

The mathematical problem solving circuit 100 and the above described embodiments thereof deliver ap ^¬ proximate solutions because there are some uncertain ^¬ ties about the configurable values of the resistive memories. Such approximations are acceptable for the greater part of the applications of the circuit it ^¬ self .

The mathematical problem solving circuit 100 and the above described embodiments thereof have the ad ^¬ vantage of being computationally simple: the compu ^¬ tation is done in just one clock without requiring multiplications and additions operations. Therefore, the described circuit operates as an algebraic com ^¬ putational accelerator.

Figure 9 refers to another example of the mathe ^¬ matical problem solving circuit, as a linear regres ^¬ sion accelerator circuit 1000.

The linear regression accelerator circuit 1000 comprises a first circuit 600 (or input circuit) con ^¬ nected to a second circuit 700 (or output circuit) . The first circuit 600 is analogous to the mathemati ^¬ cal problem solving circuit 100 described in Figure 1, expect for the fact that the plurality of analog resistive memories is indicated by reference Xi , instead of Gi . The second input circuit 700 is also analogous to the mathematical problem solving circuit 100 described in Figure 1.

More particularly, the operational amplifiers OAi of the first circuit 600 have a trans-impedance con ^¬ figuration and, specifically, the first input termi ^¬ nal INli (the inverting one "-") is connected to the output terminal OUi of each operational amplifier OAi by a respective feedback resistor having a conduct ^¬ ance equal to Go, wherein Go is a known value and is a reference conductance.

The output terminal OUi of each operational am ^¬ plifier OAi is connected to the second circuit 700 comprising an output crosspoint matrix MGOU. The out ^¬ put crosspoint matrix MGOU has the same dimensions of the crosspoint matrix MG and includes analog resis ^¬ tive memories Xij having values and circuit position identical to the ones of the crosspoint matrix MG of the first circuit 600.

In the example of Figure 9, the crosspoint matrix M _G and the output crosspoint matrix M _Go u are rectan ^¬ gular matrix. Particularly, the number of columns is less (for example, by a unit) than the number of rows.

The output terminal OUi of each operational am ^¬ plifier OAi of the first circuit 600 is connected to a respective row conductor Li of the output crosspoint matrix MGOU. Each column conductor Couj of the output crosspoint matrix M _G0 u is connected to a respective output operational amplifier OAoui . Each output operational amplifier OAoui has a respective non-inverting input terminal "+" connected to a re ^¬ spective column conductor Couj and a respective in ^¬ verting input terminal "-" connected to ground GR. Each output operational amplifier OAoui is pro ^¬ vided with a respective output terminal OUPj con ^¬ nected to a respective column conductor Cj of the crosspoint matrix MG of the first circuit 600. Each output operational amplifier OAoui is connected in the closed-loop circuit 1000.

Moreover, the linear regression acceleration circuit 1000 is supplied, in the example, by voltage generators VGi each connected to a row conductor Li of the crosspoint matrix MG and having voltages -yi at a respective terminal to which the conductance Go is connected. The voltages at the output terminals OUi of each operational amplifier AOi are indicated by reference Vi . The voltages at the output terminals OUPj of the output operational amplifiers OAoui are indicated by the symbols wj (in the example wl and w2 ) .

With reference to the linear regression problem, it is remembered that the regression formalizes and solves the problem of a functional relationship be ^¬ tween measured variables based on sampled date ex ^¬ tracted from an infinite hypothetical population. For example, it m points (input-output pairs are consid ^¬ ered) , each with n input variables:

(Xl, X2,..., Xn) (31) and an output coordinate y.

For an i-th input-output pair, there is the lin- regression relationships

y± = Wo + WlXl ¹ + W2X2 ¹ + ... + Wn n ¹ + Si (32) wherein i covers all the integers from 1 to m. By expressing the whole in a matrix form, it is obtained :

y = X · w + ε (33)

wherein: W is the matrix of the coefficients be determined) , and ε is the approximation error

In order to minimize the approximation error ε, the norm of the vector ε is minimized, therefore is expressed by:

X - w) ^T (y - X -

(34)

The minimum of the norm can be determined by set ^¬ ting to zero the derivative of the expression (34) with respect to the variable w:

ow ow

-- -2X ^T y + 2X ^T Xw = 0

It follows that: X ^T X w = X ^T y (35)

Therefore, the matrix of the coefficients w is given by:

w = (X ^T X) - ^λ Χ ^τ y (36) wherein (X ^T X) ^~1 X ^T is the pseudo-inverse of the matrix X.

Referring again to the linear regression circuit 1000, it is useable for minimizing the quadratic er ^¬ ror (34) , by solving the expression (36) .

Indeed, the output voltages Vi (vector V) of the operational amplifiers OAi (which are made to operate in a trans-impedance configuration) can be expressed by the following relationship:

V = - (X-w - Goy) /Go (37)

wherein X is the matrix of the conductances Xij and Go was defined before.

The output operational amplifiers OAou , analo ^¬ gously to what described with reference to the oper ^¬ ational amplifiers OAi, are configured to connect the respective non-inverting input terminal to the vir ^¬ tual ground, and therefore:

X ^■ V = -X ^■ (X-w - Goy) / Go = 0 (38)

The matrix Go is the unit for the matrix of the conductances X and therefore the expression (38) takes the form: XX-w - Xy = 0 (39)

which is equivalent to the equation (36) .

After correctly dimensioning the circuit 1000 based on known values and once applied the supply voltage by the voltage generators VGi, the circuit 1000 will operate so that the output terminals of the output operational amplifiers OAouj have respective voltages wj representing the sought solution (as from the expression (36)) . Analogously to what was done for the previous embodiments, also for the linear regression accelerator circuit 1000, the computation is performed in only one clock without requiring mul ^¬ tiplying and summing operations. Therefore, the de ^¬ scribed linear regression circuit 1000 operates as an algebraic computational accelerator.

The arrangement of the linear regression accel ^¬ erator circuit 1000 hereinbefore described is also adaptable to be used as a logistic regression accel ^¬ erator. As it is known, the logistic regression is a particular case regarding the cases wherein the y- dependent variable is of a dichotomic type tied to a low value (for example -1) and a high value (for example +1), as all the variables which can only take two values (for example: true or false) . For implementing the logistic regression accelerator, it is provided to modify the layout of the circuit 1000 of Figure 9 with voltage (or current) supply generators representing dichotomic values -1 (for the low value) and 1 (for the high value) . The obtained vector w identifies the coefficient of the equation (32) enabling to linearly separate the two classes of points with high and low y.

Moreover, the linear regression accelerator circuit 1000 is useable for calculating the coefficient wj of a neural network, this prevents to perform the iterative training, for example according to the back-propagation method.

For example, Figure 10 schematically shows a neu ^¬ ral network 800 comprising an inlet layer 801, and a hidden layer 802 and an outlet layer 803.

In the example of Figure 10, the inlet layer com ^¬ prises Ni = 14x14 = 196 neurons, the hidden layer comprises N2 = 784 neurons, and lastly the output or classifying layer comprises N3 = 10 neurons.

The weights of the neural network 800, associated to the input layer 801 can be represented by a matrix W ⁽¹⁾ (for example, W ⁽¹⁾ with dimensions Ni x N2 = 196 x 784) while the weights associated to the hidden layer 804 can be represented by a matrix W ⁽²⁾ (for example, W ⁽²⁾ with dimensions N2 x N3 = 784 x 10) .

The neural network 800 is useable, for example, for acknowledging/classifying the input quantities. According to a particular example, a network as the above described one can be used for acknowledging hand-written numbers (from 0 to 9) . For example, a single digit in the MNIST data-set standard repre ^¬ sented in a matrix of Ni input values (pixel) can be classified by the neural network as a digit from 0 to 9 by suitably training the network 800 for setting the values of the synaptic weights.

It is observed that the operation of training the neural network 800 is a logistic regression opera ^¬ tion, wherein the weights represent the coefficients which better linearly separate the dichotomic output of the neurons.

Consequently, the network can be trained by the linear regression circuit 1000 in Figure 9, operated according to the previously described logistic re ^¬ gression algorithm. For this purpose, the synaptic weights W ⁽¹⁾ between the first neuron layer 801 and the hidden neuron layer 802 are arbitrarily set, for example, according to a weight random distribution. Moreover, it is used a training data-set formed by a determined number M of sets of Ni input values to be presented to the neurons of the first layer 801. The matrix X to be used inside the circuit 1000 will then be formed by output values of the hidden network on different M presentations.

By defining the matrix I containing all the M*N presented input values, the matrix I is for example obtained as :

X = sigmoid (I*W< ¹ > ) (40)

wherein the sigmoid function is a possible non ^¬ linear function associated to a generic neuron. It is observed that the matrix has dimensions M*N2. The vector y± in Figure 9 is on the contrary obtained as a matrix of labels ( + 1 for label "true" and -1 for label "false") , referred to the i-th classification neuron (for i varying from 1 to N3) . The synaptic weights referred to the i-th neuron are therefore obtained by the linear regression:

_Wi <2) = _(X T x) -ΐχτ y (4!)

Therefore, the values of the weights W ⁽²⁾ of the hidden layer 802 can be obtained by using the linear regression accelerator circuit 1000, by observing that the matrix expression (41) is analogous to the above discussed matrix expression (36) . It is ob ^¬ served that the operation must be repeated N3 times, one for each classification neuron. Each operation enables to obtain the N2 synaptic weights referring to the i-th output neuron. In this way, the repetitive operations enable to obtain all the N2 x N3 synaptic weights W ⁽²⁾ between the hidden layer and the output layer .

It is observed that the linear regression accel ^¬ erator 1000 and the above described particular embodiments thereof are very useful in the data science and find possible applications in different scien ^¬ tific and engineering fields such as, for example in: economy, finance, biology, physics, automatic train ^¬ ing, robotics.