Optical multiplier implementing skip connection operation

Abstract

FIELD: optics.SUBSTANCE: invention relates to optical devices that solve problems of machine learning and artificial intelligence. An optical multiplier designed to implement the skip connection operation has N inputs and N outputs, where N≥3, and consists of one programmable interferometer, which has N inputs and N outputs, a set of splitters, consisting of N splitters, and a set of combiners, consisting of N combiners. The input of each splitter of the plurality of splitters is the input of an optical multiplier, one output of each splitter is connected to the input of a programmable interferometer, and the second output of each splitter is connected to one input of a combiner. Each output of the programmable interferometer is connected to one input of the combiner, and the output of each combiner is the output of the programmable interferometer.EFFECT: increase in energy efficiency and computing speed when implementing the skip connection operation.5 cl, 5 dwg

Description

AREA OF TECHNOLOGY

[0001] The invention relates to methods for creating optical devices that perform linear operations of multiplying large matrices by vectors. The invention can be used in classical and quantum optical computing devices, for example, those solving problems of machine learning and artificial intelligence. The invention can be used in the implementation of individual elements of communication and computing networks serving a large number of subscribers and computing nodes; these elements and networks can be both classical and quantum. In addition, the invention can be used to create devices that analyze and synthesize multimode electromagnetic fields. BACKGROUND

[0002] Transmitted or processed information can be encoded into optical signals. Optical devices can perform certain types of operations on optical fields with the speed and energy efficiency of digital electronic circuits. Therefore, the use of optical methods for implementing computational operations can improve the speed and / or energy efficiency of computations. Matrix-vector multiplication is one of the most computationally expensive and frequently used operations in modern computers, particularly in neural networks.

[0003] Optical multipliers are known that perform matrix-vector or matrix-matrix multiplication in analog format, disclosed, for example, in the article by H. Zhou et al., ((Photonics matrix multiplication lights up photonic accelerator and beyond» / / Nature Light: Science & Applications 11, 30 (2022). In optical multipliers, the vectors to be multiplied are encoded into optical field amplitudes, which are fed to the inputs of a multichannel optical circuit. If the transfer matrix of the multichannel optical circuit is set in accordance with the matrix by which the input vectors need to be multiplied, then the multiplication result can be obtained during the passage of the optical fields through the circuit. In this case, the output vectors corresponding to the input vectors multiplied by a given matrix are obtained by measuring the optical fields at the output of the circuit.The operation of multiplying a matrix by vectors encoded in optical fields is performed in a time period determined by the time it takes for the fields to pass through the optical multiplier circuit. This time can be much shorter than the time it takes to perform the multiplication on a traditional digital computer.

[0004] This approach to matrix multiplication is more energy efficient and faster than the implementation of this operation on traditional digital computers.

[0005] Fig. 1 illustrates the execution of vector multiplication by the matrix W in the optical multiplier 1. The vectors to be multiplied x (m) in digital electronic format from the electronic digital computer 2 are sent to the electro-optical converter 3 via electronic lines 4, which converts these digital vectors into optical fields with the corresponding amplitudes. For example, one vector x in electronic format, consisting of N (in) elements x i , is transformed into optical fields 5 with an amplitude vector , coming out of the outputs 6 of the electron-optical converter 3. The amplitudes of the optical fields are equal to the corresponding elements of the multiplied vector: . The electronic digital computer is also connected by electronic lines 7 to the optical multiplier 1 to enable the transfer matrix to be set in accordance with the matrix W of size by which the input vector x must be multiplied. Optical fields 5 are fed to inputs 1a of the optical multiplier 1. As a result of the propagation of optical fields through the optical multiplier, optical fields 8, described by the amplitude vector, emerge from its outputs 1b Amplitude vector has amplitudes that carry the result of the multiplication. Optical fields 8 are fed to inputs 9a of the optoelectronic conversion unit 9, which converts the amplitudes of these fields into electronic digital format. The multiplication results are fed to the electronic digital computer via electronic lines 10.

[0006] The prior art includes methods for creating multichannel programmable interferometers that can implement multiplication by arbitrary unitary matrices. For example, methods for creating universal N-channel programmable interferometers are disclosed in the works of C.W. Clements et al. ((Optimal design for universal multiport interferometers» and M. Reck et al., "Experimental realization of any discrete unitary operator"). Interferometers created using these methods are capable of multiplying N×1 vectors, encoded in the field amplitudes arriving at the interferometer inputs, by arbitrary N×N unitary matrices.

[0007] Fig. 2 illustrates the well-known Clements architecture of multichannel programmable interferometers, which are used to implement an optical multiplier 1 on arbitrary unitary matrices. Fig. 2a and Fig. 2b show the diagrams of 4- and 8-channel programmable interferometers. The interferometers consist of sequentially connected layers of Mach-Zehnder interferometers (MZIs). Fig. 2a shows one MZI layer 11 and one MZI element 12 within this layer.

[0008] In addition, the interferometers have a phase shift layer 13, which is not related to the MZI layers, and is located at the outputs of the interferometers. In Fig. 2a, one phase shift element 14 is also indicated, related to the phase shift layer 13. To perform the multiplication of vectors by the matrix set in the 4-channel programmable interferometer, the input field amplitudes 15, encoding the vector, are fed to its inputs 16. The field propagates through a 4-channel interferometer and the field amplitudes emerge from the output channels 17 18, which are equal to the vector ', multiplied by the transfer matrix of a 4-channel interferometer. Fig. 2c shows a diagram of a single MZI 12, which makes up multi-channel interferometers. The MZI consists of two programmable phase shifters 14a and 14b and two balanced dividers 19.

[0009] When implementing digital neural networks, operations of multiplying vectors or matrices by special matrices are often used. One of the important types of multiplication by special matrices are the transformations performed by linear layers of the skip connection type, which are described, for example, in the patent of Wai Kit Ricky Fok et al. ((Parallel residual neural network architecture and system and method for training a residual neural network" (US 11645536 B2, 09.05.2023). Fig. 3 shows a diagram explaining the operation of the skip connection layer used in digital neural networks. The layer multiplies the input vector of numbers x by a special matrix The action of this multiplication can be described by the following relation (written for one vector): where y is the vector obtained at the output of this layer, W is the matrix trained during the training stage of the entire digital neural network, and α and β are parameters whose values ​​determine the contribution of the multiplication by the matrix W and the unmultiplied input vector x to y. The cases with α ≠ 0 and β = 0 correspond to a traditional linear layer, which multiplies input vectors by the matrix W. In the general case, the layer transformation is specified by the matrix W and the parameters α and β.

[0010] Skip connection layers are used in neural networks to address several issues that arise during network training. The primary cause of this problem when training deep models is the difficulty of propagating the error signal back through a large number of layers. The deeper the network, the more the signal is attenuated, losing important information. This phenomenon is called the vanishing gradient problem. Skip connection layers, also known as residual connections, help overcome the vanishing gradient problem when training deep neural networks consisting of hundreds or more layers. Such layers are widely used in deep learning, for example, in ImageNet classifiers (ResNet, DenseNet), object detection models (Faster RCNN, YOLOvX), image segmentation (U-Net, SegFormer), and transformers. Thus, skip connections help maintain stable performance of a deep network by providing feedforward signal paths.Furthermore, skip connections simplify the backpropagation process and make training more efficient. Thus, the use of skip connections has become standard in modern neural network architectures, especially when working with large data sets and high levels of image detail.

[0011] Using a skip connection layer solves the following problems of deep neural networks (He K., Zhang X., Ren S., Sun J., “Deep Residual Learning for Image Recognition”)} / / Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR'16), 2016; S. Xie, et al., “ResiDual: Transformer with Dual Residual Connections”) / / arxiv:2304.14802 (2023)):• The problem of vanishing gradients. As the number of layers in a neural network increases, gradient signals begin to vanish during backpropagation. This leads to a slowdown in training and makes it difficult to optimize the weights of the lower layers. Skip connections create additional direct paths for gradients, allowing the useful signal to be transmitted directly between remote layers, thus facilitating the training of deep architectures.• Improving model quality. Research shows that deep learning without skip connections often faces the problem of "degradation": increasing the number of layers leads to a decrease in accuracy.Skip connections address this effect, maintaining stability and improving classification performance.• Simplifying training of deep models. With fast error paths, training large models (e.g., ResNets with hundreds of layers) becomes possible. Without such connections, deep structures could completely stall progress.• Improved resilience to overfitting. These mechanisms provide greater flexibility to the network, allowing it to better adapt to different data conditions and avoid overreliance on specific patterns that are characteristic of overfitted models.• Ability to use larger data sets. By improving training efficiency, larger amounts of data become more readily available to the network, improving the generalization ability of the model.

[0012] When implementing skip connection layers with given matrix W and coefficients α and β in optical multipliers, it is necessary to reprogram the transfer matrix of the multiplier in accordance with the expression (where I is the identity matrix). Before reprogramming the optical multiplier's transfer matrix, the following computational operations must be performed on a digital electronic computer: multiplying the matrix W by the number α and adding the result of the previous operation to the diagonal matrix β I. Only after this can the optical computer's transfer matrix be set / reprogrammed.

[0013] This leads to a decrease in the energy efficiency of using the optical multiplier and an increase in the execution time. ESSENCE OF THE INVENTION

[0014] The claimed invention solves the problem of the need to perform certain computational operations—matrix addition and matrix multiplication—on a digital computer when implementing skip connection layer transformation. Performing these operations on a digital computer consumes the computer's computing resources and does not fully utilize the capabilities of optical accelerators. Furthermore, the need to use a digital computer in addition to an optical multiplier leads to time delays, which reduces the computation speed.

[0015] The technical result of the invention is to increase energy efficiency and computation speed when implementing the skip connection operation by performing all computational operations necessary for implementing the skip connection layer on optical computers without the use of digital electronic processors. In the proposed invention, the skip connection layer transformation is implemented entirely on an optical accelerator, without intermediate operations performed on a digital electronic computer.

[0016] The claimed technical result is achieved by implementing an optical multiplier designed to implement a skip connection operation, having N inputs and N outputs, wherein N≥3, and consisting of one programmable interferometer, which has N inputs and N outputs, a plurality of splitters, consisting of N splitters, and a plurality of combiners, consisting of N combiners; the input of each splitter from the plurality of splitters is the input of the optical multiplier, one output of each splitter is connected to the input of the programmable interferometer, and the second output of each splitter is connected to one input of the combiner; each output of the programmable interferometer is connected to one input of the combiner, and the output of each combiner is the output of the programmable interferometer.

[0017] In one particular example of implementation, each splitter of a plurality of splitters divides the amplitude supplied to it into two amplitudes of equal power.

[0018] In another particular example of implementation, each combiner from a plurality of combiners combines the amplitudes arriving at it into one amplitude, the value of which is proportional to the sum of the incoming amplitudes.

[0019] In another particular example of implementation, at least one splitter of a plurality of splitters is configured to change the division ratio.

[0020] In another particular example of implementation, at least one combiner of a plurality of combiners is configured to change the combining coefficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Fig. 1 explains the implementation of matrix multiplication operations on analog optical multipliers.

[0022] Fig. 2 shows programmable Clements interferometers: a 4-channel interferometer (Fig. 2a) and an 8-channel interferometer (Fig. 2b). Fig. 2c shows a Mach-Zehnder interferometer used in Clements interferometers.

[0023] Fig. 3 explains the linear operation of multiplying the vector x in a skip connection layer used in digital implementations of neural networks.

[0024] Fig. 4 illustrates an optical circuit implementing the proposed optical implementation of a skip connection layer.

[0025] Fig. 5 illustrates the implementation of a variable splitter (Fig. 5a) and combiner (Fig. 5b), in which the division and combination coefficients can be varied. IMPLEMENTATION OF THE INVENTION

[0026] The invention concerns converters of electromagnetic fields, which can belong to different wavelength ranges - from radio to optical.

[0027] For a more unambiguous understanding of the essence of the claimed solution, the main terms and definitions used within the framework of the present invention are presented below.

[0028] A transformation channel is any degree of freedom of an electromagnetic field that can be uniquely associated with an independent signal characterized by a complex amplitude. For clarity, we will separately identify the following degrees of freedom that can serve as a transformation channel: 1) A spatial degree of freedom of a field can be a mode of a single-mode waveguide through which an electromagnetic signal can propagate. In this case, several single-mode waveguides form multiple channels. A spatial channel can also be a single spatial mode of a multimode waveguide or a single mode of free space. In this case, several spatial modes of a single multimode waveguide or several modes of free space form multiple independent channels. 2) A frequency degree of freedom of a field is represented by individual lines of the frequency spectrum. A set of non-overlapping spectral lines forms a set of channels.3) A temporal degree of freedom is a given time interval, characterized by the timing of its beginning and end. The presence of an electromagnetic signal in this time interval is interpreted as the presence of a signal in this channel. Thus, a set of time intervals forms a set of channels. Depending on which of the designated degrees of freedom of the electromagnetic field is used to encode information, we speak of spatial, frequency, and temporal channels.

[0029] A linear transformation is a transformation whose action can be described by the linear law of matrix multiplication by vectors. In general, multiplication operations are described by rectangular matrices W of dimensions , which multiply vectors (columns) dimensions Linear transformations are described by multiplying these matrices by these vectors, which can be written as follows: Where - elements of the vector being multiplied - elements of the vector obtained after a linear transformation Coefficients are elements of the matrix W, which defines a specific linear transformation. Expression (1.2) can be represented in matrix form: Where - columns composed of the amplitudes of the signals at the input and output of the transformation, respectively. In the general case, vectors and the matrices W are complex.

[0030] Often the optimal way to use linear multipliers, especially when training neural networks or inference, is to multiply a set of K different vectors in one go. (called a batch). The proposed invention operates regardless of the number of vectors K being multiplied simultaneously. In the following description, it will be assumed that only one vector is multiplied at a time. (i.e. K=1) on a given matrix W, which does not reduce generality.

[0031] Optical multiplication refers to the operation of multiplying matrices by vectors using linear optical circuits, such as programmable optical converters. The optical multipliers discussed in this application are analog computers.

[0032] A linear optical circuit is an optical circuit that transforms the amplitudes of the optical field arriving at it. in output amplitudes in accordance with expression (1.3).

[0033] The transformation transfer matrix or simply the transformation matrix is ​​the matrix W that relates the column of amplitudes at the transformation output to the column at its input (see expression (1.3)).

[0034] A multichannel interferometer (multichannel converter) is any device with at least three inputs or three outputs that linearly converts electromagnetic signals received at its inputs. The number of inputs may differ from the number of outputs. A special case of a multichannel interferometer is the N-channel interferometer (N-channel converter), which has an equal number of inputs and outputs. A multichannel interferometer is an example of a linear-optical circuit.

[0035] The transfer matrix of a linear-optical circuit (interferometer, converter) is the transfer matrix of the transformation implemented by this circuit.

[0036] A programmable interferometer (converter) is an interferometer that is designed with the ability to change its transfer matrix after its manufacture.

[0037] A dual-channel divider, or simply a divider, is an optical element with two inputs and two outputs that converts two optical fields arriving at its inputs, resulting in two converted amplitudes. The divider element is described by a 2×2 transfer matrix: where r is the divider's reflection coefficient, τ is the divider's transmission coefficient. The divider's reflection and transmission coefficients are related to each other by the ratio: r 2 +τ 2 =1, thus the transformation of one element of the divider is described by one parameter, for example, its reflection coefficient r.

[0038] A balanced divider is a divider in which the absolute values ​​of all four elements of its transfer matrix are equal

[0039] A phase shifter (phase shift element or phase shift element) is a single-channel element that adds a given phase θ to the amplitude of the field input to it. As a result, the field at the output of this element has an amplitude The present invention concerns interferometers that consist of static two-channel dividers and programmable phase shift elements.

[0040] A Mach-Zehnder interferometer (MZI) is a two-channel element consisting of two series-connected static divider elements with a balanced division ratio and one variable / programmable phase shift element located in one of the channels between the dividers. The Mach-Zehnder interferometer acts as a tunable divider, in which the reflection coefficient r and transmission coefficient τ can be varied by varying the phase shift. It serves as the core component of the conversion units in known methods for implementing universal and non-universal converters.

[0041] A splitter is an optical element with one input and two outputs that divides the power of the signals arriving at it in a given proportion. When a field amplitude a is received at its input, the two field amplitudes at the output are equal to τa and ra, where the coefficients τ and r obey the relation r + τ = 1. The transformation of the splitter can be specified by a single coefficient - r or τ; in what follows, the coefficient r will be used (and the notation In the particular case of a balanced splitter, when the amplitude of the field enters the input of the splitter, the amplitudes of the fields leaving it are equal to (LA. Krutov et al., ((Optimized low-loss integrated photonics silicon-nitride Y-branch splitter" / / AIP Conf. Proc. 2241, 020027 (2020)).

[0042] A combiner is an optical element with two inputs and one output that performs summation of the incoming field amplitudes. In general, when the amplitudes arrive at its inputs, the field coming out of it has an amplitude , where the coefficients τ and r obey the relation r 2 +τ 2 =1. The transformation of the combiner can be specified by one coefficient - r or τ; in what follows, the coefficient r will be used (and the notation introduced In the particular case of a balanced combiner, the field leaving it has an amplitude A splitter can act as a combiner, in which two outputs are used as combiner inputs, and one input is used as combiner output.

[0043] A two-channel divider, described by transfer matrix (1.4) with a given reflection coefficient r, can function as a splitter or combiner with a corresponding coefficient r. When performing the splitter function, only one input of the divider is used, to which the optical signal is fed. When performing the combiner function, only one of the two outputs of the divider is used, from which the amplitude of the converted incoming optical signals is output.

[0044] The present invention relates to the implementation of skip connection layers entirely in analog optical devices that can perform multiplication operations with lower energy consumption and higher speed. Fig. 4 illustrates a circuit diagram of an optical multiplier 1 that implements a linear skip connection layer. The circuit consists of one programmable optical interferometer 20, a plurality of splitters 21 and a plurality of combiners 22 (in Fig. 4, a splitter 21a and a combiner 22a are also separately designated). To implement a skip connection layer acting on vectors x of size N, the plurality of splitters 21 and the plurality of combiners must consist of N splitters and N combiners, respectively. In this case, the programmable interferometer 20 must have N inputs and N outputs. When implementing vector multiplication in the proposed optical multiplier, each input vector x is converted into field amplitudes 5, which are fed to the inputs of splitters 21. In this case, one splitter with index n (1≤n≤N) from the set of splitters divides the field amplitude fed to it in accordance with a given coefficient As a result, at the outputs of the set of splitters, two amplitude vectors 23a and 23b are obtained (also designated . The vector of amplitudes b1 is fed to a set of inputs 24 of the programmable multiplier 20. When passing through the programmable optical multiplier, the field amplitudes that encode the results of multiplying the vector of amplitudes b1 by the matrix W set in the programmable multiplier 20 emerge from its outputs 25. Fields with amplitudes 26 (b2), emerging from the outputs 25 of the optical multiplier, and amplitudes 23b are combined on a set of N combiners. In this case, the optical fields from one output of the programmable multiplier 20 and one field from one output of the splitter (one field from the amplitude vector) are fed to a pair of inputs of each combiner. ). The combiner with index k (1≤k≤N) from the set of combiners sums the field amplitudes arriving at it in accordance with its coefficient Outputs 27a and 27b of combiners are the outputs of the optical multiplier. The result of the multiplication (amplitude vector) ) is obtained in a subset of N outputs, with only one output of each divider out of 4 being used. As an example, Fig. 4 illustrates the case where the optical fields carrying information about the result of the multiplication come out of outputs 27a, while outputs 27b are not used.

[0045] The claimed optical multiplier, designed to implement the skip connection operation, contains a plurality of splitters that divide the input signal into two components: • one component is processed in a programmable optical interferometer (multiplication by the matrix W). • the second component passes directly, without changes (the second output of each splitter is connected to one input of the combiner).

[0046] This division is fundamentally important, since it is the basis of the skip connection concept, where part of the signal is transmitted directly, bypassing intermediate transformations.

[0047] In the circuit shown in Fig. 4, it is possible to implement matrix multiplication In the case where the set of splitters and the set of combiners consist of elements with the same parameters (i.e. for all and k), and the transfer matrix of the programmable interferometer is set to W, the transfer matrix of the proposed optical multiplier circuit has the following form: Where - coefficients depending on the transformations of splitters and combiners, and Dependences of the coefficients η and ζ on the parameters of the splitters and unifiers allows to implement multiplication in the optical circuit by ' with arbitrary coefficients α and β, taking values ​​from 0 to 1. In the expression for there is a scaling factor s. The adjustment of the scaling factor s consists of dividing by s the results of multiplication by the matrix s W (SC) is computationally inexpensive. The magnitude of the multiplier s is irrelevant for performing multiplication by a given matrix W (SC) because when obtaining the result of multiplication by the matrix s W (SC) the desired result of multiplication by the matrix W (SC)(without the factor s) can be easily obtained by dividing the resulting vectors or matrices by the factor s.

[0048] To demonstrate the possibility of setting arbitrary relationships between the coefficients α and β in the proposed optical circuit implementing the skip connection layer, we will consider two special cases: 1) Splitters with a balanced coefficient value In this case, the ratio of the coefficients η and ζ from the transfer matrix (1.5) takes the form: . Thus, the range of values ​​of the combiner parameters allows you to change your attitude in the range from 0 to ∞. Therefore, in (1.5) up to a scaling factor, arbitrary values ​​of α and β can be realized.2) Combiners with a balanced coefficient value In this case, the ratio of the coefficients η and ζ from the transfer matrix (1.5) takes the form: Thus, the range of values ​​of the parameters of the splitters allows you to change your attitude in the range from 0 to ∞ - similar to the previous example. Therefore, in (1.5), up to a scaling factor, it is possible to implement the multiplication operation in the skip connection layer with arbitrary values ​​of α and β.

[0049] The splitters 21 and combiners 22 can be implemented as static elements, such as Y-splitters and Y-combiners. Such optical elements are standard for integrated optical circuits (see, for example, the work of Z. Lin and W. Shi, ((Broadband, low-loss silicon photonic Y-junction with an arbitrary power splitting ratio» / / Optics Express, vol. 27, No. 10, pp. 14338-14343 (2019)). The coefficients The coefficients characterizing the transformation of these elements are specified at the design and manufacturing stage and cannot be changed after production. It is possible to implement programmable optical elements that function as splitters and combiners. In such elements, the coefficients , which define α and β in the proposed optical multiplier, can be reprogrammed after the optical circuit is manufactured. Fig. 5 illustrates the optical circuits of programmable splitters (a) and combiners (b), which represent a programmable Mach-Zehnder interferometer.

[0050] For a more unambiguous understanding of the essence of the claimed solution, the main terms and definitions used within the framework of the present invention are presented below.

[0051] In case of performing the splitter function (see Fig. 5a), the input signal 28 with amplitude c (in) is fed to only one input 29 of the MZI. From the two outputs 30 of the MZI, two amplitudes 31a and 31b (amplitudes . In this case, the power ratio between these amplitudes is determined by the phase value ψ set on the phase shift element 14a. In the case of performing the combiner function (see Fig. 5b), the input amplitudes 32a and 32b (amplitudes ) are fed to two inputs of the MCI 33. Two field amplitudes 34a and 34b (amplitudes) emerge from the two outputs. However, for the combiner functionality, only one amplitude 34a is of interest (amplitude , coming out of one output 35. Field amplitude is a weighted sum of the input amplitudes And , and the ratio of deposits is set by the phase value ψ set on the phase shift element.

[0052] The submitted application materials disclose preferred examples of the implementation of the technical solution and should not be interpreted as limiting other, particular examples of its implementation that do not go beyond the scope of the requested legal protection, which are obvious to specialists in the relevant field of technology.

Claims

1. An optical multiplier designed to implement a skip connection operation, having N inputs and N outputs, where N≥3, and consisting of one programmable interferometer that has N inputs and N outputs, a set of splitters consisting of N splitters, and a set of combiners consisting of N combiners; the input of each splitter of the plurality of splitters is the input of an optical multiplier, one output of each splitter is connected to the input of a programmable interferometer, and the second output of each splitter is connected to one input of a combiner; each output of the programmable interferometer is connected to one input of the combiner, and the output of each combiner is the output of the programmable interferometer.

2. An optical multiplier according to claim 1, in which each splitter of the plurality of splitters divides the amplitude supplied to it into two amplitudes of equal power.

3. An optical multiplier according to claim 1, in which each combiner of the plurality of combiners combines the amplitudes received by it into one amplitude, the value of which is proportional to the sum of the incoming amplitudes.

4. An optical multiplier according to claim 1, in which at least one splitter from the plurality of splitters is configured to change the division ratio.

5. An optical multiplier according to claim 1, wherein at least one combiner from a plurality of combiners is configured to change the combining coefficient.

Images