A method for judging the high-frequency working ability of laser chips based on neural networks

Abstract

The present invention provides a method for judging the high-frequency working ability of a laser chip based on a neural network. The method specifically comprises: obtaining chip sample data from multiple wafers and preprocessing the sample data to obtain a sample data set F; using an optimized AHP algorithm to process the sample data set F to obtain a new sample data set F ’ ; Take the sample data set F ’ Construct a Transformer network model for input; set the relevant parameters of the transformer network model, execute the network training process until the model converges; obtain the corresponding parameters and positions of some chips on the wafer to be evaluated, perform data preprocessing, and input the trained transformer network to obtain the evaluation results. The above method uses chip sample data to construct a high-frequency working ability evaluation model for semiconductor laser chips in various regions on the training wafer, integrates the influence of multiple parameters on the evaluation results, and uses AHP to process the data to mine the relationship between the parameters, thereby reducing the production cost while ensuring the detection accuracy.

Description

Technical Field

[0001] The present invention relates to the field of optical communication semiconductor laser chips and neural network technology, and in particular to a method for judging the high-frequency working ability of a laser chip based on a neural network. Background Art

[0002] With the increasing demand for cloud computing, AI computing power, 5G transmission, and other infrastructure construction, the demand for data centers, base stations, and other infrastructure is also increasing, and the optical communication market represented by 400G / 800G and higher speeds has also entered a period of rapid development. As the core light source in optical communications, there is an increasing demand for semiconductor laser chips with high speed, high bandwidth, and other high-frequency working capabilities.

[0003] Generally speaking, the high-frequency working ability of semiconductor laser chips is mainly determined by epitaxial design and resonant cavity size design, but on the same wafer, due to the influence of manufacturing process, the overall performance of each laser chip will be different and fluctuate. In the ordinary LIV DC test process, chips with qualified DC performance can be screened out, but due to the current high-frequency test methods and test efficiency, it is impossible to test the high-frequency characteristics of each laser chip on the wafer. According to the current method, the high-frequency characteristics of the laser chip need to be mounted on a specific substrate and fixture to test the high-frequency characteristics of the laser chip. Once the test is completed, it is difficult to remove the laser chip from the specific substrate and fixture and put it into production, which will cause the risk of appearance damage. Therefore, whether it is the sample preparation or the test process of the test piece, it takes a long time and cannot be applied in mass production. In addition, since there is no way to test and monitor the high-frequency performance of all laser chips on the wafer, there is a greater risk in the later stage. Summary of the invention

[0004] The purpose of the present invention is to provide a method for judging the high-frequency working ability of a laser chip based on a neural network, so as to solve the problems that the chip to be tested is easily damaged and the sample preparation and testing process of the chip to be tested are time-consuming in the current testing method. The specific technical solution is as follows:

[0005] A method for judging the high-frequency working capability of a laser chip based on a neural network, the method comprising the following steps:

[0006] S1, obtain chip sample data and preprocess it to obtain a sample data set F;

[0007] S2. Use the optimized AHP algorithm to process the sample data set F to obtain a new sample data set F';

[0008] S3, using the sample data set F' as input data to build a Transformer network model;

[0009] S4, set the relevant parameters of the transformer network model and execute the network training process until the model converges;

[0010] S5. Obtain the parameters and position of the chip to be evaluated, perform data preprocessing, and input the data into the trained transformer network to obtain the evaluation results.

[0011] The step S1 specifically includes:

[0012] S11, obtaining sample data of chip samples from multiple wafers; dividing each wafer into multiple regions, each region containing multiple chip samples, and obtaining multiple index numbers of each chip sample through detection;

[0013] S12, weighting the obtained chip sample data; assigning different weights to chip samples in different areas of the wafer, multiplying the index number parameter of each chip sample by the weight of the area where it is located, to obtain a weighted sample data set F.

[0014] Furthermore, the multiple indicators in step S11 include quantum well structure and material, resonant cavity size, metal layer thickness and metal layer appearance, threshold, slope efficiency, resistance, voltage, test temperature, amplitude of relaxation oscillation peak, frequency corresponding to relaxation oscillation peak and -3dB cutoff bandwidth.

[0015] Furthermore, in step S12, the weight of the chip samples in the middle area of ​​the wafer is higher than the weight of the chip samples in the peripheral area.

[0016] Furthermore, the step S1 also includes: performing conventional process evaluation on chip samples of the wafer and classifying them into four grades: I, II, III, and IV according to high-frequency working capabilities, where I is the best grade and IV is the worst grade.

[0017] Furthermore, the step S2 specifically includes:

[0018] S21, using the AHP algorithm driven by optimizing the correlation function difference to obtain the index weight vector W;

[0019] S22. Process the sample data set F in combination with the weight vector W to obtain a new sample data set F'.

[0020] Furthermore, the step S21 specifically includes: first constructing a lower side correlation function, constructing a "sample-index" information matrix through the lower side correlation function, performing center normalization processing on the information matrix to obtain a positive matrix, obtaining an orthogonal matrix through positive matrix calculation, and calculating the weight vector W of the orthogonal matrix.

[0021] Furthermore, the step S22 specifically includes: performing weighted processing on the sample data set F in combination with the weight vector W to obtain a new data sample set F'.

[0022] Furthermore, the step S3 specifically includes: designing the architecture of the Transformer network model, using the new sample data set F' as input data of the Transformer network, and using the high-frequency working capability level of the semiconductor laser chip as the expected output label of the Transformer network.

[0023] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute one or more of the steps in the above-mentioned method for determining the high-frequency working capability of a laser chip based on a neural network when loading the computer program.

[0024] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is loaded by a processor to execute one or more steps in the above-mentioned method for determining the high-frequency working capability of a laser chip based on a neural network.

[0025] The present invention provides a method for judging the high-frequency working ability of a laser chip based on a neural network, which has the following beneficial effects:

[0026] The present invention provides a method for judging the high-frequency working ability of a laser chip based on a neural network. A high-frequency working ability evaluation model of semiconductor laser chips in various regions on a training wafer is constructed using fewer samples, and the influence of multiple parameters on the evaluation results is integrated. The AHP is used to process the data to mine the relationship between the parameters, and the evaluation of the number of semiconductor laser chips for subsequent sampling and evaluation can also be achieved. This greatly reduces the impact of destructive chip detection on production, and reduces production costs while ensuring detection accuracy. In addition, in order to ensure the stability and effectiveness of the model, continuous sampling and iteration are performed during production for continuous training to adapt to subsequent production. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of a method for judging the high-frequency working capability of a laser chip based on a neural network provided by the present invention.

[0028] Figure 2 This is a schematic diagram of wafer area division in a method for determining the high-frequency working capability of a laser chip based on a neural network provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the accompanying drawings provided by the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. According to the following description, the advantages and features of the present invention will be more clear. It should be noted that the accompanying drawings are all in a very simplified form and are not in precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

[0030] In the description of the present invention, unless otherwise clearly stipulated and limited, the terms "connection" and "connected" should be interpreted in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0031] In the description of the present invention, terms such as "up", "down", "left", "right", "front", "back", "center", "horizontal", "vertical", "top", "bottom", "inside" and "outside" and other orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings and are only for the convenience of description and simplification of operation, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0032] Example 1

[0033] This embodiment provides a method for determining the high-frequency working capability of a laser chip based on a neural network. Figure 1 As shown, the method comprises the following steps:

[0034] S1, obtain chip sample data and preprocess it to obtain a sample data set F;

[0035] S2. Use the optimized AHP algorithm to process the sample data set F to obtain a new sample data set F';

[0036] S3, using the sample data set F' as input data to build a Transformer network model;

[0037] S4, set the relevant parameters of the transformer network model and execute the network training process until the model converges;

[0038] S5. Obtain the parameters and position of the chip to be evaluated, perform data preprocessing, and input the data into the trained transformer network to obtain the evaluation results.

[0039] The method for judging the high-frequency working ability of a laser chip based on a neural network provided by the present invention randomly selects some chip samples in each area on a wafer, tests their small signal bandwidth and obtains sample data, uses a neural network model to build a model, adds key parameters and weight indicators, to estimate and judge the performance and distribution of the high-frequency working ability of all laser chips in each area on the wafer, so as to judge which areas of the wafer chips can meet the indicator requirements in terms of high-frequency working ability.

[0040] The method for determining the high-frequency working capability of a laser chip based on a neural network provided by the present invention is described in detail below.

[0041] Step S1 specifically includes:

[0042] S11, obtaining sample data of chip samples from multiple wafers;

[0043] Specifically, each wafer is divided into multiple areas, each area contains multiple chip samples, and multiple index parameters of each chip sample are obtained through detection. Each wafer can contain n chip samples, n = 1, 2, 3...200, each chip sample has m index numbers, m = 1, 2, 3...11, which can be calibrated to z1, z2, z3...z11.

[0044] In this embodiment, see Figure 2 As shown in the figure, two concentric circles with a radius of twice the relationship are used to divide the wafer into 8 areas, namely A, B, C, D, E, F, G, and H. Each area has 25 chips, so each wafer has 200 chip samples. After small signal broadband detection, the corresponding 2000 index parameters can be obtained. If 100 wafers are sampled, there are a total of 20,000 chip samples that can be used for subsequent network training.

[0045] The main factors that affect the bandwidth of laser chips include: (1) internal structure and materials: quantum well structure and material selection. However, considering that they are on the same wafer, it is assumed that the internal structure of all laser chips on the wafer is the same; (2) external dimensions: resonant cavity dimensions (length, width, height), metal layer appearance (thickness of the metal layer covering the resonant cavity); (3) performance: threshold, slope efficiency (output optical power at the same bias current point), resistance and voltage; (4) test temperature: the bandwidth of the chip will change at different temperatures. Therefore, the sample data for each chip sample needs to include:

[0046] ① Quantum well structure and materials, considering the same wafer, it is assumed that the internal structure of all laser chips on the wafer is the same. The material design index is calibrated as z1 and the weight is a1.

[0047] ② Actual measured resonant cavity dimensions: Obtain the length, width, and height of the resonant cavity, because the volume of the resonant cavity (length x height x width) will affect all performance of the laser chip. The measured volume index is calibrated as z2 and the weight is a2.

[0048] ③ The actual measured thickness of the metal layer covering the resonant cavity and the appearance of the metal layer. The thickness of the metal layer is related to the heat dissipation of the chip. The thicker the metal layer, the better the heat dissipation, the smaller the self-heating of the chip, and the higher the electro-optical conversion efficiency. In the actual process, due to the non-absolute flatness of the wafer surface, as well as the influence of edge effects and process uniformity, the actual metal thickness covering the resonant cavity of each chip on the wafer is different. The appearance of the metal layer is also greatly affected by the process, including whether there is metal overflowing the design area, whether there are metal pits, whether the design range is covered, and whether there is metal collapse (gold is relatively soft) to block part of the light-emitting cavity surface. The measured metal layer thickness and appearance index number is calibrated as z3, and the weight is a3.

[0049] ④ Actual measured threshold: One of the key parameters of the laser chip, greatly affected by laser design and process factors. The measured threshold index number is calibrated as z4 and the weight is a4.

[0050] ⑤ Actual measured slope efficiency: It is the output optical power at the same bias current point, one of the key parameters of the laser chip. The higher the slope efficiency, the higher the electro-optical conversion efficiency of the laser. It is greatly affected by the laser design and process. The measured slope efficiency index is calibrated as z5 and the weight is a5.

[0051] ⑥ Actual measured resistance: One of the key parameters of the laser chip, it is greatly affected by the laser chip design and process. The smaller the resistance, the smaller the internal heat of the laser chip. The measured resistance index is calibrated as z6 and the weight is a6.

[0052] ⑦ Actual measured voltage: One of the key parameters of the laser chip, it is greatly affected by the laser chip design and process factors, and is related to resistance. The smaller the working voltage at the same bias current point, the lower the power consumption of the laser chip. The measured voltage index is calibrated as z7 and the weight is a7.

[0053] ⑧ Actual measured test temperature: The laser chip is a temperature-sensitive chip, and the ambient temperature will affect all electro-optical performance indicators of the laser, including wavelength, bandwidth, etc. The actual test temperature index number measured is calibrated as z8, and the weight is a8.

[0054] ⑨The actual measured amplitude of the relaxation oscillation peak: the frequency point where the frequency response changes by 3dB relative to the low-frequency response. If the amplitude of the relaxation oscillation peak increases by more than 3dB, then the frequency point exceeding 3dB is the actual bandwidth of the laser chip. The actual bandwidth value of the laser is lower than the designed bandwidth value, indicating that the actual bandwidth of the chip is low and the high-frequency working ability is poor. The amplitude index of the measured relaxation oscillation peak is calibrated as z9, with a weight of a9.

[0055] ⑩The frequency size corresponding to the relaxation oscillation peak actually measured: If the amplitude of the relaxation oscillation peak increases by more than 3dB, then the frequency point exceeding 3dB is the actual bandwidth of the laser chip. The actual bandwidth value is lower than the designed bandwidth value, indicating that the actual bandwidth of the chip is low and the high-frequency working ability is poor. The frequency size index corresponding to the measured relaxation oscillation peak is calibrated as z10, with a weight of a10.

[0056] Actual measured -3dB cutoff bandwidth: -3dB cutoff bandwidth data directly reflects the high-frequency working ability of the laser chip itself. The higher the bandwidth, the more it can meet the signal transmission quality requirements in the optical module. The measured -3dB cutoff bandwidth index is calibrated as z11 and the weight is a11.

[0057] In the above sample data, the index number z1 is the inherent parameter after chip design, and the index numbers z2-z11 are the parameters of sampling measurement, so the inherent parameter z1 is not considered in the subsequent analysis.

[0058] S12, performing weighted processing on the acquired chip sample data;

[0059] Specifically, different weights should be given to chip samples in different regions when obtaining chip samples. Since the wafer needs to undergo multiple processes such as photolithography, etching, and chemical grinding, the wafer will accumulate more stress in the outer circle and is very easy to be damaged during the production process, affecting the overall yield. In addition, due to the influence of the stress itself, the wafer actually has a certain degree of bending and warping. The more central the wafer is, the flatter it is, and the higher the bending and warping angles are at the edges. In the actual photolithography process, this will cause differences in chips in different areas, so in theory, the chips closer to the center of the wafer will have better performance.

[0060] In this embodiment, the weights of regions E, F, G, and H should be smaller than those of regions A, B, C, and D. Based on empirical parameters, the parameters of regions A, B, C, and D can be assigned a higher weight of a=0.6, while the weight assigned to regions E, F, G, and H is b=0.4.

[0061] The index number parameter of each chip sample is multiplied by the weight of the region where it is located to obtain the weighted sample data set F.

[0062] Step S1 also includes: performing conventional process evaluation on chip samples of the wafer and classifying them into grades, and classifying the chips into four grades I, II, III, and IV according to high-frequency working capabilities, where I is the best and IV is the worst.

[0063] Step S2 specifically includes:

[0064] S21, using the AHP algorithm driven by optimizing the correlation function difference to obtain the index weight vector W;

[0065] Specifically, the analytic hierarchy process (AHP) is a multi-objective method that divides complex problems into multiple hierarchical structures. This method can combine qualitative and quantitative methods. When there are many parameters and there are certain relationships, the AHP method can be used to achieve indicator data fusion. The AHP algorithm driven by the correlation function difference can obtain the corresponding fusion input data. The method for optimizing the AHP algorithm driven by the correlation function difference is as follows:

[0066] Construct the lower correlation function: convert the qualitative evaluation into quantitative values, and consider the relationship between the evaluation value and the benchmark value of the indicator. This function has different expressions in different intervals, which can better quantify the degree of correlation between the evaluation value of a certain indicator and the benchmark value of the indicator. By defining the correlation function to quantify the relative importance of factors, the function uses different calculation methods according to different value ranges, so as to achieve a detailed description of the relationship between factors. The lower correlation function k ij (x) are as follows:

[0067]

[0068] Where, i represents the number of chip samples, i=n=1,2,3...200; j represents the index number, j=m=z2,z3,...z11; x ij represents the evaluation value of the jth indicator number of the i-th chip sample; xj(1), xj(2), xj(3), and xj(4) are four benchmark values ​​related to the j-th indicator, which define different correlation intervals of the evaluation values.

[0069] Construct the "sample-indicator" information matrix based on the lower correlation function:

[0070]

[0071] Perform central normalization on the information matrix: adjust the value of the correlation function to make it more reasonably reflect the relative importance of factors. The information matrix is ​​center normalized, where

[0072] Get the positive matrix Through the center normalization process, a positive matrix is ​​constructed Provides a basis for subsequent weight calculation. ij =k i ' j -t j +ε, where t j =min(k ij )<0, a positive matrix can be obtained for

[0073]

[0074] Calculate the orthogonal matrix: through the positive matrix The orthogonal matrix COR is calculated, which helps to further optimize the weight determination process. The orthogonal matrix COR is

[0075]

[0076] Calculate the weight vector W of the symmetric matrix COR

[0077] W=(w1,w2,…,w n ) T

[0078] Among them, Wi is the entropy weight of each indicator, and the weight vector W is the set of Wi. The weight vector W is used for comprehensive evaluation of multiple indicators. The calculation method of entropy weight Wi is as follows

[0079] w i =o i / o

[0080]

[0081] Among them, Oij is the relationship between the ith sample and the jth indicator, which is used to indicate the importance of the ith sample relative to the jth indicator. Oi is obtained by summing Oij, which indicates the comprehensive score or importance of the ith sample on all relevant indicators.

[0082] S22. Process the sample data set F in combination with the weight vector W to obtain a new sample data set F'.

[0083] Specifically, the fusion input data is obtained by combining the weights W. The purpose is to weight the original data by combining the weights of different indicators to obtain a new data sample set F' for more accurate analysis and decision-making. The data sample set F' is as follows

[0084]

[0085] Step S3 specifically includes: designing the architecture of the Transformer network model, including the attention mechanism, encoder-decoder structure, etc., using the weighted sample data set F' as the input data of the Transformer network, and using the high-frequency working capability level of the semiconductor laser chip as the expected output label of the Transformer network.

[0086] Step S4 specifically includes: setting transformer network related parameters and executing the network training process until the model converges.

[0087] Specifically, common parameter settings in the Transformer model include: embedding dimension: specifies the dimension of the input word vector, which is used to convert discrete data into a vector of fixed size so that the neural network can process it; number of encoder layers and decoder layers: specifies the number of Transformer layers in the encoder and decoder. Deeper layers can improve the representation ability of the model, but also increase the training and inference time; number of attention heads: specifies the number of multi-head attention in each attention mechanism. A larger number of heads can improve the model's ability to pay attention to different positions and semantics; feedforward neural network hidden layer dimension: specifies the dimension of the feedforward neural network hidden layer. A larger dimension can improve the model's nonlinear modeling ability. Dropout probability: specifies the probability of randomly discarding neurons during training. Appropriate dropout can reduce overfitting, while too high a dropout may lead to information loss. Learning rate: specifies the learning rate of the model during training. A smaller learning rate can make the model more stable, but the training time may be longer. A larger learning rate can speed up the training, but may lead to unstable model. Batch size: specifies the number of samples input for each training. A larger batch size can increase the training speed, but it will also increase memory consumption. Maximum sequence length: specifies the maximum length of the input sequence. Sequences exceeding this length will be truncated or otherwise processed.

[0088] Step S5 specifically includes: obtaining the corresponding parameters and location areas of some chips on the wafer to be evaluated, and after data preprocessing, inputting them into the trained transformer network to obtain the corresponding wafer high-frequency working ability evaluation results.

[0089] Example 2

[0090] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute one or more steps of the method for determining the high-frequency working capability of a laser chip based on a neural network in Embodiment 1 when loading the computer program.

[0091] Example 3

[0092] This embodiment provides a computer-readable storage medium on which a computer program is stored. The computer program is loaded by a processor to execute one or more steps in the method of determining the high-frequency working capability of a laser chip based on a neural network in Embodiment 1.

[0093] Those skilled in the art should understand that the present invention can be implemented in many other specific forms without departing from the spirit and scope of the present invention. Based on the embodiments of the present invention, any changes and modifications made by ordinary technicians in the field of the present invention according to the above disclosure are within the scope of protection of the claims.

Claims

1. A method for judging the high-frequency working ability of a laser chip based on a neural network, characterized in that: The method comprises the following steps: S1, obtain chip sample data and preprocess it to obtain a sample data set F; S2. Use the optimized AHP algorithm to process the sample data set F to obtain a new sample data set F'; S3, using the sample data set F' as input data to build a Transformer network model; S4, set the relevant parameters of the transformer network model and execute the network training process until the model converges; S5. Obtain the parameters and position of the chip to be evaluated, perform data preprocessing, and input the data into the trained transformer network to obtain the evaluation result; The step S1 specifically includes: S11, obtaining sample data of chip samples from multiple wafers; dividing each wafer into multiple regions, each region containing multiple chip samples, and obtaining multiple index numbers of each chip sample through detection; S12, weighting the obtained chip sample data; assigning different weights to chip samples in different areas of the wafer, multiplying the index number parameter of each chip sample by the weight of the area where it is located, to obtain a weighted sample data set F.

2. The method for determining the high frequency working capability of a laser chip based on a neural network according to claim 1, characterized in that: The multiple indicators in step S11 include quantum well structure and material, resonant cavity size, metal layer thickness and metal layer appearance, threshold, slope efficiency, resistance, voltage, test temperature, amplitude of relaxation oscillation peak, frequency corresponding to relaxation oscillation peak and -3dB cutoff bandwidth.

3. The method for determining the high frequency working capability of a laser chip based on a neural network according to claim 1, characterized in that: In step S12, the weight of the chip samples in the middle area of ​​the wafer is higher than the weight of the chip samples in the peripheral area.

4. The method for determining the high frequency working capability of a laser chip based on a neural network according to any one of claims 1 to 3, characterized in that: The step S1 also includes: performing conventional process evaluation on the chip samples of the wafer and classifying them into grades, and classifying the chips into I There are four levels: I, II, III, and IV, among which I is the best level and IV is the worst level.

5. The method for determining the high frequency working capability of a laser chip based on a neural network according to claim 1, characterized in that: The step S2 specifically includes: S21, using the AHP algorithm driven by optimizing the correlation function difference to obtain the index weight vector W; S22. Process the sample data set F in combination with the weight vector W to obtain a new sample data set F'.

6. The method for determining the high frequency working capability of a laser chip according to a neural network according to claim 5, characterized in that: The step S21 specifically includes: first constructing a lower correlation function, constructing a "sample-index" information matrix through the lower correlation function, performing center normalization processing on the information matrix to obtain a positive matrix, obtaining an orthogonal matrix through positive matrix calculation, and calculating the weight vector W of the orthogonal matrix.

7. The method for determining the high frequency working capability of a laser chip according to a neural network according to claim 6, characterized in that: The step S22 specifically includes: performing weighted processing on the sample data set F in combination with the weight vector W to obtain a new data sample set F'.

8. The method for determining the high frequency working capability of a laser chip according to a neural network according to claim 6, characterized in that: The step S3 specifically includes: designing a Transformer network model, using the new sample data set F' as input data of the Transformer network, and using the high-frequency working capability level of the semiconductor laser chip as the expected output label of the Transformer network.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute one or more steps of the method for determining the high-frequency working capability of a laser chip based on a neural network as described in any one of claims 1 to 8 when the computer program is loaded.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program is loaded by a processor to execute one or more steps in the method for determining the high-frequency working capability of a laser chip based on a neural network as described in any one of claims 1 to 8.

Images