BSIM-BULK model parameter extraction method based on neural network

Abstract

The invention belongs to the technical field of parameter extraction, and discloses a BSIM-BULK model parameter extraction method based on a neural network, and the method comprises the following steps: 1, automatically and intelligently completing the parameter extraction through machine learning; by combining the advantages of physical modeling and a data-driven model, physical constraints and priori knowledge are introduced in a parameter extraction process; and 2, constructing a parameter extraction model by adopting a CNN architecture: converting one-dimensional characteristic data into a two-dimensional matrix, combining multiple groups of device characteristics into multi-channel input by utilizing the advantage of the CNN on two-dimensional characteristic extraction, and enabling the CNN to process the multi-dimensional characteristics at the same time. According to the method, the CNN neural network is used for replacing an ANN neural network for fitting, so that the local fitting precision is improved; the CNN significantly reduces the number of parameters through the convolution kernel, and the convolution and pooling operations process the local region, thereby improving the calculation efficiency, enhancing the generalization ability of the model, and improving the robustness of the model.

Description

Technical Field

[0001] The invention belongs to the technical field of parameter extraction, and in particular relates to a BSIM-BULK model parameter extraction method based on neural network. Background Art

[0002] As semiconductor technology develops towards deep submicron and nanometer levels, the BSIM-BULK model provides a reliable basis for device modeling. The BSIM-BULK model is a standard compact model for deep submicron CMOS devices. Compared with other BSIM models (such as BSIM3, BSIM4, and BSIM-CMG), the BSIM-BULK model provides higher modeling accuracy for key physical phenomena (including short channel effects and subthreshold characteristics), so it is particularly suitable for advanced planar CMOS technology.

[0003] For the parameter extraction of the BSIM-BULK model, the traditional parameter extraction method first processes the data, then builds a parameter extraction template, manually analyzes the curve characteristics, selects the characteristic parameter range, and extracts specific parameters through fitting tools. Then the extracted parameters are substituted into the model, and the simulation and experimental results are compared for further adjustment. However, manual analysis and fitting take a lot of time and are inefficient, and the quality of parameter extraction is highly dependent on the experience level of engineers, which may lead to insufficient consistency and reliability. When dealing with high-dimensional and complex nonlinear problems, the problem of insufficient efficiency and accuracy is exposed. The neural network-based parameter extraction method provides a better solution for the parameter extraction of complex models through its efficient nonlinear fitting ability, automation characteristics, adaptive ability and powerful data processing performance. An article proposes to use the ANN architecture to realize the parameter extraction of the BSIM-BULK model. Compared with the manual extraction of model parameters, each parameter in this method is repeatedly adjusted until an acceptable error is reached. This method can directly generate a set of parameters without iteration. The steps of the method are as follows:

[0004] 1. Obtain a large amount of initial data sets through simulation and testing, and divide the data into current voltage (IV) data and capacitance voltage (CV) data. Convert the discrete S parameters into Y parameters, calculate Cgg.s (simulation results) and Cgg.m (test results), extract Ids, Vgs, Vds (simulation and test results), and calculate Gm.s, Gds.s (simulation results) and Gm.m, Gds.m (test results) based on the results.

[0005] 2. Parameter value normalization

[0006] The normalization is performed in the following way:

[0007]

[0008] Among them, θ min and θ max are the minimum and maximum values ​​of the parameters respectively. The normalized parameters are limited to between 0 and 1, allowing the algorithm to process these parameters more effectively.

[0009] 3. Adopt ANN architecture. The input layer inputs the current, voltage, and capacitance of the device; the hidden layer is used to capture the nonlinear relationship in the data, and the activation function is ReLU; the output layer outputs the target parameters of the HEMT model (such as threshold voltage, saturation rate, etc.). A set of CV parameters that Cgg.s is closest to Cgg.m is obtained, and a set of IV parameters that IdVd.s, IdVg.s, Gm.s, Gds.s and IdVd.m, IdVg.m, Gm.m, Gds.m are obtained.

[0010] 4. Train the neural network. Divide the initial data into 80% of the data as the training set and 20% of the data as the test set. Save the trained model and use the data in the test set as the input and output of the model to verify the effectiveness of the model training. Train the neural network. Divide the initial data into 80% of the data as the training set and 20% of the data as the test set. Save the trained model and use the data in the test set as the input and output of the model to verify the effectiveness of the model training. By adjusting the weights of each optimization objective (IdVd, IdVg, Gds, Gm, Cgg), a set of model parameters with the best overall effect can be obtained.

[0011] 5. Save the output parameters of the trained neural network, bring the parameter combination back to the original model netlist for verification, and compare it with the measured data to verify the fitting accuracy of the parameter combination.

[0012] However, the above method has the following technical disadvantages:

[0013] 1. The input of the neural network is very long and has problems such as sequential dependency;

[0014] 2. The fully connected structure of ANN requires weight calculation for all input features, which has low computational efficiency. Summary of the invention

[0015] The purpose of the present invention is to provide a BSIM-BULK model parameter extraction method based on neural network to solve the above technical problems.

[0016] In order to solve the above technical problems, the specific technical solution of a BSIM-BULK model parameter extraction method based on a neural network of the present invention is as follows:

[0017] A BSIM-BULK model parameter extraction method based on neural network includes the following steps: Step 1: Automatically and intelligently complete parameter extraction through machine learning; by combining the advantages of physical modeling and data-driven model, physical constraints and prior knowledge are introduced in the parameter extraction process;

[0018] Step 2: Use the CNN architecture to build a parameter extraction model: convert the one-dimensional characteristic data into a two-dimensional matrix, use the advantage of CNN in extracting two-dimensional characteristics, combine multiple groups of device characteristics into multi-channel inputs, and let CNN process multi-dimensional characteristics simultaneously.

[0019] Furthermore, the step 1 comprises the following steps:

[0020] Step 1.1: Introduce data-driven approach;

[0021] Step 1.2: Hybrid modeling approach and automated workflow.

[0022] Furthermore, the step 1.1 specifically includes:

[0023] CNN is used to automatically extract features from device characteristic curves, replacing manual analysis. The core parameters of the BSIM-BULK model are predicted directly from characteristic data through training models, and the generalization ability of the model is improved through data amplification. Furthermore, the step 1.2 specifically includes:

[0024] Use the physical model to provide initial parameter estimates, and then use the neural network to fit the residuals to improve accuracy; add physical consistency as a constraint in the trained model to ensure that the prediction results conform to physical laws; and implement an integrated automated process from data collection to parameter extraction through programming to achieve normalized preprocessing of characteristic data, and subsequently combine with the optimizer to transform the parameter extraction process into an optimization problem, and use genetic algorithms, particle swarm optimization or Bayesian optimization to automatically fit parameters.

[0025] Furthermore, the step 2 includes the following specific steps:

[0026] Step 2.1: Obtain a large amount of simulation data by calling automatic simulation;

[0027] Step 2.2: Data processing: After obtaining a large amount of data through the netlist, the data is divided into DC data and S parameter data, and the data is processed to obtain the required C gg Value, g ds and g m , and normalize the obtained data; Step 2.3: Implementation of CNN architecture: CNN captures the local correlation and spatial structure of data through convolution kernels, and processes the superposition of IV curves under multiple bias conditions from two-dimensional characteristic data in model parameter extraction.

[0028] Furthermore, the step 2.1 includes the following specific steps:

[0029] Step 2.1.1: Get the netlist of the BSIM-BULK model;

[0030] First determine the circuit topology of the BSIM-BULK model netlist. The netlist clearly lists the connection relationship of the components in the circuit. Then determine the complete netlist based on actual parameter adjustment experience. After obtaining the netlist, use it for automated simulation.

[0031] Step 2.1.2: Call the automated simulation to determine the model parameters to be modified; by automatically modifying the model parameters, mobilize the automated simulation to generate a large amount of simulation experimental data;

[0032] Given a certain range of variation for the parameter to be adjusted, different IV data and CV data are obtained. It is difficult to obtain these data through actual testing, but the required amount of data can be easily obtained through automatic simulation, and then the data can be used as training data and test data for the neural network.

[0033] Furthermore, the step 2.2 includes the following specific steps:

[0034] First, convert the discrete S parameters into Y parameters, and then use the formula:

[0035]

[0036] Calculate C gg.s Simulation results and C gg.m Test results, where f is the frequency, Im(Y 11 ) is (Y 11 ); then extract I ds ,V gs ,V ds Simulation and test results, by formula:

[0037]

[0038] Calculate g m.s ,g ds.s Simulation results and g m.m ,g ds.m Test results.

[0039] And normalization is performed in the following way:

[0040]

[0041] Among them, θ min and θ maxare the minimum and maximum values ​​of the parameters respectively. The normalized parameters are limited to between 0 and 1, so that the algorithm can process these parameters more effectively.

[0042] Furthermore, the step 2.3 includes the following specific steps:

[0043] The basic structure of CNN consists of input layer, convolution layer, pooling layer, fully connected layer and output layer. The convolution operation in convolutional neural network can be regarded as the inner product operation of input sample and convolution kernel. Each element of convolution kernel corresponds to a weight coefficient and a bias, which is similar to the neurons of a feedforward neural network. Each neuron in the convolution layer is connected to multiple neurons in the previous layer in the area with similar positions. The size of the area depends on the size of the convolution kernel, which is called "receptive field". Its meaning is similar to the receptive field of visual cortical cells. When the convolution kernel is working, it will regularly scan the input features, perform matrix element multiplication and sum the input features in the receptive field and superimpose the bias:

[0044]

[0045] Where b is the deviation, Z l and Z l+1 Represents the convolution input and output of the l+1th layer, also known as the feature map;

[0046] After the first convolution layer performs convolution operation on the input sample, the feature map can be obtained. The same convolution kernel is used to perform convolution operation on each input sample in the convolution layer. In the second and subsequent convolution layers, the feature map of the previous layer is used as input data and the convolution operation is also performed.

[0047] The function of the pooling layer is to reduce the size of the feature map generated by the convolutional layer, select a region, and obtain a new feature map based on the feature map of the region. This process is called a pooling operation, and its general representation is:

[0048]

[0049] Where the step size is s0, p is a pre-specified parameter. When p = 1, the pooling takes the average value in the pooling area, which is called mean pooling; when p → ∞, the pooling takes the maximum value in the area, which is called maximum pooling;

[0050] The fully connected layer is the same as the multi-layer perceptron. It first calculates the activation value and then calculates the output value of each unit through the activation function. The activation function is selected as ReLU, which replaces all negative values ​​in the feature map with 0 and keeps the positive values ​​unchanged.

[0051] Change, that is:

[0052] f(x)=max(0,x)

[0053] The output layer takes different forms depending on the task.

[0054] The BSIM-BULK model parameter extraction method based on a neural network of the present invention has the following advantages:

[0055] The present invention proposes a method for parameter extraction through a neural network, which is particularly suitable for parameter extraction of a BSIM-BULK model. The local fitting accuracy is improved by introducing CNN. The present invention extracts parameters of the BSIM-BUIK model for the first time; a CNN neural network is used to replace an ANN neural network for fitting, thereby improving the local fitting accuracy; CNN significantly reduces the number of parameters through a convolution kernel, and convolution and pooling operations are used to process local areas, thereby improving computational efficiency, enhancing the generalization ability of the model, and improving the robustness of the model.

[0056] Experimental results show that by combining the powerful nonlinear fitting ability of neural networks and the physical properties of the BSIM-BULK model, fast and accurate extraction of device model parameters can be achieved, making it suitable for modeling needs of complex semiconductor devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the CNN structure of the present invention;

[0058] Figure 2 It is a characteristic curve diagram of the present invention verified in ICCAP. DETAILED DESCRIPTION

[0059] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of a BSIM-BULK model parameter extraction method based on a neural network in conjunction with the accompanying drawings.

[0060] like Figure 1 As shown, a BSIM-BULK model parameter extraction method based on a neural network of the present invention comprises the following steps:

[0061] Step 1: Automatically and intelligently complete parameter extraction through machine learning; by combining the advantages of physical modeling and data-driven models, physical constraints and prior knowledge are introduced in the parameter extraction process.

[0062] Step 1.1: Introduce data-driven approach;

[0063] The present invention changes the traditional parameter extraction process through machine learning technology, and completes parameter extraction automatically and intelligently. The specific improvement is to use CNN to automatically extract features from device characteristic curves (such as IV and CV curves) instead of manual analysis. The core parameters of the BSIM-BULK model can be predicted directly from the characteristic data through the training model, and the generalization ability of the model can be improved through data amplification to solve the problem of insufficient data.

[0064] Step 1.2: Hybrid modeling approach and automated workflow;

[0065] The present invention introduces physical constraints and prior knowledge in the parameter extraction process by combining the advantages of physical modeling and data-driven models. The specific improvement is to use the physical model to provide initial parameter estimates, and then use a neural network to fit the residuals to improve accuracy. Physical consistency is added as a constraint in the trained model to ensure that the prediction results conform to physical laws. In addition, an integrated automated process from data acquisition to parameter extraction is implemented through programming, and normalized preprocessing of characteristic data is achieved. It can then be combined with an optimizer to transform the parameter extraction process into an optimization problem, and automatically fit the parameters using genetic algorithms, particle swarm optimization, or Bayesian optimization.

[0066] Step 2: Build a parameter extraction model using CNN architecture

[0067] The present invention converts one-dimensional characteristic data (such as IV curves and CV curves) into a two-dimensional matrix. By taking advantage of CNN's ability to extract two-dimensional characteristics, multiple sets of device characteristics can be combined into multi-channel inputs, allowing CNN to process multi-dimensional characteristics simultaneously. For data such as IV and CV curves, CNN can identify important patterns such as turning points, linear regions, and nonlinear regions.

[0068] Step 2.1: Get a large amount of simulation data by calling automatic simulation.

[0069] Step 2.1.1: Get the netlist of the BSIM-BULK model. During the parameter adjustment process, the netlist provides a standardized input form, which facilitates quick parameter adjustment and observation of its impact on circuit performance.

[0070] In this embodiment, the circuit topology of the netlist is first determined, and the connection relationship between the components in the circuit is clearly listed in the netlist. Then, the complete netlist is determined based on actual parameter adjustment experience. After the netlist is obtained, it is used for automated simulation.

[0071] Step 2.1.2: Call the automated simulation and determine the model parameters to be modified. Traditional model parameter adjustment is to obtain results through manual simulation. This method automatically modifies the model parameters and mobilizes the automated simulation to generate a large amount of simulation experimental data.

[0072] In this embodiment, the following table shows the model parameters to be modified. Given a certain range of variation of the parameters to be adjusted, different IV data and CV data can be obtained. The data required for this experiment is as high as 100,000. It is difficult to obtain these data through actual testing, but the required number of data can be easily obtained through automatic simulation, and then the data can be used as training data and test data for the neural network.

[0073] IVParameters Description NFACTOR Subthreshold swing factor U0 Low field mobility ETAMOB Effective electric field parameters UA Phonon / surface roughness scattering parameters EU Phonon / surface roughness scattering parameters. RDW Zero bias drain spread resistance unit width VSAT Saturation rate PTWG Correction factor for rate saturation in forward mode PSATX Correction parameters for PTWG effect

[0074]

[0075]

[0076] Step 2.2: Data processing. After obtaining a large amount of data through the netlist, the data is divided into DC data and S parameter data, and the data is processed to obtain the required C gg Value, g ds and g m , and normalize the obtained data.

[0077] In this embodiment, the discrete S parameters are first converted into Y parameters, and then the formula is:

[0078]

[0079] Calculate C gg.s (Simulation results) and C gg.m (Test results), where f is the frequency, Im(Y 11 ) is (Y 11 )’s imaginary part.

[0080] Re-extraction I ds ,V gs ,V ds (Simulation and test results), by the formula:

[0081]

[0082] Calculate g m.s ,g ds.s (Simulation results) and g m.m ,g ds.m (Test results).

[0083] And normalization is performed in the following way:

[0084]

[0085] Among them, θ min and θ maxare the minimum and maximum values ​​of the parameters respectively. The normalized parameters are limited to between 0 and 1, allowing the algorithm to process these parameters more effectively.

[0086] Step 2.3: Implementation of CNN architecture. CNN can capture the local correlation and spatial structure of data through convolution kernels. In model parameter extraction, it can process the superposition of IV curves under multiple bias conditions from two-dimensional feature data, and has higher applicability in parameter extraction tasks.

[0087] In this embodiment, the basic structure of CNN is as follows: Figure 1 As shown in the figure, it consists of an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer. The convolution operation in a convolutional neural network can be regarded as the inner product operation of the input sample and the convolution kernel. Each element that makes up the convolution kernel corresponds to a weight coefficient and a bias, similar to a neuron in a feedforward neural network. Each neuron in the convolution layer is connected to multiple neurons in a similar area in the previous layer. The size of the area depends on the size of the convolution kernel, which is called the "receptive field", and its meaning is analogous to the receptive field of visual cortical cells. When the convolution kernel is working, it will regularly scan the input features, perform matrix element multiplication on the input features within the receptive field, and superimpose the bias:

[0088]

[0089] Where b is the deviation, Z l and Z l+1 Represents the convolution input and output of the l+1th layer, also known as a feature map.

[0090] After the first convolution layer performs convolution operation on the input sample, the feature map can be obtained. In the convolution layer, the same convolution kernel is used to perform convolution operation on each input sample. In the second and subsequent convolution layers, the feature map of the previous layer is used as input data and the convolution operation is also performed.

[0091] The role of the pooling layer is to reduce the size of the feature map generated by the convolutional layer. Selecting a region and obtaining a new feature map based on the feature map of the region is called a pooling operation. Its general representation is:

[0092]

[0093] Where the step size is s0, and p is a pre-specified parameter. When p = 1, the pooling takes the average value in the pooling area, which is called mean pooling; when p → ∞, the pooling takes the maximum value in the area, which is called maximum pooling. Mean pooling and maximum pooling are pooling methods that have been used for a long time in the design of convolutional neural networks. The pooling operation reduces the dimension of the feature map, making the feature representation robust to the position changes of the input data.

[0094] The fully connected layer is the same as the multi-layer perceptron. It also first calculates the activation value and then calculates the output value of each unit through the activation function. The activation function selected by the present invention is ReLU. It replaces all negative values ​​in the feature map with 0 and keeps the positive values ​​unchanged. That is:

[0095] f(x)=max(0,x)

[0096] The output layer can take different forms depending on the task.

[0097] This implementation provides a BSIM-BULK model parameter extraction method based on neural network.

[0098] Parameter extraction through neural network greatly shortens the parameter extraction time. Figure 2 In order to put the parameters extracted by the neural network back into the characteristic curve diagram verified in ICCAP in the original model netlist, the use of CNN can give full play to its advantages in automatic feature extraction, spatial perception, multimodal processing, etc., and is particularly suitable for parameter extraction tasks of complex physical models such as BSIM-BULK. Combining physical priors, data enhancement, network optimization and result interpretability design in the application can further improve model performance and realize innovative applications.

[0099] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A BSIM-BULK model parameter extraction method based on neural network, characterized in that: The steps include: Step 1: Automatically and intelligently extract parameters through machine learning; by combining the advantages of physical modeling and data-driven models, physical constraints and prior knowledge are introduced in the parameter extraction process; Step 2: Use the CNN architecture to build a parameter extraction model: convert the one-dimensional characteristic data into a two-dimensional matrix, use the advantage of CNN in extracting two-dimensional characteristics, combine multiple groups of device characteristics into multi-channel inputs, and let CNN process multi-dimensional characteristics simultaneously.

2. The BSIM-BULK model parameter extraction method based on neural network according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Introduce data-driven approach; Step 1.2: Hybrid modeling approach and automated workflow.

3. The BSIM-BULK model parameter extraction method based on neural network according to claim 2 is characterized in that: The step 1.1 specifically includes: CNN is used to automatically extract features from device characteristic curves instead of manual analysis. The core parameters of the BSIM-BULK model can be predicted directly from the characteristic data through training models, and the generalization ability of the model can be improved through data amplification.

4. The BSIM-BULK model parameter extraction method based on neural network according to claim 2, characterized in that: The step 1.2 specifically includes: Use the physical model to provide initial parameter estimates, and then use the neural network to fit the residuals to improve accuracy; add physical consistency as a constraint in the trained model to ensure that the prediction results conform to physical laws; and implement an integrated automated process from data collection to parameter extraction through programming to achieve normalized preprocessing of characteristic data, and subsequently combine with the optimizer to transform the parameter extraction process into an optimization problem, and use genetic algorithms, particle swarm optimization or Bayesian optimization to automatically fit parameters.

5. The BSIM-BULK model parameter extraction method based on neural network according to claim 1, characterized in that: The step 2 comprises the following specific steps: Step 2.1: Obtain a large amount of simulation data by calling automatic simulation; Step 2.2: Data processing: After obtaining a large amount of data through the netlist, the data is divided into DC data and S parameter data, and the data is processed to obtain the required C gg Value, g ds and g m , and normalize the obtained data; Step 2.3: Implementation of CNN architecture: CNN captures the local correlation and spatial structure of data through convolution kernels, and processes the superposition of IV curves under multiple bias conditions from two-dimensional characteristic data in model parameter extraction.

6. The BSIM-BULK model parameter extraction method based on neural network according to claim 5, characterized in that: The step 2.1 includes the following specific steps: Step 2.1.1: Get the netlist of the BSIM-BULK model; First determine the circuit topology of the BSIM-BULK model netlist. The netlist clearly lists the connection relationship of the components in the circuit. Then determine the complete netlist based on actual parameter adjustment experience. After obtaining the netlist, use it for automated simulation. Step 2.1.2: Call the automated simulation and determine the model parameters to be modified; By automatically modifying model parameters, automated simulation is mobilized to generate large quantities of simulation experimental data; Given a certain range of variation for the parameter to be adjusted, different IV data and CV data are obtained. It is difficult to obtain these data through actual testing, but the required amount of data can be easily obtained through automatic simulation, and then the data can be used as training data and test data for the neural network.

7. The BSIM-BULK model parameter extraction method based on neural network according to claim 5, characterized in that: The step 2.2 includes the following specific steps: First, convert the discrete S parameters into Y parameters, and then use the formula: Calculate C gg.s Simulation results and C gg.m Test results, where f is the frequency, Im(Y 11 ) is (Y 11 ) Re-extraction I ds ,V gs ,V ds Simulation and test results, by formula: Calculate g m.s ,g ds.s Simulation results and g m.m ,g ds.m Test results. And normalize it in the following way: Among them, θ min and θ max are the minimum and maximum values ​​of the parameters respectively. The normalized parameters are limited to between 0 and 1, so that the algorithm can process these parameters more effectively.

8. The method for extracting parameters of the BSIM-BULK model based on a neural network according to claim 5, characterized in that: The step 2.3 includes the following specific steps: The basic structure of CNN consists of input layer, convolution layer, pooling layer, fully connected layer and output layer. The convolution operation in convolutional neural network can be regarded as the inner product operation of input sample and convolution kernel. Each element of convolution kernel corresponds to a weight coefficient and a bias, which is similar to the neurons of a feedforward neural network. Each neuron in the convolution layer is connected to multiple neurons in the previous layer in a similar area. The size of the area depends on the size of the convolution kernel, which is called "receptive field". Its meaning is similar to the receptive field of visual cortical cells. When the convolution kernel is working, it will regularly scan the input features, perform matrix element multiplication and sum the input features in the receptive field and superimpose the bias: Where b is the deviation, Z l and Z l+1 Represents the convolution input and output of the l+1th layer, also known as the feature map; After the first convolution layer performs convolution operation on the input sample, the feature map can be obtained. The same convolution kernel is used to perform convolution operation on each input sample in the convolution layer. In the second and subsequent convolution layers, the feature map of the previous layer is used as input data and the convolution operation is also performed. The function of the pooling layer is to reduce the size of the feature map generated by the convolutional layer, select a region, and obtain a new feature map based on the feature map of the region. This process is called a pooling operation, which is expressed as: Where the step size is s0, p is a pre-specified parameter. When p = 1, the pooling takes the average value in the pooling area, which is called mean pooling; when p → ∞, the pooling takes the maximum value in the area, which is called maximum pooling; The fully connected layer is the same as the multi-layer perceptron. It first calculates the activation value, and then calculates the output value of each unit through the activation function. The activation function is selected as ReLU, which replaces all negative values ​​in the feature map with 0 and keeps the positive values ​​unchanged, that is: f(x)=max(0,x) The output layer takes different forms depending on the task.

Images