Integrated network training method and device based on modular self-coding

By constructing equal variable constraints in ensemble learning, the homogeneity problem between basic models is solved, the generalization ability and robustness of ensemble learning are improved, and functional differentiation and specialization are realized .

CN119026708AActive Publication Date: 2024-11-26BEIJING ACAD OF ARTIFICIAL INTELLLIGENCE
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410940660.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-13
Publication Date
2024-11-26
Estimated Expiration
2044-07-13

AI Technical Summary

Technical Problem

The homogeneity problem between basic models in integrated learning leads to poor integration results, which is even worse than that of a single basic learner, making it difficult to effectively make up for errors and amplify the impact of errors.

Method used

By constructing isovariant constraints, the convolutional network autoencoder is trained to enhance the correlation and independence within the module, and the feature modules with functional differentiation are used for integrated learning, reducing the homogeneity problem between basic models.

Benefits of technology

It improves the generalization ability, robustness and integration effect of integrated learning, promotes the specialization of network module functions, reduces homogeneity problems, and enhances the generalization ability and robustness of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026708A_ABST
    Figure CN119026708A_ABST
Patent Text Reader

Abstract

The invention discloses a training method and device of an integrated network based on modular self-coding, and belongs to the technical field of artificial intelligence. The training method comprises the following steps: acquiring training data, and training a convolutional network auto-encoder by using isovariant constraints based on the training data to obtain a trained auto-encoder; inputting an image I in the training data into the trained auto-encoder to obtain f (i); accessing the obtained f (i) into a base model G (i) to obtain G (i) (f (i)); training the ensemble learning model E by using G (i) (f (i)) to obtain a trained base model G (i) and an ensemble learning model E; according to the method, firstly, the modular auto-encoder is trained by constructing the isovariant constraint, so that the specialization of network module functions is promoted; and then the feature module with function differentiation and the base model are used for ensemble learning, so that the problem of homogenization between the base models of the ensemble learning is reduced, and the generalization ability, robustness and integration effect of the ensemble learning are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a training method and device for an integrated network based on modular autoencoding. Background Art

[0002] Ensemble learning is a machine learning method that combines multiple learners (called base learners) to achieve better performance than a single learner. Compared with a single learner, ensemble learning can reduce the overfitting of the model to the training data and improve the generalization ability of the model by combining the prediction results of multiple different learners. Different base learners may make different mistakes. By integrating the prediction results of multiple learners, the probability of a single learner making a mistake can be reduced and the accuracy of the overall model can be improved. When there are some abnormal data in the data set, a single learner may be misled by these abnormal data, while ensemble learning can reduce the impact of these abnormal data. While improving generalization ability and robustness, ensemble learning is easy to implement and can be combined with any base model, with good flexibility.

[0003] However, ensemble learning is also prone to homogeneity problems, that is, the capabilities of each base model are too similar, which will lead to low ensemble effect, or even worse than a single base learner. Since homogeneous base learners often make similar mistakes, ensembles cannot effectively make up for these mistakes, but will amplify the impact of the errors. Therefore, reducing the homogeneity between base models is crucial to improving ensemble effects. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention provides the following technical solutions.

[0005] The first aspect of the present invention provides a training method for an integrated network based on modular autoencoding, comprising:

[0006] Acquire training data, wherein the training data includes an image I and an image I′ obtained after the image I is transformed;

[0007] Based on the training data, the convolutional network autoencoder is trained using equivariant constraints to obtain a trained autoencoder; wherein the equivariant constraint L EC The calculation is performed using the following formula:

[0008]

[0009] L equ =∑ i ||f′ (i) -M (i) (δ)f (i) ||2 , i∈{0, 1, ..., k-1},

[0010]

[0011] L EC =L recon +λ1L equ +λ2L sym ,

[0012] f=I*W,

[0013] f′=I′*W,

[0014] Where I′=L δ (I), Lδ is a transformation with a transformation parameter δ; W is the convolution kernel of the convolutional network autoencoder, and W = [W (0) , W (1) , ..., W (k-1) ], that is, the convolution kernel is divided into k modules; f is the feature image obtained after the image I is convolved, and f = [f (0) , f (1) , ..., f (k-1) ]; f′ is the feature image obtained after convolution of image I′, f′=[f′ (0) , f′ (1) , ..., f′ (k-1) ]; * indicates convolution operation, represents the deconvolution operation; f (i) is the feature map obtained after image I is convolved by the i-th module of the convolution kernel; f′ (i) is the feature map obtained after the image I′ is convolved by the i-th module of the convolution kernel; M (i) (δ) is the prediction matrix of the i-th module with parameter δ, which is obtained by random initialization and model training optimization; L recon is the reconstruction loss, L equ is the constant variable loss, L sym is the symmetric loss, L EC is an equivariant constraint; λ1 and λ2 are the weight coefficients of equivariant loss and symmetric loss respectively; l represents the number of dimensions of each module; It means that the feature map of the i-th module has only the m-th dimension as 1 and the other dimensions as 0. It indicates that the feature map of the i-th module has only the n-th dimension as 1 and the other dimensions as 0; For arrive The optimal transformation parameters of The prediction matrix M is parameterized by δ (i) The value of the mth row and nth column of (δ); is the prediction matrix M with parameter δ′(i) The value of the mth row and nth column of (δ′); T is a hyperparameter greater than 0; The transformation parameters are The prediction matrix at this time; δ′ is the integral variable, which is used to distinguish the transformation parameter δ under different integral terms;

[0015] Input the image I in the training data into the trained autoencoder to obtain the feature map f of each module (i) , i = 0, 1, 2, ..., k-1; each feature graph f (i) Access a base model G (i) The output result G is obtained (i) (f (i) );Use the output results G of each base model (i) (f (i) ) Train the integrated learning model E to obtain the trained base model G (i) and ensemble learning model E;

[0016] The trained autoencoder and base model G (i) Together with the integrated learning model E, they constitute an integrated network based on modular autoencoding.

[0017] Preferably, the transformation comprises a translation transformation or a translation-rotation transformation.

[0018] The second aspect of the present invention provides an integrated network based on modular autoencoding, which is trained using the method described in the first aspect.

[0019] A third aspect of the present invention provides a training device for an integrated network based on modular autoencoding, comprising:

[0020] A training data acquisition module, used to acquire training data, wherein the training data includes an image I and an image I′ obtained after the image I is transformed;

[0021] The autoencoder training module is used to train the convolutional network autoencoder based on the training data using equivariant constraints to obtain a trained autoencoder; wherein the equivariant constraint L EC The calculation is performed using the following formula:

[0022]

[0023] L equ =∑ i ||f′ (i) -M (i) (δ)f (i) || 2 , i∈{0, 1, ..., k-1},

[0024]

[0025] L EC =L recon +λ1L equ +λ2L sym ,

[0026] f=I*W,

[0027] f′=I′*W,

[0028] Where I′=L δ (I), Lδ is a transformation with a transformation parameter δ; W is the convolution kernel of the convolutional network autoencoder, and W = [W (0) , W (1) , ..., W (k-1) ], that is, the convolution kernel is divided into k modules; f is the feature image obtained after the image I is convolved, and f = [f (0) , f (1) , ..., f (k-1) ]; f′ is the feature image obtained after convolution of image I′, f′=[f′ (0) , f′ (1) , ..., f′ (k-1) ]; * indicates convolution operation, represents the deconvolution operation; f (i) is the feature map obtained after image I is convolved by the i-th module of the convolution kernel; f′ (i) is the feature map obtained after the image I′ is convolved by the i-th module of the convolution kernel; M (i) (δ) is the prediction matrix of the i-th module with parameter δ, which is obtained by random initialization and model training optimization; L recon is the reconstruction loss, L equ is the constant variable loss, L sym is the symmetric loss, L EC is an equivariant constraint; λ1 and λ2 are the weight coefficients of equivariant loss and symmetric loss respectively; l represents the number of dimensions of each module; It means that the feature map of the i-th module has only the m-th dimension as 1 and the other dimensions as 0. It indicates that the feature map of the i-th module has only the n-th dimension as 1 and the other dimensions as 0; For arrive The optimal transformation parameters of The prediction matrix M is parameterized by δ (i) The value of the mth row and nth column of (δ); is the prediction matrix M with parameter δ′ (i) The value of the mth row and nth column of (δ′); T is a hyperparameter greater than 0; The transformation parameters are The prediction matrix at this time; δ′ is the integral variable, which is used to distinguish the transformation parameter δ under different integral terms;

[0029] The base model and ensemble learning model training module is used to input the image I in the training data into the trained autoencoder to obtain the feature map f of each module (i) , i = 0, 1, 2, ..., k-1; each feature graph f (i) Access a base model G (i) The output result G is obtained (i) (f (i) );Use the output results G of each base model (i) (f (i) ) Train the integrated learning model E to obtain the trained base model G (i) and ensemble learning model E;

[0030] The integrated network based on modular autoencoder is used to integrate the trained autoencoder and base model G (i) Together with the integrated learning model E, they constitute an integrated network based on modular autoencoding.

[0031] Preferably, the transformation comprises a translation transformation or a translation-rotation transformation.

[0032] A fourth aspect of the present invention provides a memory storing a plurality of instructions, wherein the instructions are used to implement the training method of the integrated network based on modular autoencoding as described in the first aspect.

[0033] The fifth aspect of the present invention provides an electronic device, comprising a processor and a memory connected to the processor, wherein the memory stores a plurality of instructions, which can be loaded and executed by the processor so that the processor can execute the training method of the integrated network based on modular autoencoding as described in the first aspect.

[0034] The beneficial effects of the present invention are as follows: the training method and device of the integrated network based on modular autoencoder provided by the present invention first construct equivariant constraints, and use the equivariant constraints to train the modular network autoencoder on the training data of the homologous image pair to obtain a trained autoencoder, thereby enhancing the correlation within the autoencoder module and the independence between modules, and promoting the specialization of the network module function. Then, the functionally differentiated feature modules and base models are used to perform integrated learning. Due to the functional differentiation between modules, the homogeneity problem between the integrated learning base models can be greatly reduced, thereby improving the generalization ability, robustness and integration effect of the integrated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1This is a flow chart of the training method of the integrated network based on modular autoencoder according to the present invention;

[0036] Figure 2 Schematic diagram of the training process of modular autoencoder;

[0037] Figure 3 Schematic diagram of the training process of the integrated network based on modular autoencoder;

[0038] Figure 4 It is a schematic diagram of the functional differentiation effect under modular equivariant constraints;

[0039] Figure 5 This is a functional structural diagram of the training device of the integrated network based on modular autoencoding according to the present invention. DETAILED DESCRIPTION

[0040] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0041] The method provided by the present invention can be implemented in the following terminal environment, and the terminal may include one or more of the following components: a processor, a memory, and a display screen. The memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the method described in the following embodiment.

[0042] The processor may include one or more processing cores. The processor uses various interfaces and lines to connect various parts of the entire terminal, and executes various functions of the terminal and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory, and calling data stored in the memory.

[0043] The memory may include random access memory (RAM) or read-only memory (ROM). The memory may be used to store instructions, programs, codes, code sets or instructions.

[0044] The display screen is used to display the user interface of each application.

[0045] In addition, those skilled in the art can understand that the structure of the above terminal does not constitute a limitation on the terminal, and the terminal may include more or fewer components, or combine certain components, or arrange the components differently. For example, the terminal also includes components such as a radio frequency circuit, an input unit, a sensor, an audio circuit, and a power supply, which will not be described in detail here.

[0046] Embodiment 1

[0047] like Figure 1As shown, an embodiment of the present invention provides a training method for an integrated network based on modular autoencoding, comprising:

[0048] S101, obtaining training data, wherein the training data includes an image I and an image I′ obtained after the image I is transformed;

[0049] S102, based on the training data, train the convolutional network autoencoder using equivariant constraints to obtain a trained autoencoder; wherein the equivariant constraint L EC The calculation is performed using the following formula:

[0050]

[0051] L equ =∑ i ||f′ (i) -M (i) (δ)f (i) || 2 , i∈{0, 1, ..., k-1},

[0052]

[0053] L EC =L recon +λ1L equ +λ2L sym ,

[0054] f=I*W,

[0055] f′=I′*W,

[0056] Where I′=L δ (I), Lδ is a transformation with a transformation parameter δ; W is the convolution kernel of the convolutional network autoencoder, and W = [W (0) , W (1) , ..., W (k-1) ], that is, the convolution kernel is divided into k modules; f is the feature image obtained after the image I is convolved, and f = [f (0) , f (1) , ..., f (k-1) ]; f′ is the feature image obtained after convolution of image I′, f′=[f′ (0) , f′ (1) , ..., f′ (k-1) ]; * indicates convolution operation, represents the deconvolution operation; f (i) is the feature map obtained after image I is convolved by the i-th module of the convolution kernel; f′ (i) is the feature map obtained after the image I′ is convolved by the i-th module of the convolution kernel; M (i)(δ) is the prediction matrix of the i-th module with parameter δ, which is obtained by random initialization and model training optimization; L recon is the reconstruction loss, L equ is the constant variable loss, L sym is the symmetric loss, L EC is an equivariant constraint; λ1 and λ2 are the weight coefficients of equivariant loss and symmetric loss respectively; l represents the number of dimensions of each module; It means that the feature map of the i-th module has only the m-th dimension as 1 and the other dimensions as 0. It indicates that the feature map of the i-th module has only the n-th dimension as 1 and the other dimensions as 0; For arrive The optimal transformation parameters of The prediction matrix M is parameterized by δ (i) The value of the mth row and nth column of (δ); is the prediction matrix M with parameter δ′ (i) The value of the mth row and nth column of (δ′); T is a hyperparameter greater than 0; The transformation parameters are The prediction matrix at this time; δ′ is the integral variable, which is used to distinguish the transformation parameter δ under different integral terms;

[0057] S103, input the image I in the training data into the trained autoencoder to obtain the feature map f of each module (i) , i = 0, 1, 2, ..., k-1; each feature graph f (i) Access a base model G (i) The output result G is obtained (i) (f (i) );Use the output results G of each base model (i) (f (i) ) Train the integrated learning model E to obtain the trained base model G (i) and ensemble learning model E;

[0058] S104, the trained autoencoder and base model G (i) Together with the integrated learning model E, they constitute an integrated network based on modular autoencoding.

[0059] It should be noted that The subscript "m" in The subscript "m" in , although it represents the number of dimensions in the feature map and the number of rows in the matrix, but since the feature map and the matrix are corresponding, The subscript "m" in The subscript "m" in is corresponding. In order to illustrate the corresponding relationship between the two, the same letter "m" is used to represent them in the present invention. Similarly, The subscript "n" in The subscript "n" in , although it represents the number of dimensions in the feature map and the number of columns in the matrix, but since the feature map and the matrix are corresponding, The subscript "n" in The subscript "n" in is corresponding. In order to illustrate the corresponding relationship between the two, the same letter "n" is used to represent them in the present invention.

[0060] In a preferred embodiment of the present invention, in step S101, the transformation may include a translation transformation or a translation rotation transformation, so as to obtain a homologous data pair image I and image I'.

[0061] In the above training method, the training process of the modular autoencoder can be as follows Figure 2 As shown in Figure 2, the training process of the integrated network based on modular autoencoders can be as follows: Figure 3 shown.

[0062] The present invention can be implemented in the following steps during practical application:

[0063] Step 1: construct an autoencoder consisting of convolutions. And divide the convolution kernel into k modules. When training the autoencoder, the untransformed image I and the transformed image I′ can be simultaneously input into the constructed autoencoder, and the k modules can be used to perform convolution operations respectively to obtain the feature image corresponding to each module. In addition, an equivariant constraint condition can be constructed, and the equivariant constraint condition can be used to train and optimize the convolution kernel W and prediction matrix M of the constructed convolutional network autoencoder. After the training is completed, all parameters are solidified to obtain a modular autoencoder with differentiated functions.

[0064] Step 2: Get the modular autoencoder with k modules trained in step 1, and the convolution kernel is W = [W (0) , W (1) , ..., W (k-1) ].

[0065] Step 3: construct k base models. The base models can be downstream task networks such as classification, segmentation, and detection. The k base models can be expressed as [G (0) , G (1) ,,..,G (k-1) ].

[0066] Step 4: Input the image I in the training data into the autoencoder trained in step 1 to obtain a modular feature map f = I * W, where W = [W (0) , W (1) , ..., W(k-1) ],f=[f (0) , f (1) , ..., f (k-1) ].

[0067] Step 5: Connect the feature graphs obtained from each module to a base model to obtain G (i) (f (i) ).

[0068] Step 6: Use G (i) (f (i) ) to conduct ensemble learning training E([G (0) (f (0) ), G (1) (f (1) ), ..., G (k-1) (f (k -1) )]), where E is the ensemble learning model.

[0069] Step 7: Perform supervised training such as stochastic gradient descent according to steps 2 to 6 to obtain the final trained base model [G (0) , G (1) , ..., G (k-1) ] and integrated network model E. The trained modular autoencoder W = [W (0) , W (1) , ..., W (k-1) ], base model [G (0) , G (1) , ..., G (k-1) ] and the integrated learning model E together constitute an integrated network based on modular autoencoding.

[0070] In the present invention, equivariant constraints are used to enhance the correlation within the modules and the independence between the modules, so that the modules can form spontaneous functional differentiation. Figure 4 The effect of functional differentiation under modular equivariant constraints is demonstrated. Figure 4In the figure, each row in Figure a, Figure b and Figure c is a module, and each block is a visualization of a learned convolution kernel. It can be seen from the figure that there is an obvious functional differentiation between the modules, such as differentiation of different spatial frequencies, brightness and color. Figure d shows the visualization of the reconstruction of the features extracted by each module in Figure c, and the differentiation of the functions between the modules in terms of color, brightness and edges of different spatial frequencies can be more intuitively seen. This differentiated modular feature is conducive to providing a better representation space for downstream tasks, such as classification tasks, detection tasks and associative generation tasks. At the same time, it is conducive to improving the robustness and generalization of downstream tasks. For example, when the style of the input image of the recognition task changes greatly, the conventional classification network may be unable to recognize the phenomenon, that is, the generalization ability is poor. However, for the modular representation in the present invention, even if the image style changes, it will only affect the representation of a few modules, and will not have much impact on other modules. Therefore, it can still be better recognized through integrated learning, thereby improving the generalization ability of the network.

[0071] Embodiment 2

[0072] like Figure 5 As shown, another aspect of the present invention also includes a functional module architecture that is completely consistent with the aforementioned method flow, that is, the embodiment of the present invention also provides a training device for an integrated network based on modular autoencoder, including:

[0073] A training data acquisition module 501 is used to acquire training data, wherein the training data includes an image I and an image I′ obtained after the image I is transformed;

[0074] The autoencoder training module 502 is used to train the convolutional network autoencoder based on the training data using equivariant constraints to obtain a trained autoencoder; wherein the equivariant constraints L EC The calculation is performed using the following formula:

[0075]

[0076] L equ =∑ i ||f′ (i) -M (i) (δ)f (i) || 2 , i∈{0, 1, ..., k-1},

[0077]

[0078] L EC =L recon +λ1L equ +λ2L sym ,

[0079] f=I*W,

[0080] f′=I′*W,

[0081] Where I′=L δ (I), Lδ is a transformation with a transformation parameter δ; W is the convolution kernel of the convolutional network autoencoder, and W = [W (0) , W (1) , ..., W (k-1) ], that is, the convolution kernel is divided into k modules; f is the feature image obtained after the image I is convolved, and f = [f (0) , f (1) , ..., f (k-1) ]; f′ is the feature image obtained after convolution of image I′, f′=[f′ (0) , f′ (1) , ..., f′ (k-1) ]; * indicates convolution operation, represents the deconvolution operation; f (i) is the feature map obtained after image I is convolved by the i-th module of the convolution kernel; f′ (i) is the feature map obtained after the image I′ is convolved by the i-th module of the convolution kernel; M (i) (δ) is the prediction matrix of the i-th module with parameter δ, which is obtained by random initialization and model training optimization; L recon is the reconstruction loss, L equ is the constant variable loss, L sym is the symmetric loss, L EC is an equivariant constraint; λ1 and λ2 are the weight coefficients of equivariant loss and symmetric loss respectively; l represents the number of dimensions of each module: It means that the feature map of the i-th module has only the m-th dimension as 1 and the other dimensions as 0. It indicates that the feature map of the i-th module has only the n-th dimension as 1 and the other dimensions as 0; For arrive The optimal transformation parameters of The prediction matrix M is parameterized by δ (i) The value of the mth row and nth column of (δ); is the prediction matrix M with parameter δ′ (i) The value of the mth row and nth column of (δ′); T is a hyperparameter greater than 0; The transformation parameters are The prediction matrix at this time; δ′ is the integral variable, which is used to distinguish the transformation parameter δ under different integral terms;

[0082] The base model and ensemble learning model training module 503 is used to input the image I in the training data into the trained autoencoder to obtain the feature map f of each module. (i) , i = 0, 1, 2, ..., k-1; each feature graph f (i) Access a base model G (i) The output result G is obtained (i) (f (i) );Use the output results G of each base model (i) (f (i) ) Train the integrated learning model E to obtain the trained base model G (i) and ensemble learning model E;

[0083] The integrated network based on modular autoencoder constitutes a module 504, which is used to integrate the trained autoencoder, the base model G (i) Together with the integrated learning model E, they constitute an integrated network based on modular autoencoding.

[0084] The device can be implemented by the training method of the integrated network based on modular autoencoding provided in the above-mentioned embodiment 1. The specific implementation method can be found in the description of embodiment 1 and will not be repeated here.

[0085] The present invention also provides a memory storing a plurality of instructions, wherein the instructions are used to implement the training method of the integrated network based on modular autoencoding as described in the first embodiment.

[0086] The present invention also provides an electronic device, comprising a processor and a memory connected to the processor, wherein the memory stores a plurality of instructions, and the instructions can be loaded and executed by the processor so that the processor can execute the training method of the integrated network based on modular autoencoding as described in Example 1.

[0087] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A training method for an integrated network based on modular autoencoders, characterized in that: include: Acquire training data, wherein the training data includes an image I and an image I′ obtained after the image I is transformed; Based on the training data, the convolutional network autoencoder is trained using equivariant constraints to obtain a trained autoencoder; wherein the equivariant constraint L EC The calculation is performed using the following formula: L equ =∑ i ||f′ (i) -M (i) (δ)f (i) || 2 ,i∈{0,1,...,k-1}, L EC =L recon +λ1L equ +λ2L sym , f=I*W, f′=I′*W, Where I′=L δ (I), Lδ is a transformation with a transformation parameter δ; W is the convolution kernel of the convolutional network autoencoder, and W = [W (0) , W (1) , ..., W (k-1) ], that is, the convolution kernel is divided into k modules; f is the feature image obtained after the image I is convolved, and f = [f (0) , f (1) , ..., f (k-1) ]; f′ is the feature image obtained after convolution of image I′, f′=[f′ (0) , f′ (1) , ..., f′ (k-1) ]; * indicates convolution operation, represents the deconvolution operation; f (i) is the feature map obtained after image I is convolved by the i-th module of the convolution kernel; f′ (i) is the feature map obtained after the image I′ is convolved by the i-th module of the convolution kernel; M (i) (δ) is the prediction matrix of the i-th module with parameter δ, which is obtained by random initialization and model training optimization; L recon is the reconstruction loss, L equ is the constant variable loss, L sym is the symmetric loss, L EC is an equivariant constraint; λ1 and λ2 are the weight coefficients of equivariant loss and symmetric loss respectively; l represents the number of dimensions of each module; It means that the feature map of the i-th module has only the m-th dimension as 1 and the other dimensions as 0. It indicates that the feature map of the i-th module has only the n-th dimension as 1 and the other dimensions as 0; For arrive The optimal transformation parameters of The prediction matrix M is parameterized by δ (i) The value of the mth row and nth column of (δ); is the prediction matrix M with parameter δ′ (i) The value of the mth row and nth column of (δ′); T is a hyperparameter greater than 0; The transformation parameters are The prediction matrix at this time; δ′ is the integral variable, which is used to distinguish the transformation parameter δ under different integral terms; Input the image I in the training data into the trained autoencoder to obtain the feature map f of each module (i) , i = 0, 1, 2, ..., k-1; each feature graph f (i) Access a base model G (i) The output result G is obtained (i) (f (i) );Use the output results G of each base model (i) (f (i) ) Train the integrated learning model E to obtain the trained base model G (i) and the trained ensemble learning model E; The trained autoencoder and the trained base model G (i) Together with the trained integrated learning model E, it constitutes an integrated network based on modular autoencoding.

2. The training method of the integrated network based on modular autoencoder according to claim 1, characterized in that: The transformation includes a translation transformation or a translation-rotation transformation.

3. An integrated network based on modular autoencoding, characterized in that: The method is trained by any one of claims 1 to 2.

4. A training device for an integrated network based on modular autoencoders, characterized in that: include: A training data acquisition module, used to acquire training data, wherein the training data includes an image I and an image I′ obtained after the image I is transformed; The autoencoder training module is used to train the convolutional network autoencoder based on the training data using equivariant constraints to obtain a trained autoencoder; wherein the equivariant constraint L EC The calculation is performed using the following formula: L equ =∑ i ||f′ (i) -M (i) (δ)f (i) || 2 ,i∈{0,1,...,k-1}, L EC =L recon +λ1L equ +λ2L sym , f=I*W, f′=I′*W, Where I′=L δ (I), Lδ is a transformation with a transformation parameter δ; W is the convolution kernel of the convolutional network autoencoder, and W = [W (0) , W (1) , ..., W (k-1) ], that is, the convolution kernel is divided into k modules; f is the feature image obtained after the image I is convolved, and f = [f (0) , f (1) , ..., f (k-1) ]; f′ is the feature image obtained after convolution of image I′, f′=[f′ (0) , f' (1) , ..., f′ (k-1) ]; * indicates convolution operation, represents the deconvolution operation; f (i) is the feature map obtained after image I is convolved by the i-th module of the convolution kernel; f′ (i) is the feature map obtained after the image I′ is convolved by the i-th module of the convolution kernel; M (i) (δ) is the prediction matrix of the i-th module with parameter δ, which is obtained by random initialization and model training optimization; L recon is the reconstruction loss, L equ is the constant variable loss, L sym is the symmetric loss, L EC is an equivariant constraint; λ1 and λ2 are the weight coefficients of equivariant loss and symmetric loss respectively; l represents the number of dimensions of each module: It means that the feature map of the i-th module has only the m-th dimension as 1 and the other dimensions as 0. It indicates that the feature map of the i-th module has only the n-th dimension as 1 and the other dimensions as 0; For arrive The optimal transformation parameters of The prediction matrix M is parameterized by δ ( i ) The value of the mth row and nth column of (δ); is the prediction matrix M with parameter δ′ ( i ) The value of the mth row and nth column of (δ′); T is a hyperparameter greater than 0; The transformation parameters are The prediction matrix at this time; δ′ is the integral variable, which is used to distinguish the transformation parameter δ under different integral terms; The base model and ensemble learning model training module is used to input the image I in the training data into the trained autoencoder to obtain the feature map f of each module (i) , i = 0, 1, 2, ..., k-1; connect each feature graph f(i) obtained to a base model G(i) to obtain the output result G (i) (f (i) );Use the output results G of each base model (i) (f (i) ) Train the integrated learning model E to obtain the trained base model G (i) and the trained ensemble learning model E; The integrated network based on modular autoencoder is used to integrate the trained autoencoder and the trained base model G (i) Together with the trained integrated learning model E, it constitutes an integrated network based on modular autoencoding.

5. The training device of the integrated network based on modular autoencoder according to claim 4, characterized in that: The transformation includes a translation transformation or a translation-rotation transformation.

6. A memory, characterized in that: Multiple instructions are stored, and the instructions are used to implement the training method of the integrated network based on modular autoencoding as described in any one of claims 1-2.

7. An electronic device, characterized in that: It includes a processor and a memory connected to the processor, wherein the memory stores a plurality of instructions, and the instructions can be loaded and executed by the processor so that the processor can execute the training method of the integrated network based on modular autoencoding as described in any one of claims 1-2.

Citation Information

Patent Citations

  • Gating feature attention equivariant segmentation method based on weak supervised learning

    CN114373094A

  • Training and reasoning to predict direction of object in image using neural network

    CN114787879A

  • Data and compute efficient equivariant convolutional networks

    US20210248467A1