A black-box model explainable attribution method and device based on visual editing path

Abstract

The application provides a black box model explainable attribution method and device based on a visual editing path, relates to the field of black box model explanation, and comprises the following steps: obtaining a trained black box image classification model, and selecting a path method to explain the black box image classification model; constructing a minimum sufficiency principle and a visual editing path set; under the minimum sufficiency principle, a compressed sensing reconstruction SAMP algorithm for quickly and approximately solving an optimal path is obtained through the visual editing path set; the SAMP algorithm is initialized, the SAMP algorithm is iteratively run through a greedy search strategy, and an explainable attribution result is obtained, wherein an L1 norm constraint and a momentum updating mechanism are introduced in the running process; and the explainable attribution result is normalized and visualized to obtain a contribution value of each pixel in an original image to a result of the black box image classification model. The application can intuitively and completely display the decision logic of the black box model, and improves the credibility of the black box model for users.

Description

Technical Field

[0001] This application relates to the field of black box model interpretation, and in particular to an interpretable attribution method and apparatus for black box models based on visual editing paths. Background Technology

[0002] Deep learning has facilitated applications in the field of artificial intelligence in recent years. However, the black-box nature of deep neural networks (DNNs) hinders the understanding of the decision-making mechanisms of complex models, posing risks to safety-critical industry applications such as autonomous driving and healthcare.

[0003] Currently, some methods attribute a model's output to each component of the input and interpret the corresponding contribution by assigning attribution values. Attribution methods, as a debugging tool, can be used to monitor errors in the model and data. However, users still seem to have little trust in attribution results, and research suggests this distrust stems from a lack of rigor and understandability. Therefore, there is an urgent need to enhance the rigor of attribution methods to eliminate the influence of the method's own errors on the results, thereby gaining user trust. Simultaneously, to improve understandability, other criteria are equally important in aligning attribution with its actual meaning.

[0004] In the field of game theory, existing research has shown that the path method is the only method that simultaneously satisfies additivity, sensitivity, and effectiveness. However, rigor alone is not enough to obtain a credible explanation. Different path choices greatly affect the attribution results of existing path methods. Currently, some path methods use basic axioms to ensure the rigor of attribution, but the actual content of the attribution is unclear due to the different path choices, thus leading to ambiguity in the explanation. Summary of the Invention

[0005] To address the aforementioned issues, a method and apparatus for interpretable attribution of black-box models based on visual editing paths are proposed to enhance users' understanding of the model's decision-making logic, thereby improving the reliability of black-box models in fields such as autonomous driving and healthcare.

[0006] The first aspect of this application proposes an interpretable attribution method based on a black-box model of visual editing paths, including:

[0007] Obtain the trained black box image classification model, and select a path method to interpret the trained black box image classification model;

[0008] A minimum sufficiency principle and a set of visual editing paths are constructed. Under the minimum sufficiency principle, the SAMP algorithm for compressive sensing reconstruction, which quickly approximates the optimal path, is obtained through the set of visual editing paths.

[0009] The SAMP algorithm is initialized and iteratively run using a greedy search strategy to obtain interpretable attribution results. During the process, L1 norm constraints and momentum update mechanisms are introduced.

[0010] The interpretable attribution results are normalized and visualized to obtain the contribution value of each pixel in the original image to the results generated by the trained black-box image classification model.

[0011] Optionally, obtaining the trained black-box image classification model includes:

[0012] Obtain an arbitrarily differentiable image classification model;

[0013] Given a preset classification loss function, the parameters of the image classification model are iteratively optimized using the gradient descent algorithm until the loss function basically converges, at which point the training ends and a trained black-box image classification model is obtained.

[0014] Optionally, the path selection method interprets the trained black-box image classification model, including:

[0015] Select starting point image and endpoint image The change in output of the trained black-box image classification model is expressed using a line integral equation. ,as follows:

[0016] ,

[0017] in, Indicates the output quantity at the starting point. Indicates the output quantity at the termination point. This represents a trained black-box image classification model. Represents a path function, and ;

[0018] By using the Riemann equation and its approximation, we obtain:

[0019] ,

[0020] ,

[0021] in, The output change amount Approximate value, Image of the current point;

[0022] Based on the contribution value equation and the approximate line integral equation, the relative relationship between the image classification contribution value and the approximate value is obtained as follows:

[0023] ,

[0024] The contribution value equation is as follows:

[0025] .

[0026] Optionally, the minimum sufficiency principle is to allocate the necessary contribution to the pixels based on the image classification contribution value set, so that there exists a minimum set of high contribution pixels that can completely cover the changes in the image output result;

[0027] The physical meaning of the visual editing path set is that each editing step involves deleting or adding a finite number of image pixels. Specifically, for the visual editing path set... any path function its k-th segment satisfy:

[0028] ,

[0029] in, And all This constitutes the set of all indicators. A non-overlapping partition, namely:

[0030] .

[0031] Optionally, initializing the SAMP algorithm includes:

[0032] Current indicators Initialize to 0, and initialize the current path segment set to an empty set. ;

[0033] Image of the current point Initialize to the starting point image , cumulative attribution results Initialize as a vector of all zeros ;

[0034] The current accumulated gradient Initialize to the gradient at the original point .

[0035] Optionally, the SAMP algorithm is iteratively run using a greedy search strategy to obtain interpretable attribution results. This includes introducing L1 norm constraints and momentum update mechanisms during the process, including:

[0036] The current accumulated gradient is updated using the momentum update mechanism as follows:

[0037] ,

[0038] in, It is the momentum weighting coefficient, and ;

[0039] The projection of the gradient onto the overall editing direction is calculated as follows:

[0040] ,

[0041] in, For gradient;

[0042] according to Choose the largest size. The indexes constitute a set ;

[0043] Construct the k-th path segment as follows:

[0044] ;

[0045] Update the current point image based on the k-th path segment. and cumulative attribution results ,as follows:

[0046] ,

[0047] ,

[0048] ;

[0049] Repeat the above iterative loop until the current point image is reached. Image with termination point When they overlap, the current one will be... As the final, explainable attribution result.

[0050] Optionally, the method further includes:

[0051] The modulus of the k-th path segment is compared with a preset threshold. If the modulus is greater than the preset threshold, the modulus is trimmed. The trimming equation is as follows:

[0052] ,

[0053] in, The preset threshold is defined as follows.

[0054] Optionally, the normalization and visualization processing of the explainable attribution results includes:

[0055] The explainable attribution results are normalized to a range of 0 to 1;

[0056] The normalized interpretable attribution results are visualized in the form of a heatmap.

[0057] The second aspect of this application proposes an interpretable attribution device based on a black-box model of visual editing paths, comprising:

[0058] The interpretation module is used to obtain the trained black box image classification model and select a path method to interpret the trained black box image classification model.

[0059] A construction module is used to construct the minimum sufficiency principle and the set of visual editing paths. Under the minimum sufficiency principle, the compressed sensing reconstruction SAMP algorithm, which quickly approximates the optimal path, is obtained through the set of visual editing paths.

[0060] The iterative running module is used to initialize the SAMP algorithm and iteratively run the SAMP algorithm through a greedy search strategy to obtain interpretable attribution results. During the running process, L1 norm constraints and momentum update mechanisms are introduced.

[0061] The display module is used to normalize and visualize the interpretable attribution results to obtain the contribution value of each pixel in the original image to the results generated by the trained black-box image classification model.

[0062] A third aspect of this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements any of the methods described in the first aspect above.

[0063] The technical solutions provided by the embodiments of this application have at least the following beneficial effects:

[0064] The ability to compute interpretable attributions for differentiable complex black-box models helps to reveal the model's decision-making process, enhances users' understanding of the model's decision-making logic, and allows for tracing back the attention mechanisms that generate results in typical classification models. This facilitates increased trust in complex models among users in visual analysis and recognition fields such as autonomous driving and healthcare.

[0065] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0066] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0067] Figure 1 This is a flowchart illustrating an interpretable attribution method based on a visual editing path black box model according to an embodiment of this application;

[0068] Figure 2 This is a block diagram illustrating an interpretable attribution device based on a visual editing path black box model according to an embodiment of this application;

[0069] Figure 3 It is a block diagram of an electronic device. Detailed Implementation

[0070] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0071] Figure 1 This is a flowchart illustrating an interpretable attribution method based on a visual editing path black box model according to an embodiment of this application, including:

[0072] Step 101: Obtain the trained black box image classification model and select a path method to interpret the trained black box image classification model.

[0073] A black box model is a purely empirical model that establishes functional relationships between variables based on information provided by the system's input-output data or various types of output variable data, without investigating the mechanisms of internal state changes of the system.

[0074] This application aims to generate interpretable attribution results for differentiable black-box models. Since fields such as autonomous driving and healthcare have high requirements for the trustworthiness of model results, this application takes a typical image classification task in the vision field as an example to reveal the attention mechanism that generates the model classification results, which helps to improve users' trust in the classification results.

[0075] In this embodiment of the application, for any differentiable image classification model The input image data is The output image classification labels are The parameters of the classification model are Given a standard classification loss function The model parameters are iteratively optimized using the gradient descent algorithm until the loss function basically converges, thus ending the training and obtaining a trained black-box image classification model.

[0076] It should be noted that, similar to existing related methods, this application uses a path selection method to interpret the trained black-box image classification model. The specific interpretation process is as follows:

[0077] Select starting point image and endpoint image The change in output of the trained black-box image classification model is expressed using a line integral equation. ,as follows:

[0078] ,

[0079] in, Indicates the output quantity at the starting point. Indicates the output quantity at the termination point. This represents a trained black-box image classification model. Represents a path function, and ;

[0080] By using the Riemannian and approximate line integral equations, we obtain:

[0081] ,

[0082] Expanding it, we get:

[0083] ,

[0084] ,

[0085] Therefore, there is .

[0086] in, To output the change Approximate value, For the current point image, This represents the contribution value of each pixel in the original input image to the final image classification result.

[0087] Step 102: Construct the minimum sufficiency principle and the set of visual editing paths. Under the minimum sufficiency principle, the compressed sensing reconstruction SAMP algorithm, which quickly approximates the optimal path, is obtained through the set of visual editing paths.

[0088] In existing technologies, different line integral solution paths can lead to drastically different pixel contribution value allocation results, thereby undermining users' trust in the interpretation of classification results. To eliminate the ambiguity of pixel contribution value allocation, embodiments of this application propose the minimum sufficiency principle, which defines a clear path selection objective, improving the clarity of the path's meaning. Furthermore, to efficiently solve for the optimal path under the minimum sufficiency principle, this application defines a typical path candidate set, namely, the visual editing path set.

[0089] Specifically, the minimum sufficiency principle is to allocate the necessary contribution to the pixels based on the image classification contribution value set, so that there exists a minimum set of high-contribution pixels that can completely cover the changes in the image output result;

[0090] The physical meaning of the visual editing path set is that each editing step involves deleting or adding pixels to a finite number of images. Specifically, for the visual editing path set... any path function its k-th segment satisfy:

[0091] ,

[0092] in, And all This constitutes the set of all indicators. A non-overlapping partition, namely:

[0093] .

[0094] Step 103: Initialize the SAMP algorithm and iteratively run the SAMP algorithm using a greedy search strategy to obtain interpretable attribution results. During the process, L1 norm constraints and momentum update mechanisms are introduced.

[0095] In this embodiment, the SAMP algorithm is first initialized, and the initialization process is as follows:

[0096] Current indicators Initialize to 0, and initialize the current path segment set to an empty set. ;

[0097] Image of the current point Initialize to the starting point image , cumulative attribution results Initialize as a vector of all zeros ;

[0098] The current accumulated gradient Initialize to the gradient at the original point .

[0099] After initialization, the SAMP algorithm will run iteratively, and the index k will automatically increase by 1 after each iteration.

[0100] In this embodiment of the application, in order to improve the rigor of the method, an L1 norm constraint is applied in each iteration to ensure that the step size of each iteration is less than a tolerable upper bound. At the same time, in order to alleviate the local optimal solution dilemma, a gradient update strategy is adopted, thereby helping the solution algorithm to get rid of ill-conditioned points in the solution space.

[0101] Specifically, the current accumulated gradient is updated using a momentum update mechanism, as follows:

[0102] ,

[0103] in, It is the momentum weighting coefficient, and ;

[0104] The projection of the gradient onto the overall editing direction is calculated as follows:

[0105] ,

[0106] in, For gradient;

[0107] according to Choose the largest size. The indexes constitute a set ;

[0108] Construct the k-th path segment as follows:

[0109] ;

[0110] Update the current point image based on the k-th path segment. and cumulative attribution results ,as follows:

[0111] ,

[0112] ,

[0113] ;

[0114] Repeat the above iterative loop until the current point image is reached. Image with termination point When they overlap, the current one will be... As the final explainable attribution result.

[0115] It should be added that the L1 norm constraint proposed in this application uses a modulus pruning strategy to ensure that the modulus of each iteration is less than a given upper bound, thus guaranteeing the strictness of the path integral. That is, the modulus of the k-th path segment is compared with a preset threshold. If the modulus is greater than the preset threshold, the modulus is pruned. The pruning equation is as follows:

[0116] ,

[0117] in, This is a preset threshold, also known as the upper bound.

[0118] Step 104: Normalize and visualize the interpretable attribution results to obtain the contribution value of each pixel in the original image to the results generated by the trained black box image classification model.

[0119] In this embodiment, the explainable attribution results are normalized to between 0 and 1, and the normalized explainable attribution results are visualized in the form of a heatmap, so that the contribution value of each pixel in the original image to the result generated by the image classification model can be obtained.

[0120] It should be noted that, according to the principle of minimum sufficiency, the data components corresponding to larger attribution values ​​contribute more to the change in the final output of the model.

[0121] In this embodiment, for fields with reliability requirements (such as autonomous driving and healthcare), users can intuitively analyze the degree of attention the original image classification model pays to different pixels in the image during the decision-making process by observing heatmaps with interpretable attributions. Specifically, for correctly classified images, it helps users understand the model's reasoning logic; for misclassified images, it helps users monitor and debug errors in the classification model, facilitating iterative elimination of errors generated by the visual classification model, and ultimately increasing users' trust in complex image classification models.

[0122] The embodiments of this application can calculate the interpretable attribution of differentiable complex black-box models, which helps to reveal the model decision-making process, enhance users' understanding of the model's decision-making logic, trace back the attention mechanism of typical classification models, and facilitate the improvement of users' trust in complex models in fields such as autonomous driving and medical health visual analysis and recognition.

[0123] Figure 2 This is a block diagram of an interpretable attribution device 200 based on a visual editing path black box model, according to an embodiment of this application, comprising:

[0124] The interpretation module 210 is used to obtain the trained black box image classification model and select a path method to interpret the trained black box image classification model.

[0125] Module 220 is used to construct the minimum sufficiency principle and the set of visual editing paths. Under the minimum sufficiency principle, the compressed sensing reconstruction SAMP algorithm for quickly approximating the optimal path is obtained through the set of visual editing paths.

[0126] The iterative running module 230 is used to initialize the SAMP algorithm and iteratively run the SAMP algorithm through a greedy search strategy to obtain interpretable attribution results. During the running process, L1 norm constraints and momentum update mechanisms are introduced.

[0127] The display module 240 is used to normalize and visualize the interpretable attribution results, and obtain the contribution value of each pixel in the original image to the results generated by the trained black box image classification model.

[0128] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0129] Figure 3 A schematic block diagram of an example electronic device 700 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0130] like Figure 3 As shown, device 700 includes a computing unit 701, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 702 or a computer program loaded into random access memory (RAM) 703 from storage unit 708. The RAM 703 may also store various programs and data required for the operation of device 700. The computing unit 701, ROM 702, and RAM 703 are interconnected via bus 704. Input / output (I / O) interface 705 is also connected to bus 704.

[0131] Multiple components in device 700 are connected to I / O interface 705, including: input unit 706, such as keyboard, mouse, etc.; output unit 707, such as various types of monitors, speakers, etc.; storage unit 708, such as disk, optical disk, etc.; and communication unit 709, such as network card, modem, wireless transceiver, etc. Communication unit 709 allows device 700 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0132] The computing unit 701 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 701 performs the various methods and processes described above, such as the voice command response method. For example, in some embodiments, the voice command response method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 708. In some embodiments, part or all of the computer program may be loaded and / or installed on device 700 via ROM 702 and / or communication unit 709. When the computer program is loaded into RAM 703 and executed by the computing unit 701, one or more steps of the voice command response method described above may be performed. Alternatively, in other embodiments, the computing unit 701 may be configured to perform the voice command response method by any other suitable means (e.g., by means of firmware).

[0133] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0134] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0135] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0136] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0137] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), the Internet, and blockchain networks.

[0138] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server) services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0139] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0140] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An interpretable attribution method based on a black-box model of visual editing paths, characterized in that, include: Obtain the trained black box image classification model, and select a path method to interpret the trained black box image classification model; A minimum sufficiency principle and a set of visual editing paths are constructed. Under the constraint of the minimum sufficiency principle, the optimal visual editing path is quickly approximated using the Compressed Sensing Reconstruction SAMP algorithm based on the set of visual editing paths. The minimum sufficiency principle allocates pixels that make necessary contributions based on the image classification contribution value set, such that there exists a minimum set of high-contribution pixels that can completely cover the changes in the image output result. The set of visual editing paths performs deletion or addition operations on a finite number of image pixels for each editing step. The compressed sensing reconstruction SAMP algorithm is initialized, and the compressed sensing reconstruction SAMP algorithm is run iteratively through a greedy search strategy to obtain interpretable attribution results. During the operation, L1 norm constraints and momentum update mechanisms are introduced. The interpretable attribution results are normalized and visualized to obtain the contribution value of each pixel in the original image to the results generated by the trained black box image classification model. The path selection method interprets the trained black-box image classification model, including: Select starting point image and endpoint image The change in output of the trained black-box image classification model is expressed using a line integral equation. ,as follows: ， in, Indicates the output quantity at the termination point. Indicates the output quantity at the starting point. This represents a trained black-box image classification model. Represents a path function, and ; By using the Riemann equation and its approximation, we obtain: ， ， in, The change in output Approximate value, Image of the current point; Based on the contribution value equation and the approximate line integral equation, the relative relationship between the image classification contribution value and the approximate value is obtained as follows: ， The contribution value equation is as follows: 。 2. The method according to claim 1, characterized in that, The process of obtaining the trained black-box image classification model includes: Obtain an arbitrarily differentiable image classification model; Given a preset classification loss function, the parameters of the image classification model are iteratively optimized using the gradient descent algorithm until the loss function basically converges, at which point the training ends and a trained black-box image classification model is obtained.

3. The method according to claim 1, characterized in that, For the set of visual editing paths any path function its k-th segment satisfy: ， in, And all This constitutes the set of all indicators. A non-overlapping partition, namely: 。 4. The method according to claim 1, characterized in that, The initialization of the compressed sensing reconstruction SAMP algorithm includes: Current indicators Initialize to 0, and initialize the current path segment set to an empty set. ; Image of the current point Initialize to the starting point image The cumulative attribution results Initialize as a vector of all zeros ; The current accumulated gradient Initialize to the gradient at the original point .

5. The method according to claim 4, characterized in that, The compressed sensing reconstruction SAMP algorithm is iteratively run using a greedy search strategy to obtain interpretable attribution results. During the process, L1 norm constraints and momentum update mechanisms are introduced. The L1 norm constraints employ a modulus pruning strategy to ensure that the modulus in each iteration is less than a given upper bound, including: The current accumulated gradient is updated using the momentum update mechanism as follows: ， in, It is the momentum weighting coefficient, and ; The projection of the gradient onto the overall editing direction is calculated as follows: ， in, For gradient; according to Choose the largest size. The indexes constitute a set ; Construct the k-th path segment as follows: ； Update the current point image based on the k-th path segment. and cumulative attribution results ,as follows: ， ， ； Repeat the above iterative loop until the current point image is reached. Image with termination point When they overlap, the current one will be... As the final, explainable attribution result.

6. The method according to claim 5, characterized in that, The method further includes: The modulus of the k-th path segment is compared with a preset threshold. If the modulus is greater than the preset threshold, the modulus is trimmed. The trimming equation is as follows: ， in, The preset threshold is defined as follows.

7. The method according to claim 5, characterized in that, The normalization and visualization processing of the explainable attribution results includes: The explainable attribution results are normalized to a range of 0 to 1; The normalized interpretable attribution results are visualized in the form of a heatmap.

8. A black-box model interpretable attribution device for visual editing paths based on the method of any one of claims 1-7, characterized in that, include: The interpretation module is used to obtain the trained black box image classification model and select a path method to interpret the trained black box image classification model. The construction module is used to construct the minimum sufficiency principle and the set of visual editing paths. Under the constraint of the minimum sufficiency principle, based on the set of visual editing paths, the optimal visual editing path is quickly approximated by the compressed sensing reconstruction SAMP algorithm. The minimum sufficiency principle is to allocate pixels that make necessary contributions based on the image classification contribution value set, so that there exists a minimum set of high contribution pixels that can completely cover the changes in the image output result. The set of visual editing paths performs deletion or addition operations on a finite number of image pixels for each editing step; The iterative running module is used to initialize the compressed sensing reconstruction SAMP algorithm and iteratively run the compressed sensing reconstruction SAMP algorithm through a greedy search strategy to obtain interpretable attribution results. During the running process, L1 norm constraints and momentum update mechanisms are introduced. The display module is used to normalize and visualize the interpretable attribution results to obtain the contribution value of each pixel in the original image to the results generated by the trained black box image classification model. The path selection method interprets the trained black-box image classification model, including: Select starting point image and endpoint image The change in output of the trained black-box image classification model is expressed using a line integral equation. ,as follows: ， in, Indicates the output quantity at the termination point. Indicates the output quantity at the starting point. This represents a trained black-box image classification model. Represents a path function, and ; By using the Riemann equation and its approximation, we obtain: ， ， in, The change in output Approximate value, Image of the current point; Based on the contribution value equation and the approximate line integral equation, the relative relationship between the image classification contribution value and the approximate value is obtained as follows: ， The contribution value equation is as follows: 。 9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method as described in any one of claims 1-7.

Images