Visual analysis system for interpretable decoupling representation learning

Abstract

The invention discloses a visual analysis system for interpretable decoupling representation learning, and the system comprises a training configuration panel module which is constructed at the front end and facilitates the importing of a model and data, the parameter setting and the display of a training state; the projection view module is used for presenting sample embedding distribution auxiliary concept grouping by using a scatter diagram; the group management module is used for maintaining groups and binding training strategies; the image library module is used for displaying samples according to the similarity sequence and supporting interpretation loading; the image level interpretation module and the feature level interpretation module respectively present a model attention area and a diagnosis problem dimension; the parallel coordinate graph module is used for comparing model similarities and carrying out associated traceability; and the historical tree module records training branches and check points to support backtracking. And a back-end calculation module is constructed, training is executed according to configuration, and embedded and output interpretation data is generated. Through cooperation of the front end and the rear end, interaction control and training optimization are realized, deep neural network decoupling representation learning is guided, and model interpretability, concept controllability and semantic alignment capability are enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of intersection of artificial intelligence and visual analysis technology, and specifically relates to a visual analysis system for interpretable decoupled representation learning. Background Art

[0002] With the widespread application of deep learning technology, neural networks have achieved breakthroughs in image recognition, natural language processing, medical diagnosis, and other fields. However, despite the superior performance of these models, their internal representation structures are highly complex and opaque, resulting in a serious "black box" problem, making them difficult for human experts to understand and control. Especially in high-risk or highly sensitive scenarios (such as medical diagnosis, algorithmic decision-making, and social fairness analysis), the lack of interpretable and human-aligned representations not only reduces the model's credibility but also may raise ethical and safety issues. Therefore, how to ensure that models "learn correctly" and, more importantly, "learn correctly" has become a key research focus in the field of explainable AI.

[0003] To alleviate this problem, researchers have proposed the concept of decoupled representation learning, which aims to partition the latent representation learned by a neural network into several independent conceptual dimensions with clear semantic interpretations. This decoupled representation not only improves model interpretability but also makes downstream tasks such as transfer learning, generative modeling, and human-computer collaboration more controllable. However, existing mainstream decoupled learning methods mostly rely on strict supervision (i.e., providing clear labels for each conceptual dimension) or prior assumptions (such as independent and identically distributed and structural causality). This leads to bottlenecks in practical large-scale applications, such as expensive labeling, incomplete semantics, and bias transmission.

[0004] To this end, recent research has proposed interactive decoupled representation learning. This approach emphasizes user engagement, such as clicks, markup, and selections, to inject human intuition and prior knowledge into the representation space, guiding the model to align with human semantic divisions. For example, an expert can guide the model to understand visual attributes such as "skin color," "gender," or "background" by selecting a group of similar samples and assigning them conceptual labels. However, this type of interaction often presents the following three challenges: (1) High-dimensional representations are difficult to understand and operate: The potential representations of model outputs are often high-dimensional vectors, which are difficult to directly correspond to human semantics. Traditional dimensional visualization (such as t-SNE and UMAP) only provides two-dimensional projections and lacks explanation and feedback paths at the semantic level. (2) Fragmented and weakly closed-loop model interpretation mechanisms: Although some interpretation methods have emerged, such as LRP, GradCAM, and TCAV, they are mostly used for static visualization and lack linkage mechanisms with the model training process. They cannot adjust the model representation in a timely manner according to user annotations or corrections. (3) The associations between concepts are complex, and intervention may cause side effects: the representations learned by the model are often the composite results of multiple factors. Modifying one concept dimension may have unpredictable effects on other dimensions. Existing methods lack tracking and diagnostic mechanisms to deal with such "concept interference."

[0005] To address these challenges, research has attempted to incorporate visual analytics methods into the decoupled representation learning process. For example, systems such as ConceptExplainer, Visual Concept Programming, and ConceptExtract combine projected visualization, concept activation vectors, and interactive annotations to assist users in understanding and intervening in the model learning process. However, most of these systems lack closed-loop training capabilities. This means that user interaction results cannot be used to reverse-engineer the original model parameters, limiting the true utility of explanations for model improvement.

[0006] Furthermore, in the area of ​​model learning, some scholars have proposed joint modeling approaches that integrate multiple supervisory mechanisms, such as supervised contrastive learning and invariant risk minimization. These approaches, respectively, guide models to better identify and separate semantic concepts from the perspectives of equivariance and invariance. However, embedding these approaches into interactive visualization systems to truly achieve "human-machine collaborative concept modeling" remains a challenge in current research.

[0007] In summary, existing research on disaggregated representation learning and visualization analysis still has significant deficiencies in the three areas of "understandability, controllability, and feedback." There is a lack of a unified system that can both support experts in exploring the representations learned by the model through a visual interface and enable fine-grained control and reverse optimization of the model through interpretation mechanisms and interactive operations. Therefore, there is an urgent need to design a multi-view visual analysis system for interactive interpretable learning tasks that integrates representation projection, inter-group interpretation, dimensional clustering diagnosis, training version management, and model feedback, in order to establish a closed-loop process for interpretable disaggregated representation learning that is co-built by humans and machines. Summary of the Invention

[0008] In view of the above, the purpose of the present invention is to provide a visual analysis system for interpretable decoupled representation learning, by constructing eight front-end functional modules for performing visualization of representation space, sample grouping management, training strategy binding, image interpretation and model version comparison, and constructing a back-end module for performing model training, embedding generation and interpreting data output according to user configuration, realizing interactive control and training optimization in the representation learning process, and guiding deep neural networks to perform decoupled representation learning visualization system through human-computer interaction, thereby enhancing the model's interpretability, conceptual controllability and semantic alignment capabilities.

[0009] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: An embodiment of the present invention provides a visual analysis system for interpretable decoupled representation learning, including: The training configuration panel module provides image recognition model and image dataset import, training parameter setting, and displays the model training status through subsequent views based on the parameter setting results; The projection view module includes a t-SNE dimension reduction scatter plot and a triple similarity scatter plot, which is used to present the sample embedding distribution generated by the selected model in a single iteration and perform preliminary concept grouping; The group management module is used to maintain the preliminary concept groupings and bind supervised contrastive learning or invariant risk minimization training strategies to the concept groupings to obtain different concept groupings; The image library module displays sample thumbnails in order of cosine similarity between the sample and the center of the concept group representation, i.e., the group centroid, and supports triggering interpretation loading to load the sample into subsequent views; An image-level explanation module that renders weighted bidirectional layer-wise correlation propagation explanations for selected pairs of samples, helping users identify areas of interest in the model's input. The feature-level interpretation module performs hierarchical clustering of selected sample pairs on different dimensions based on the dimension-level layer-by-layer correlation propagation heat map and presents the clustering results for diagnosing problem dimensions; The parallel coordinates graph module is used to compare the similarity distribution of different models on multiple concept axes, supports conditional filtering, and links back to other modules for traceability; The history tree module records model training branches and checkpoints and supports retrospective loading.

[0010] Preferably, the training parameter settings in the training configuration panel module include: Provides input boxes for batch size, number of training epochs, temperature, learning rate, and a weight slider for supervised contrastive learning or invariant risk minimization training strategies.

[0011] Preferably, the t-SNE dimension reduction scatter plot in the projection view module is used to globally display the sample embedding distribution and support lasso generation of concept grouping; The triple similarity scatter plot is used to calculate the average cosine similarity of all samples to the two anchor points specified by the user and redraw it in two-dimensional coordinates for analyzing concept separation; At the same time, the projection view module provides an automatic grouping button to automatically divide clusters according to the scatter distribution to obtain preliminary concept groupings and push them to the group management module.

[0012] Preferably, in the group management module, the group names of all concept groups and the current training strategy labels are displayed in the form of cards, and support is provided for renaming concept groups, merging multiple groups, hiding or showing groups, and deleting groups. The status of the current training strategy label supports setting between three training strategies: no setting, supervised contrastive learning, or invariant risk minimization.

[0013] Preferably, in the image library module, supporting triggering interpretation loading includes: Sample image pairs are loaded for generating image-level explanations in the image-level explanation module, and explanation results are generated in batches in the image-level explanation module according to the current intra-group mode or cross-group mode.

[0014] Preferably, in the image-level interpretation module, the following steps are performed on the selected sample: First, the dimension-level layer-by-layer correlation propagation heat map based on the selected sample pairs is pooled and normalized, and the low-contribution value areas are thinned; then, the heat map is threshold-cropped; the heat map is weightedly combined according to the contribution weight of each dimension, the similarity explanation matrix is ​​calculated, and the pixel-level alignment relationship is constructed through the outer product; finally, the edge clustering algorithm is used to draw the connecting lines, with the line color representing the contribution symbol and the transparency representing the contribution strength; at the same time, the image-level interpretation module provides a parameter configuration interface for adjusting the threshold, sparsity and clustering strength to control the visualization style of the interpretation results.

[0015] Preferably, in the feature-level interpretation module, the Pearson correlation coefficient is used to perform hierarchical clustering on the dimension-level layer-by-layer correlation propagation heat map of the selected samples, generate dimension clustering groups and display the dimension similarity matrix, and support clicking to view single-dimensional explanations and mark the problem dimensions. The marked information is written back to the backend for excluding or downgrading the corresponding dimensions in subsequent training.

[0016] Preferably, in the parallel coordinate graph module, the similarity of each sample in the selected concept grouping is used as the coordinate axis, which supports interval screening, condition combination query and traceability highlighting linkage with the image library module, image-level interpretation module and feature-level interpretation module to realize sample retrieval and model comparison.

[0017] Preferably, in the history tree module, the results of each round of model training are displayed in a tree structure. Each node of the tree structure contains the parameter configuration and model snapshot used in that round of training. Users can click on the node to view the parameter details, and click to load the historical model associated with any node for retrospective analysis, version switching and effect comparison, and synchronously refresh the views of each module, while providing log viewing and node deletion functions.

[0018] Preferably, the system further comprises: The back-end computing module applies the user-configured training strategy to the image recognition model, uses a fusion supervised contrastive learning or invariant risk minimization mechanism for training optimization, and pushes embedding results, performance indicators and explanation data during and after training. It interacts with the training configuration panel module, projection view module, group management module, image library module, image-level interpretation module, feature-level interpretation module, parallel coordinate graph module, and history tree module for data and real-time updates.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The system of the present invention combines embedding visualization, sample grouping, training strategy binding and model training to build a complete closed-loop process from concept discovery to training optimization, enabling users to visually explore and controllable tune the model representation space; the system adopts two interpretation mechanisms, image-level and feature-level, to provide similarity contribution explanations between sample pairs and representation clustering analysis at the dimension level, respectively, enabling experts to understand model behavior and accurately identify problem dimensions from both macro and micro levels; the system supports flexible configuration of two types of strategies, supervised contrastive learning and invariant risk minimization, at the sample group level, thereby effectively enhancing the equivariance and invariance of representation and improving concept independence; in addition, the system provides a tree-like management and backtracking mechanism for training versions, allowing users to easily compare the differences in model representation and strategy effects under different training paths, with good traceability and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 This is a schematic diagram of the interface structure of a visual analysis system for interpretable decoupled representation learning provided by an embodiment of the present invention; Figure 2 It is a scatter plot of triple similarity in the projection view module provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0023] The inventive concept of the present invention is as follows: to address the problems in the existing decoupled representation learning process, such as the unclear concept construction process, the lack of interpretable feedback in the model representation structure, and the inability of user interaction to effectively affect the training process, the embodiment of the present invention provides a visual analysis system for interpretable decoupled representation learning, including a front-end training configuration panel module, a projection view module, a group management module, an image library module, an image-level interpretation module, a feature-level interpretation module, a parallel coordinate graph module, and a history tree module, as well as a back-end computing module. The system integrates functional modules such as representation display, group management, interpretation analysis, and training configuration panels. User interaction can affect the model training process. Through front-end and back-end collaboration, interactive control and training optimization are achieved, guiding the decoupled representation learning of deep neural networks and enhancing the model's interpretability, concept controllability, and semantic alignment capabilities.

[0024] Figure 1 This is a schematic diagram of the interface structure of a visual analysis system for interpretable decoupled representation learning provided by an embodiment of the present invention. Figure 1 As shown, the embodiment provides a visual analysis system for interpretable decoupled representation learning, including: a training configuration panel module, a projection view module, a group management module, an image library module, an image-level interpretation module, a feature-level interpretation module, a parallel coordinate graph module, and a history tree module. In the interface of the visual analysis system, A is the training configuration panel module, B is the projection view module (wherein B1 is the view switching button, B2 is the lasso button, B3 is the cursor button, and B4 is the automatic grouping button), C is the group management module, D is the image library module (wherein D1 is the intra-group interpretation button, D2 is the inter-group interpretation button, and D3 is the load interpretation button), E is the image-level interpretation module, F is the feature-level interpretation module, G is the parallel coordinate graph module, and H is the history tree module. The following is a detailed introduction to each module.

[0025] The training configuration panel, Module A, imports image recognition models and image datasets and allows users to set various model-related training parameters, including the base dataset, batch size, number of training epochs, temperature, learning rate, and penalty weights (primarily Supervised Contrastive Learning (SCL) and Invariant Risk Minimization (IRM) weights in this system). Once the training parameters are configured according to the user's settings, the system trains the model and generates sample representations. After training is complete, the generated representations can be visualized and analyzed using subsequent views.

[0026] In the projection view module B, Figure 1As shown in the figure, the t-SNE dimensionality reduction scatter plot (t-distributed Stochastic Neighbor Embedding, t-SNE) is displayed. Click the Switch View button B1 to switch directly between the t-SNE dimensionality reduction scatter plot and the triple similarity scatter plot. The t-SNE dimensionality reduction scatter plot is used to present the relative distance relationship between samples in the high-dimensional representation space. Each point corresponds to an original image, and the system supports presenting the corresponding image information in the image display area. Click the Lasso button B2 to enable the lasso function. Users can use the lasso tool to manually select multiple samples to form a preliminary concept grouping. Click the Cursor button B3 to enable the cursor function. Users can switch from the lasso to the cursor. When the cursor hovers over an anchor point, a thumbnail image corresponding to the anchor point will be displayed. Click the Auto Grouping button B4 to enable the Auto Partition function. The system will automatically divide high-density areas based on the sample distribution and generate preliminary concept groupings.

[0027] In the projection view module B, Figure 2 As shown in the figure, it shows a triplet similarity scatter plot (TripletPlot), which is used to display the cosine similarity of the sample relative to the two anchor groups and redrawn with (x, y) coordinates. The X-axis represents the similarity of the sample to the first group, and the Y-axis represents the similarity to the second group. If the sample points are distributed near the diagonal, it means that the model has difficulty distinguishing the concepts represented by the two anchor groups; if the sample points deviate from the diagonal, it indicates that the model has a strong ability to distinguish the concept. This view can help users evaluate the adequacy of the model's learning of specific concepts. Through the t-SNE dimensionality reduction scatter plot and triplet similarity scatter plot Figure 2 The authors jointly supported the creation of sample groups and the identification of conceptual boundaries.

[0028] The Group Management Module C displays all concept groups created by the current user. Each group can be renamed, hidden, shown, deleted, and merged. Users can also specify the type of supervisory signal (such as IRM or SCL) for a specific group. In the next round of training, the system will guide the model to enforce invariance or equivariance based on the user's specified method. This policy configuration is then passed to the backend computation module to influence subsequent training. When a user selects a group, the Image Library Module D automatically displays thumbnails of all samples in that group for subsequent interpretation.

[0029] Image Library Module D displays the sample images in the currently selected group, sorted by their cosine similarity to the group center by default. Users can click on a thumbnail to view the original image of the sample separately; click on the "Load Interpretation" icon in the upper right corner of a thumbnail to transfer it and the currently selected reference sample (in the intra-group / inter-group mode) to Image Library Module E to generate a BiLRP interpretation map. Figure 1As shown in the image library module D, clicking the Intra-Group Interpretation button D1 loads the intra-group BiLRP interpretation results for the image pair within that group; clicking the Inter-Group Interpretation button D2 loads the inter-group interpretation between the two selected groups; and clicking the Load Interpretation button D3 loads the current image pair into the Image-Level Interpretation module for further analysis. This module provides image preview and interactive selection, allowing users to intuitively understand the differences between samples in the representation space.

[0030] The image-level explanation module E displays the pixel-level aligned explanations between image pairs computed using the weighted bidirectional layer-wise relevance propagation (BiLRP) algorithm. Figure 1 The system calculates the similarity of the "background," "subject," and so on, defined in the group management module C, back to the input pixel layer, calculating the "correlation" (positive or negative contribution) of each pixel between the two images. It then uses curves (edge ​​bundles) to connect the corresponding high-contribution pixels in the two images. This allows for docking, zooming, and linked highlighting, forming a visual explanation map. Color indicates directional differences, and transparency indicates the strength of similarity. To reduce the visual redundancy and computational burden of the explanation map, the system introduces 4×4 pooling, normalization, sparsification, thresholding (values ​​limited to the range [-7, 7]), and edge convergence during rendering. This module supports detailed visual hierarchical interpretation of specific image pairs, helping users identify the model's focus areas on the input.

[0031] The feature-level explanation module F displays explanation results for each representation dimension. The system first performs hierarchical clustering on the layer-wise relevance propagation (LRP) heatmaps for all dimensions. Dimensions with similar explanation patterns are grouped together and their average explanation maps are displayed. A heatmap matrix is ​​also displayed below each cluster. In the LRP heatmap, light and dark colors represent numerical magnitude: dark colors correspond to highly similar LRP heatmaps for the two dimensions (indicating that they focus on similar regions of the image). Light colors correspond to less correlated LRP heatmaps for the two dimensions (focusing on different regions), potentially even representing erroneous dimensions. If some dimensions have very low similarity to the majority of dimensions (appearing as "light-colored isolated blocks" in the LRP heatmap), these dimensions may not have learned any meaningful concepts or have only learned noise. Users can expand the LRP heatmaps for specific dimensions within each cluster and identify potentially invalid or erroneous dimensions by combining the displayed dimension numbers with their contribution weights. The system allows users to label specific dimensions and exclude or reduce their influence in subsequent training.

[0032] The parallel coordinate graph module G supports users to compare the distribution of representations of multiple models under different conceptual dimensions. Each line represents a data point: In the parallel coordinate graph module G, there are multiple parallel coordinate axes horizontally, and each axis corresponds to a "concept" (i.e., a group). The position of the line on each axis represents the distance between the data point and the corresponding "concept" (group) anchor point: The center of the "Anchor Group" is used in this article to represent a typical sample of a concept, and the feature vector of the data point (i.e., an image) will calculate the cosine similarity (Cosine Similarity) with these typical samples. In the parallel coordinate graph module G, the intersection of each line and an axis corresponds to the similarity measure (or distance measure) between the data point and the group of anchor points. In addition to selecting different models ( Figure 1 In addition to the numbers (M1, M2, etc.) and groups, the interface typically features a slider ranging from 0 to 1. This slider filters the lines in the parallel coordinate graph, making it easy to identify samples that are most sensitive or have the highest preference for a concept, allowing for interpretable retrieval or comparative analysis. This module allows users to quickly identify whether different trained models have good decoupled representation capabilities for specific concepts. It also supports filtering samples by group, locating anomalous samples, and linking them back to the image interpretation module for tracing.

[0033] The History Tree module H records the results of each round of model training in the system, displaying it as a tree structure. Each node contains the parameter configuration and model snapshot used in that round of training. Users can right-click to view parameter details and double-click to load any historical model for retrospective analysis, version switching, and effect comparison, forming a controllable and traceable training iteration loop.

[0034] In actual use, users can initiate model training through the Training Configuration Panel module. After training is complete, they can view the resulting distribution in the Projection View module, select samples using the lasso to form initial groups, manage groups in the Group Management module, and view samples in the Image Library module. Users can further load interpretation information, identify problematic dimensions or anomalous concepts in the representation in the Image-Level Interpretation and Feature-Level Interpretation modules, explore multiple data sets using the Parallel Coordinates Plot module, and finally recall the model in the History Tree module and iterate repeatedly, guiding the model to learn structured and interpretable concept representations and optimizing their controllability and semantic consistency over time.

[0035] The system also includes a backend computing module, which applies a user-configured training strategy combining SCL and IRM to the image recognition model. The SCL submodule utilizes concept groups provided by the group management module to enhance equivariant learning, while the IRM submodule utilizes environment partitioning to enhance invariant learning. The backend computing module pushes sample embedding results, performance metrics, and interpretation data during and after training. It also distributes this data via interfaces to the training configuration panel module, projection view module, group management module, image library module, image-level interpretation module, feature-level interpretation module, parallel coordinate plot module, and history tree module, supporting data interaction and real-time updates.

[0036] Through the above-mentioned module design and collaborative operation, the system realizes a complete "representation discovery - concept identification - explanation verification - user correction - model iteration" process. While ensuring the model's representation capabilities, it realizes the alignment and regulation of human-understandable concepts, effectively improving the model's interpretability and robustness.

[0037] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A visual analysis system for interpretable decoupled representation learning, characterized by: The system comprises: The training configuration panel module provides image recognition model and image dataset import, training parameter setting, and displays the model training status through subsequent views based on the parameter setting results; The projection view module includes a t-SNE dimension reduction scatter plot and a triple similarity scatter plot, which is used to present the sample embedding distribution generated by the selected model in a single iteration and perform preliminary concept grouping; The group management module is used to maintain the preliminary concept groupings and bind supervised contrastive learning or invariant risk minimization training strategies to the concept groupings to obtain different concept groupings; The image library module displays sample thumbnails in order of cosine similarity between the sample and the center of the concept group representation, i.e., the group centroid, and supports triggering interpretation loading to load the sample into subsequent views; An image-level explanation module that renders weighted bidirectional layer-wise correlation propagation explanations for selected pairs of samples, helping users identify areas of interest in the model's input. The feature-level interpretation module performs hierarchical clustering of selected sample pairs on different dimensions based on the dimension-level layer-by-layer correlation propagation heat map and presents the clustering results for diagnosing problem dimensions; The parallel coordinates graph module is used to compare the similarity distribution of different models on multiple concept axes, supports conditional filtering, and links back to other modules for traceability; The history tree module records model training branches and checkpoints and supports retrospective loading.

2. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: The training parameter settings in the training configuration panel module include: Provides input boxes for batch size, number of training epochs, temperature, learning rate, and a weight slider for supervised contrastive learning or invariant risk minimization training strategies.

3. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: The t-SNE dimension reduction scatter plot in the projection view module is used to globally display the sample embedding distribution and supports lasso generation of concept grouping; The triple similarity scatter plot is used to calculate the average cosine similarity of all samples to the two anchor points specified by the user and redraw it in two-dimensional coordinates for analyzing concept separation; At the same time, the projection view module provides an automatic grouping button to automatically divide clusters according to the scatter distribution to obtain preliminary concept groupings and push them to the group management module.

4. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: In the group management module, the group names and current training strategy labels of all concept groups are displayed in the form of cards. It also supports renaming concept groups, merging multiple groups, hiding or showing groups, and deleting groups. The status of the current training strategy label supports setting between three training strategies: no setting, supervised contrastive learning, or invariant risk minimization.

5. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: In the image library module, triggering interpretation loading is supported including: Sample image pairs are loaded for generating image-level explanations in the image-level explanation module, and explanation results are generated in batches in the image-level explanation module according to the current intra-group mode or cross-group mode.

6. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: In the image-level interpretation module, the following steps are performed on the selected samples: First, the dimension-level layer-by-layer correlation propagation heat map based on the selected sample pairs is pooled and normalized, and the low-contribution value areas are thinned; then, the heat map is threshold-cropped; the heat map is weightedly combined according to the contribution weight of each dimension, the similarity explanation matrix is ​​calculated, and the pixel-level alignment relationship is constructed through the outer product; finally, the edge clustering algorithm is used to draw the connecting lines, with the line color representing the contribution symbol and the transparency representing the contribution strength; at the same time, the image-level interpretation module provides a parameter configuration interface for adjusting the threshold, sparsity and clustering strength to control the visualization style of the interpretation results.

7. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: In the feature-level interpretation module, the Pearson correlation coefficient is used to perform hierarchical clustering on the dimension-level layer-by-layer correlation propagation heat map of the selected samples, generate dimension cluster groups and display the dimension similarity matrix. It also supports clicking to view single-dimensional explanations and mark the problem dimensions. The marked information is written back to the backend for excluding or downgrading the corresponding dimensions in subsequent training.

8. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: In the parallel coordinate graph module, the similarity of each sample in the selected concept grouping is used as the coordinate axis. It supports interval screening and condition combination query, and realizes traceability highlighting linkage with the image library module, image-level interpretation module and feature-level interpretation module to realize sample retrieval and model comparison.

9. The visual analysis system for interpretable decoupled representation learning according to claim 1, characterized in that: In the history tree module, the results of each round of model training are displayed in a tree structure. Each node of the tree structure contains the parameter configuration and model snapshot used in that round of training. Users can click on a node to view parameter details, and click to load the historical model associated with any node for retrospective analysis, version switching, and effect comparison. It also refreshes the views of each module synchronously and provides log viewing and node deletion functions.

10. The visual analysis system for interpretable decoupled representation learning according to any one of claims 1 to 9, characterized in that: The system further comprises: The back-end computing module applies the user-configured training strategy to the image recognition model, uses a fusion supervised contrastive learning or invariant risk minimization mechanism for training optimization, and pushes embedding results, performance indicators and explanation data during and after training. It interacts with the training configuration panel module, projection view module, group management module, image library module, image-level interpretation module, feature-level interpretation module, parallel coordinate graph module, and history tree module for data and real-time updates.

Images