A model fusion system and method based on interpretable information quantitative evaluation

Abstract

The application provides a model fusion system based on explainable information quantitative evaluation, which is used for fusing multiple trained models and comprises a weight alignment predictor, which is used for processing weight spaces of two models to be fused to obtain a weight alignment relationship between the two models, wherein the weight alignment relationship is expressed as a set of permutation matrices, and each element in the permutation matrix corresponds to a neuron at a level for describing a one-to-one correspondence of the two models in the neuron; a model evaluation unit, which is used for calculating weight contribution degrees of each parameter in each model weight set in the weight space according to the weight alignment relationship between the two models; and a model fusion unit, which is used for calculating fusion proportions of each layer weight of each model based on the feature contribution degrees of each parameter in each model weight, and fusing the weights of the two models based on the fusion proportions.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer technology, specifically to model fusion and model collaboration technology in the fields of artificial intelligence and computer technology, and more specifically, to a model fusion system and method based on interpretable information quantitative evaluation. Background Technology

[0002] With the continuous development of artificial intelligence and machine learning technologies, model fusion technology has gradually become an important means to improve model performance, stability, and generalization ability, and has been widely used in fields such as healthcare, finance, and industrial intelligence. By integrating the prediction results or internal knowledge of multiple models, model fusion can alleviate the performance bottleneck of a single model to some extent when faced with data distribution bias, insufficient samples, or high task complexity. Therefore, research on model fusion has become an important direction in current artificial intelligence applications.

[0003] From a technical perspective, existing model fusion methods mainly include fusion methods that rely on training data and fusion methods that are based on model parameter operations.

[0004] Methods relying on training data typically integrate model capabilities through joint training, model fine-tuning, or knowledge distillation. These methods can fully utilize data distribution information during the fusion process and achieve good fusion results in a controlled environment. However, in practical applications, models are often already trained and deployed across different systems or institutions. The original training data may not be available again or shared due to privacy protection, compliance requirements, or storage costs, significantly limiting the above methods in multi-center collaboration or model reuse scenarios. Furthermore, retraining or fine-tuning models incurs additional computational overhead and operational costs, hindering rapid model fusion and flexible deployment.

[0005] To reduce reliance on training data, some existing techniques attempt to directly fuse multiple trained models at the model parameter level, such as through weighted averaging, linear combination, or weighted summation to generate a fused model. These methods are relatively simple to implement and computationally inexpensive, achieving some success when models have similar structures or training conditions. However, due to potential inconsistencies in parameter arrangement, structural representation, and optimization paths between different models during training, direct parameter fusion can easily introduce parameter space misalignment, leading to decreased performance or instability in the fused model. To address these issues, some improved methods further introduce parameter alignment or structural matching mechanisms to rearrange or match model parameters before fusion, mitigating the impact of parameter space inconsistencies. However, these methods typically rely on complex combinatorial optimization or iterative calculations, incurring significant computational overhead and limited adaptability to different model combinations.

[0006] While existing model fusion techniques have made some progress in theoretical research and certain application scenarios, they still face numerous challenges in real-world environments with heterogeneous model sources, limited data, and high requirements for the controllability and reliability of the fusion process. Current technologies primarily focus on improving the performance of the fusion results, paying less attention to the structural analysis and decision-making basis of the fusion process itself. Their applicability and engineering scalability still need further improvement.

[0007] In short, existing model fusion technologies still have significant shortcomings in terms of applicability under data-free conditions, stability of the fusion process, and interpretability of fusion decisions, making it difficult to simultaneously meet the comprehensive requirements of efficiency, reliability, and trustworthiness in multi-model collaboration scenarios. In application environments with limited data availability, model fusion methods that rely on training data or retraining processes are difficult to deploy in practice. When models have been trained and deployed across different systems or institutions, the original training data is often inaccessible due to privacy protection, compliance requirements, or storage costs, making related fusion methods infeasible in multi-center or cross-system scenarios. Furthermore, the computational resource consumption and system complexity resulting from retraining or fine-tuning models also limit the flexible application of model fusion in engineering practice.

[0008] It should be noted that the background information presented here is only for illustrating relevant information about the present invention to aid in understanding the technical solution of the present invention, and does not imply that the relevant information is necessarily prior art. The relevant information was submitted and disclosed together with the present invention, and should not be considered prior art unless there is evidence that the relevant information was disclosed before the filing date of the present invention. Summary of the Invention

[0009] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a new model fusion system and method based on interpretable information quantitative evaluation.

[0010] According to a first aspect of the present invention, a model fusion system based on interpretable information quantification evaluation is provided for fusing multiple trained models, comprising: a weight alignment predictor for processing the weight spaces of two models to be fused to obtain a weight alignment relationship between the two models, wherein the weight alignment relationship is represented as a set of permutation matrices, each element of the permutation matrix corresponding to a layer of neurons to describe the one-to-one correspondence between the two models in that neuron; a model evaluation unit for calculating the weight contribution of each parameter in the weight set of each model in the weight space according to the weight alignment relationship between the two models; and a model fusion unit for calculating the fusion ratio of the weights of each layer of each model based on the feature contribution of each parameter in the weights of each model, and fusing the weights of the two models based on the fusion ratio.

[0011] Preferably, the weight alignment predictor is pre-trained as follows: S1. Obtain the weight parameter sets of multiple trained models and construct an alignment transformation matrix for each model; S2. Rearrange the weight parameter sets of each trained model using its corresponding alignment transformation matrix, and perform a perturbation operation on the rearranged weights using a preset enhancement function to obtain new weight parameter combinations for each trained model. Combine the new weight parameter sets of each trained model with its original weight parameter sets to form a weight parameter set pair, and use its corresponding alignment transformation matrix as an alignment label to construct a sample. All samples form a training set; S3. Construct an initial weight alignment predictor; S4. Use the weight parameter pair as input and the predicted alignment arrangement matrix as output, and train the initial weight alignment predictor until convergence using the training set, wherein the parameters of the weight alignment predictor are updated with the goal of minimizing a preset loss function.

[0012] Preferably, the preset loss function is:

[0013]

[0014] in, The weighting coefficient for the loss term. Let be the supervision loss between the predicted alignment matrix and the alignment labels, and:

[0015]

[0016] in, This represents the alignment matrix corresponding to the j-th sample predicted by the alignment predictor. Let be the alignment label for the j-th sample; The alignment error loss between the weight parameter set after the alignment permutation matrix mapping transformation based on the alignment predictor and the original weight parameter set is:

[0017]

[0018] in, Represents the original set of weight parameters. express The new set of weight parameters after rearranging and perturbing their corresponding alignment transformation matrices. express The set of weight parameters after the alignment permutation matrix mapping transformation based on the alignment predictor's prediction. Represents the calculation of the square of the L2 norm; To adapt the loss to the task, and:

[0019]

[0020] This is the weighting coefficient, ranging from 0 to 1.

[0021] Preferably, the weight alignment predictor includes: a weight space encoding module, a weight mapping module, a feature relationship calculation module, and a permutation matrix generation module, wherein: the weight space encoding module encodes the original weight parameter sets of the two models to be fused to obtain weight encoding representations corresponding to the two models to be fused respectively; the weight mapping module maps the weight encoding representations of the two models to be fused to the activation space to obtain activation space feature representations corresponding to the two models to be fused respectively; the feature relationship calculation module calculates the similarity of the weight parameters of the two models to be fused in the same layer based on the activation space feature representations of the two models to be fused to obtain a feature relationship matrix; and the permutation matrix generation module maps the feature relationship matrix of the two models to be fused to the neuron alignment result to obtain an alignment permutation matrix.

[0022] Preferably, the model evaluation unit evaluates the weight contribution of each parameter in each model weight set in the weight space in the following manner:

[0023]

[0024] in, Represents the parameters in the model weight set Contribution Represents the set of model weight parameters. It does not include parameters A subset of model weight parameters This indicates adding parameters to the subset of type weight parameters. Post-performance metrics Performance metrics for a subset of representational weight parameters This represents the Shapley weight.

[0025] Preferably, the model fusion unit calculates the fusion ratio of the weights of each layer of each model in the following manner:

[0026]

[0027] in, Indicates the first The fusion ratio of layer weight parameters, Indicates the first The weight contribution of layer weight parameters. This represents the sum of the weight contributions of all layer weight parameters. This represents the total number of layers with weights, where:

[0028]

[0029] Indicates the first The total number of parameters in the layer weight parameters; whereby the model fusion unit fuses the weight parameters of two trained models in the following way:

[0030]

[0031] in, Indicates the fused first Layer weight parameters, Indicates the model to be fused The Layer weight parameters, Indicates the model to be fused The Layer weight parameters, The first term of the aligned predictor output Layered arrangement matrix, Indicates the first The transpose of the permutation matrix of the layer above the layer.

[0032] Preferably, the system is configured to perform multiple rounds of fusion on a two-model basis until fusion is complete, wherein each round of fusion involves arbitrarily selecting a model from the trained models that have never participated in fusion before, and fusing it with the model fused in the previous round.

[0033] According to a second aspect of the present invention, a model fusion method based on interpretable information quantification evaluation is provided, the method comprising: T1, acquiring multiple trained models to be fused; T2, fusing the multiple models acquired in step T1 using a system as described in the first aspect of the present invention to obtain a fused model.

[0034] Compared with existing technologies, the advantages of this invention are as follows: This application proposes to achieve effective fusion of multiple models without relying on the original training data or changing the training state of the existing models, and provides clear, quantifiable, and interpretable decision-making basis for the model fusion process. This technical solution overcomes the limitations of existing model fusion technologies in terms of limited data, heterogeneous model sources, and uninterpretable fusion processes, transforming model fusion from an experience-driven and result-oriented process into an assessable, controllable, and traceable technical process. Compared with existing technologies, this application proposes to significantly reduce the computational complexity and engineering implementation cost of the model fusion process while ensuring the performance stability of the fused model, and improve the versatility and reusability of the fusion method under different model combinations and application scenarios. Simultaneously, by introducing an interpretable information quantification evaluation mechanism, the transparency and credibility of the model fusion process are enhanced, providing feasible technical support for application scenarios with high reliability and high security requirements. Attached Figure Description

[0035] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0036] Figure 1 This is a schematic diagram illustrating the technical principle of a model fusion system based on interpretable information quantitative evaluation according to an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0038] As described in the background section, existing model fusion techniques still have significant shortcomings in terms of applicability under data-free conditions, stability of the fusion process, and interpretability of fusion decisions, making it difficult to simultaneously meet the comprehensive requirements of efficiency, reliability, and trustworthiness in multi-model collaboration scenarios. In application environments with limited data availability, model fusion methods that rely on training data or retraining processes are difficult to deploy in practice. When models have been trained and deployed across different systems or institutions, the original training data is often inaccessible due to privacy protection, compliance requirements, or storage costs, making related fusion methods infeasible in multi-center or cross-system scenarios. Furthermore, the computational resource consumption and system complexity resulting from retraining or fine-tuning models also limit the flexible application of model fusion in engineering practice.

[0039] Research on existing technologies reveals that while fusion methods based on model parameter manipulation reduce reliance on data, their stability and generalization ability are still affected by differences in model structure and inconsistent training paths. Different models exhibit variations in parameter arrangement and internal representation during training; direct parameter-level fusion can easily introduce parameter space misalignment, leading to decreased performance or instability in the fused model. Parameter alignment or structure matching techniques introduced to alleviate these issues typically require complex optimization or iterative calculations for specific model combinations, resulting in significant computational overhead and difficulty in achieving efficient reuse in scenarios with frequently changing model combinations. In practical applications, as the number of models and the complexity of application scenarios continue to increase, the main problems manifest in the following aspects:

[0040] (1) Existing model fusion methods struggle to form stable and universal fusion mechanisms when original training data is lacking. In multi-model collaboration scenarios, models are typically already trained and deployed across different systems or institutions, and the original training data cannot be obtained again due to privacy protection, compliance requirements, or practical limitations. Fusion methods that rely on training data or retraining processes are difficult to implement in such environments, while parametric fusion methods that do not rely on data have high requirements for model structure consistency and training path similarity, making it difficult to maintain stable fusion results when model sources are heterogeneous or training conditions differ significantly, and thus difficult to form a reusable and scalable fusion process.

[0041] (2) Existing model fusion methods based on parameter manipulation have high computational complexity in the model alignment stage and lack generalization ability for different model combinations. Different models have different parameter arrangements and internal representations formed during training. Direct parameter-level fusion can easily introduce parameter space misalignment, thus affecting fusion performance. Parameter alignment or structure matching techniques introduced to alleviate this problem usually require complex optimization or iterative calculations for specific model combinations. Their computational cost increases significantly with model size and structural complexity, making them difficult to apply efficiently in scenarios with a large number of models or frequent changes in model combinations. At the same time, such alignment processes are often designed for specific model pairs and are difficult to transfer or reuse to other model combinations, limiting the engineering scalability of model fusion technology.

[0042] (3) Existing model fusion techniques lack a quantitative evaluation mechanism for interpretable information about the fusion process itself. The fusion process usually involves setting hyperparameters and selecting appropriate ones from a large number of fusion results. There is no analysis of the internal information of the model, the contribution of these parameters to the fusion result is unclear, the mechanism cannot be explained, and it is difficult to provide a clear basis for fusion decisions. Existing methods generally use the overall performance of the fused model as the main evaluation index, which is difficult to characterize the specific role of different model parameters or different hierarchical structures in the fusion process, and it is also impossible to quantify the degree of influence of each component on the fusion result. In the absence of an effective evaluation mechanism, fusion strategies often rely on experience setting or manual parameter tuning, resulting in low controllability, repeatability, and reliability of fusion behavior.

[0043] Therefore, the core challenge of model fusion technology in the absence of original training data lies in the lack of effective criteria for judging the rationality of the fusion process. This problem is particularly prominent in application scenarios involving multi-model collaboration and heterogeneous model sources. When models have been trained and deployed in a distributed manner, and the fusion results cannot be re-validated through data means, relying solely on empirical rules or static parameter operations makes it difficult to determine whether different models are suitable for fusion and whether the fusion behavior is reasonable, thus hindering the further development of model fusion technology. The main reason is that viewing model fusion simply as a parameter alignment or combination problem is insufficient to meet the fusion requirements of complex model structures. Different models exhibit differences in parameter arrangements and internal representations formed during training, and the roles of different parameters or hierarchical structures in the overall model function are significantly unbalanced. In the absence of external data constraints, adopting a uniform fusion strategy for all parameters or levels often weakens the information expression ability of key structures, leading to unstable performance of the fused model. This indicates that model fusion not only needs to address the consistency of the parameter space but also the identification and differentiation of the differences in the roles of parameters during the fusion process. Furthermore, the model fusion process implicitly contains a large amount of information closely related to model behavior, which has potential value for quantification and analysis. However, effectively characterizing the impact of different structures in the fusion process within a high-dimensional parameter space without relying on training data presents significant technical challenges in terms of both computational complexity and analytical methods. On the one hand, the large scale and deep structural hierarchy of model parameters make comprehensive analysis directly impractical in engineering. On the other hand, the effects of parameters often manifest as overall behavioral effects, which are difficult to accurately describe using local or static indicators, making it less than obvious to directly introduce interpretability analysis into the model fusion process.

[0044] Therefore, the technical approach to model fusion should gradually shift from focusing on the performance of the fusion result to focusing on the assessability and interpretability of the fusion process itself. Only by introducing quantifiable evaluation criteria into the fusion process can the fusion behavior of different models and their internal structures be effectively constrained and a clear basis for fusion decisions be provided, even without access to the original training data. Based on this, this invention, after systematically analyzing the limitations of existing model fusion technologies, redefines the core technology of the model fusion problem by introducing the quantitative evaluation of interpretable information into the entire model fusion process. This results in a solution suitable for parametric-level fusion of multiple trained models without access to the original training data, providing quantifiable and interpretable evaluation and decision support for the model fusion process.

[0045] This application aims to address the key technical problem of existing model fusion technologies struggling to achieve stable, efficient, and clearly defined model fusion without original training data. It provides a universally applicable technical approach, particularly for applications with heterogeneous multi-model sources, complex fusion processes, and high reliability requirements. Specifically, the main objectives of this invention are to solve the following technical problems: how to reasonably fuse multiple trained models while ensuring controllable differences in model structure when original training data is unavailable or retraining is not possible; how to reduce the computational complexity of the model alignment process and improve the reusability of the fusion process when model parameter arrangements are inconsistent and internal representations differ significantly; and how to quantitatively analyze the roles of different model parameters and hierarchical structures in the fusion process, thereby providing a clear, interpretable, and traceable basis for fusion decisions. The present invention analyzes and constrains the model fusion process by introducing interpretable information for quantitative evaluation without accessing the original training data, and completes the implementation plan for model alignment and fusion accordingly, thereby improving the controllability, stability and reliability of the model fusion process. Specifically, it mainly uses multi-model association analysis to mine potential feature sharing patterns between models, constructs model alignment relationships in a unified weight space, and introduces a multi-scale quantitative evaluation mechanism to quantitatively analyze the role of different parameters and different hierarchical structures in the model fusion process. Thus, it achieves model fusion with process interpretability and fusion stability without accessing the original training data.

[0046] In principle, such as Figure 1 As shown, in the present invention, a weight alignment predictor is trained by training an alignment predictor, and the trained weight alignment predictor is used to perform multi-model association analysis, unified weight space mining and multi-scale quantization evaluation, and model association fusion is performed based on the evaluation results.

[0047] In summary, this invention provides a model fusion system based on interpretable information quantification evaluation for fusing multiple trained models, comprising: a weight alignment predictor for processing the weight spaces of two models to be fused to obtain a weight alignment relationship between the two models, wherein the weight alignment relationship is represented as a set of permutation matrices, each element of the permutation matrix corresponding to a layer of neurons to describe the one-to-one correspondence between the two models in that neuron; a model evaluation unit for calculating the weight contribution of each parameter in the weight set of each model in the weight space according to the weight alignment relationship between the two models; and a model fusion unit for calculating the fusion ratio of the weights of each layer of each model based on the feature contribution of each parameter in the weights of each model, and fusing the weights of the two models based on the fusion ratio.

[0048] To better understand the present invention, the various modules of the system of the present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, in the embodiments of the present invention... It is mainly used to represent a variable index and may be used repeatedly in different formulas. The specific meaning is subject to the interpretation of each formula.

[0049] I. Weight-aligned predictor

[0050] (1) Working principle of weight-aligned predictor

[0051] In this invention, a weight alignment predictor is constructed to perform model association and build a unified weight space. By analyzing the similarity of parameter structures and feature representations of different models when processing similar tasks, shared structural units that may exist between models are identified. The introduced weight alignment predictor is used to learn the mapping relationship between the weight spaces of different models, enabling the models to be fused to perform alignment operations within a unified weight space, thereby reducing fusion errors caused by differences in model structure.

[0052] In this invention, to achieve correlation modeling and fusion of different trained models in a unified weight space, a weight alignment predictor is introduced to characterize the correspondence between model weights. The alignment prediction function is defined as follows:

[0053]

[0054] in, and Let each represent the set of weight parameters for the two models to be fused. This represents the set of trainable parameters for learning the weight alignment relationships. The output of this function is a permutation matrix. This is used to describe the one-to-one correspondence between neurons in each hierarchical structure of the two models, where This represents the total number of layers in the model. This function allows for the establishment of explicit correspondences between the weights of different models without altering the original model structure, thus providing a unified operational space for subsequent fusion. Matrix P... l (l=1…m) represents the arrangement matrix of neurons in the l-th layer, used to describe the one-to-one correspondence between neurons in the two models at this layer. Its specific form is a binary or sparse matrix, where rows represent source model neurons and columns represent target model neurons, with elements P... l [i,j]=1 indicates that the i-th neuron of the source model corresponds to the j-th neuron of the target model; P l[i,j]=0 indicates no correspondence. This matrix aligns neurons with the same function in two models, enabling fusion under a unified weight space. The set of permutation matrices {P1, …, Pm} across all layers constitutes the output of the alignment predictor, ensuring clear correspondence between neurons in each layer during fusion, thereby reducing the impact of parameter space misalignment on the fusion result.

[0055] To ensure that the alignment relationship remains consistent when structural transformations occur in the weight space, the prediction function satisfies the isovariability constraint:

[0056]

[0057] in and The matrix multiplication symbol "" represents a structural transformation acting on the weight space. "and transpose operator" This function describes the corresponding changes in alignment before and after the transformation. Furthermore, to avoid biases introduced by different model input orders, it also satisfies input symmetry constraints.

[0058]

[0059] Ensuring the model alignment process maintains consistency with the input order. These two types of constraints together ensure the stability and generalization of the weight alignment predictor in multi-model fusion scenarios.

[0060] (2) Construction of the weight-aligned predictor

[0061] In this invention, to achieve neuron-level alignment of different pre-trained models in a unified weight space, an alignment prediction function is used. A weight alignment predictor is designed. Its core objective is to solve the parameter misalignment problem caused by differences in initialization and gradient paths during model training by automatically predicting the optimal permutation matrix, while providing interpretable quantitative evidence for the fusion process. Weight spatial encoding is used to extract weight structure features, activation space mapping enhances the characterization of hierarchical functions, and permutation matrix generation realizes the alignment of corresponding neurons. The three work together to form the foundation for efficient and interpretable model fusion.

[0062] According to an embodiment of the present invention, a weight alignment predictor includes: a weight space encoding module, a weight mapping module, a feature relationship calculation module, and a permutation matrix generation module, wherein: the weight space encoding module encodes the original weight parameter sets of two models to be fused to obtain weight encoding representations corresponding to the two models to be fused respectively; the weight mapping module maps the weight encoding representations of the two models to be fused to the activation space to obtain activation space feature representations corresponding to the two models to be fused respectively; the feature relationship calculation module calculates the similarity of the weight parameters of the two models to be fused in the same layer based on the activation space feature representations of the two models to be fused to obtain a feature relationship matrix; and the permutation matrix generation module maps the feature relationship matrix of the two models to be fused to the neuron alignment result to obtain an alignment permutation matrix. This can be expressed in functional form as:

[0063]

[0064] Among them, the function composite symbol " This indicates that the modules act sequentially according to a logical order, but it is not a simple serial pipeline. The modules are functionally interdependent and collaborative: First, The original weight set is encoded, structured features are extracted, and an equivariant representation is formed, providing a stable foundation for subsequent analysis; subsequently... Mapping the weight space representation to the activation space, supplementing it with neuron bias and hierarchical characteristic information, ensures that the alignment process considers not only weight values ​​but also functional representation; then... Feature relationships between neurons are computed in the activation space, and a hierarchical similarity matrix is ​​generated through generalized outer product to quantify the degree of correspondence between different models at the same layer; finally... By mapping the similarity matrix to the permutation matrix and selecting the optimal neuron alignment scheme, accurate alignment of different models in a unified weight space can be achieved.

[0065] Each module will be described below.

[0066] First, let's introduce the weight space encoding module. .

[0067] Weight Space Encoding Module Used for the original weight set and Encode the corresponding weighted representation. and To extract the structured features contained in the weights, the mapping relationship is expressed as follows:

[0068]

[0069] in, This represents the weight encoding function, used to extract structured features from the weights and map the original weights to a unified, comparable, and aligned representation space. This encoding process includes necessary weight space transformations, such as those applied to the weights of convolutional layer filters. L2 normalization and row or column rearrangement are performed to place neurons or filters with similar functions at the same index position. This type of transformation does not change the original function of the model, but only adjusts the parameter arrangement and representation order, which is used to eliminate the arrangement differences between models and provide stable input for the aligned predictor.

[0070] To ensure that the encoding result maintains a consistent correspondence with the original weight structure, the weight encoding function satisfies the equivariance constraint:

[0071]

[0072] in, and This represents a weight space transformation operation. During the encoding process, weight space transformations (such as linear mapping, normalization, and row / column rearrangement) are applied to the weights of convolutional filters or fully connected layers, causing neurons with similar functions to be located at the same index position. This transformation does not change the model's functionality; it only adjusts the parameter arrangement and representation order, providing a stable foundation for subsequent activation mapping and feature relationship calculations, while reducing computational complexity.

[0073] Secondly, the weight mapping module is introduced. .

[0074] Weight mapping module This is used to map weights to the activation space based on weight encoding to enhance the characterization of hierarchical features:

[0075]

[0076] in, These represent the activation space features of the two models, respectively. This represents the bias term of the corresponding neuron. By introducing activation information, the predictor considers both weight relationships and functional differences between neurons at different levels during the alignment process, providing a richer hierarchical representation for feature relationship calculation.

[0077] Third, the feature relationship calculation module is introduced. .

[0078] Feature Relationship Calculation Module To introduce a generalized outer product operation in the activation space to characterize the correspondence between different models, in this embodiment of the invention, the activation space feature relationship calculation function is defined as follows:

[0079]

[0080] in, This represents a direct sum combination of results from different levels. Indicates the first The number of neurons in a layer is calculated as follows:

[0081]

[0082] These represent the two models at the [missing information] in the [missing information] year. The first in the layer The and the first Each corresponding weight feature and These represent the two models at the [missing information] in the [missing information] year. The first in the layer The and the first The activation characteristics of each corresponding neuron This is a symmetric similarity measure function. This operation is used to generate similarity matrices for each level, thereby quantifying the structural similarity of different models at the same level. According to one embodiment of the present invention, the similarity function is defined as:

[0083]

[0084] Among them, the formula It only represents two elements, unlike the previous formula. There is no direct connection. This represents the vector dot product operation. Represents the L2 norm, This is a learnable scaling parameter used to adjust the numerical range of the similarity metric. Through this similarity calculation, neuronal structures with high correspondence in the same layer of different models can be effectively identified.

[0085] Fourth, the permutation matrix generation module is introduced. .

[0086] Permutation matrix generation module Used to map the feature relation matrix to neuron-level alignment results This matrix is ​​the alignment predictor. The inference output.

[0087] (3) Training of the weight-aligned predictor

[0088] The core of training a weight-aligned predictor is optimizing the predictor parameters. This enables the generator to produce high-quality permutation matrices under different input weights, achieving neuron alignment without training data. According to one embodiment of the invention, the weight alignment predictor is trained through the following steps:

[0089] S1. Obtain the weight parameter set of multiple trained models and construct an alignment transformation matrix for each model.

[0090] S2. The weight parameter set of each trained model is rearranged using its corresponding alignment transformation matrix, and a pre-defined enhancement function is used to perturb the rearranged weights to obtain new weight parameters for each trained model. The new weight parameter set and its original weight parameter set of each trained model are combined to form a weight parameter set pair, and its corresponding alignment transformation matrix is ​​used as the alignment label to construct a sample. All samples constitute the training set; that is, the training samples are generated from the pre-trained task model weights and combined with a manually set alignment transformation matrix. and perturbation enhancement Constructing a labeled sample set:

[0091]

[0092] in, This represents the weight vector sampled from the pre-trained model. This is an artificially constructed alignment transformation matrix used to simulate the differences in neuron arrangement between different models. Based on this, it is first transformed... The weights are rearranged and then passed through an enhancement function. By introducing perturbations, a new weight representation is obtained. To improve the model's robustness to different permutations and noise, a sample set is constructed accordingly. ,in As a supervisory label, it is used to guide the alignment predictor in learning the mapping relationship between weight spaces.

[0093] S3. Construct the initial weight-aligned predictor.

[0094] S4. Using the alignment matrix of the input and prediction weight parameters as output, train the initial weight alignment predictor until convergence using the training set. The parameters of the weight alignment predictor are updated with the objective of minimizing a preset loss function. The preset loss function is:

[0095]

[0096] in, The weighting coefficient for the loss term. Let be the supervision loss between the predicted alignment matrix and the alignment labels, and:

[0097]

[0098] in, This represents the alignment matrix corresponding to the j-th sample predicted by the alignment predictor. Let be the alignment label for the j-th sample;

[0099] The alignment error loss between the weight parameter set after the alignment permutation matrix mapping transformation based on the alignment predictor and the original weight parameter set is:

[0100]

[0101] in, Represents the original set of weight parameters. express The new set of weight parameters after rearranging and perturbing their corresponding alignment transformation matrices. express The set of weight parameters after the alignment permutation matrix mapping transformation based on the alignment predictor's prediction. Represents the calculation of the square of the L2 norm; To adapt the loss to the task, and:

[0102]

[0103] This is the weighting coefficient, ranging from 0 to 1.

[0104] Convergence is determined by monitoring the trend of the loss function and the stability of the predictor output. When the combined loss stabilizes over several training rounds and its decrease is below a preset threshold, and the permutation matrix generated on the validation samples shows little change while the performance of the fusion model shows no significant fluctuations, the alignment predictor can be considered to have converged. Through this training process, the weight alignment predictor learns the ability to map weights and activation features to the optimal permutation matrix. (Permutation matrix) Generated only during the inference phase, it is used for actual neuron-level alignment and serves as input for multi-scale quantization evaluation and adaptive fusion, ensuring that key features are preserved and redundant features are controlled, thereby improving the stability, interpretability and overall performance of the fusion model.

[0105] During the inference phase, the permutation matrix generation module generates the feature relation matrix. Mapped to neuron-level alignment results:

[0106]

[0107] in, Indicates the first The set of all possible permutation matrices for the layer. It is worth noting that... It is a predictor The output result indicates that the training objective is not to directly optimize this formula, but rather to optimize the predictor parameters through the loss function. This enables it to accurately generate the permutation matrix during the reasoning stage.

[0108] II. Model Evaluation Unit

[0109] After model alignment is completed, a multi-scale quantization evaluation mechanism is introduced to analyze the role of different parameters in the model fusion process. To this end, a parameter contribution evaluation function is defined:

[0110]

[0111] in, Represents the feature set of model parameters. This represents a single parameter characteristic. This indicates a subset that does not contain this feature. This represents the expectation operation. This function characterizes the marginal impact of different parameters on the alignment result, thus providing a quantitative basis for fusion decisions. Based on this, the calculation form of feature contribution is introduced:

[0112]

[0113] This means that after model alignment is completed, a multi-scale quantitative evaluation mechanism is introduced to analyze the role of different parameters in the model fusion process, so as to provide a quantitative basis for fusion decision-making. Let be a set of parameter features for a certain layer, where a single parameter feature is: The contribution of this parameter is calculated by examining all subsets that do not contain this feature. Poor fusion performance before and after addition get.

[0114] in, This indicates that only the fusion result metrics (such as accuracy or AUC) of a subset S are used, while This means adding to subset S. The performance metrics after the parameter are determined. To fairly quantify the marginal contribution of each parameter, Shapley weights are introduced. The parameter-level contribution is obtained by weighted summation over all subsets. Furthermore, the hierarchical contribution can be calculated by summing the parameters of the entire layer. :

[0115]

[0116] Indicates the first The total number of parameters in the layer weight parameters is normalized at multiple levels and the model level to achieve multi-scale quantitative analysis. This guides the adaptive weight allocation of different levels and parameters during the fusion process, ensuring that key features are preserved and the influence of weakly correlated or redundant features is reasonably controlled, thereby improving the stability, interpretability and overall performance of the fusion model.

[0117] III. Model Fusion Unit

[0118] Based on the preceding alignment predictions and evaluations, interpretable and guided model fusion is performed. Specifically, based on the contribution analysis results described above, the fusion ratios at different levels of the model are adaptively adjusted. Let the... The contribution of layer parameters is The corresponding fusion ratio is defined as follows:

[0119]

[0120] in, This represents the total number of layers in the model. Under a unified weight space and alignment constraints, the fused weight matrix is ​​expressed as:

[0121] in, Indicates the fused first Layer weight parameters, Indicates the model to be fused The Layer weight parameters, Indicates the model to be fused The Layer weight parameters, The first term of the aligned predictor output Layered arrangement matrix, Indicates the first The transpose of the permutation matrix of the previous layer is used to rearrange the input dimensions of the weights of the current layer in model B, thereby ensuring that the connection structure between adjacent layers remains consistent after alignment. This fusion expression allows layers that contribute more to the fusion result to receive higher weights while maintaining model structural consistency, thus achieving interpretable and structurally controllable model fusion.

[0122] The above embodiment describes how to fuse two models. However, in actual application scenarios, the fusion of multiple models is involved. The specific implementation method is as follows: take two models as the basic unit and perform multiple rounds of fusion. First, fuse any two trained models using the scheme of this invention to obtain a fused model. Then, fuse this fused model with another unfused model. Repeat this process multiple times until the fusion is completed.

[0123] In summary, this solution addresses the challenges of direct fusion of trained models in multi-center collaborative scenarios, as well as the lack of interpretability and controllability in the fusion process. It proposes a model association fusion scheme based on multi-scale quantitative evaluation. This scheme does not rely on original training data or require retraining of existing models. Instead, it constructs relationships between models, aligns them within a unified weight space, and introduces a quantitative evaluation mechanism for the contribution of model parameters. This explicitly constrains key structures and parameters in the fusion process, thereby guiding the formulation of the fusion strategy.

[0124] This approach transforms the model fusion process from an experience-driven process into an analyzable, evaluable, and traceable engineering process. By conducting multi-scale quantitative analysis of the role of different level parameters in the fusion behavior, this approach can rationally allocate fusion weights for different models and different levels while ensuring model structural consistency and fusion stability. This avoids the loss of key features or interference from irrelevant structures, thereby improving the reliability and generalization ability of the fused model.

[0125] Therefore, this solution provides a general, scalable technical approach with clear engineering constraints for model fusion in a multi-model collaborative environment. It is particularly suitable for model collaboration application scenarios among multiple institutions and systems. Under the premise of ensuring data and model security, it effectively improves the efficiency of model collaboration and the quality of fusion, and has clear engineering value and application prospects.

[0126] Compared with the prior art, the solution of the present invention has the following characteristics:

[0127] First, this invention provides a technical mechanism for achieving model fusion without requiring original training data. This application achieves the fusion of multiple trained models by utilizing the models' own information without accessing the original training data or retraining or fine-tuning the models. This avoids dependence on external data and enables model fusion to be implemented in scenarios where data is limited, privacy is sensitive, or models are already deployed. It significantly reduces the dependence of model fusion on data sharing and computing resources, improving the feasibility and deployment flexibility of model fusion in multi-center collaboration and engineering applications.

[0128] Secondly, this invention proposes a reusable parameter alignment and fusion control mechanism addressing model structural differences. This invention proposes a fusion control mechanism to handle inconsistencies in parameter arrangement and structural representation between different models, ensuring model comparability before fusion and thus reducing the impact of parameter misalignment on the fusion results. By effectively constraining model structural differences during the fusion process, it avoids repeatedly performing highly complex processing for each group of models, reducing the computational complexity of the model fusion process, improving the stability and reusability of the fusion process under different model combinations, and enhancing the engineering scalability of the model fusion method.

[0129] Third, this invention proposes a fusion decision support mechanism based on interpretable information quantitative evaluation. By introducing interpretable information quantitative evaluation into the entire model fusion process, this invention characterizes the differences in the roles of different model parameters and hierarchical structures during the fusion process, thereby providing a clear basis for fusion decisions. This transforms the model fusion process from experience-driven to an assessable and traceable technical process, improving the interpretability and controllability of the model fusion process, enhancing the credibility and stability of the fusion results, and meeting the practical needs of application scenarios with high model reliability requirements.

[0130] In summary, this application proposes an effective fusion of multiple models without relying on the original training data or altering the existing model training state, and provides clear, quantifiable, and interpretable decision-making criteria for the model fusion process. This technical solution overcomes the limitations of existing model fusion technologies, such as limited data, heterogeneous model sources, and the lack of interpretability in the fusion process. It transforms model fusion from an experience-driven and result-oriented process into an assessable, controllable, and traceable technical process. Compared to existing technologies, this application significantly reduces the computational complexity and engineering implementation costs of the model fusion process while ensuring the performance stability of the fused models, and improves the versatility and reusability of the fusion method under different model combinations and application scenarios. Furthermore, by introducing an interpretable information quantification evaluation mechanism, it enhances the transparency and credibility of the model fusion process, providing feasible technical support for application scenarios with high reliability and security requirements.

[0131] This invention is applicable to multi-model collaboration and model reuse scenarios, particularly suitable for situations where models have been trained but the original training data is unavailable or unsuitable for sharing, necessitating the fusion of multiple trained models. In such scenarios, models typically originate from different systems, institutions, or training processes, exhibiting differences in parameter arrangement, structural representation, and internal representation. Furthermore, the fusion results cannot be verified or corrected through retraining or data-driven methods. The main challenges of model fusion in these scenarios are: lacking original training data as an external constraint, it is difficult to determine whether different models are suitable for fusion, and the extent to which different model parameters or hierarchical structures influence the fusion results; simultaneously, differences in model structure make direct parameter fusion prone to introducing instability, while existing fusion methods lack quantitative evaluation and interpretation mechanisms for the fusion process itself, making fusion decisions reliant on experience and hindering the formation of stable, reusable technical processes.

[0132] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0133] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A model fusion system based on interpretable information-based quantitative evaluation, used to fuse multiple trained models, characterized in that, include: The weight alignment predictor is used to process the weight space of the two models to be fused to obtain the weight alignment relationship between the two models. The weight alignment relationship is represented as a set of permutation matrices, where each element of the permutation matrix corresponds to a neuron of a layer to describe the one-to-one correspondence between the two models in that neuron. The model evaluation unit is used to calculate the weight contribution of each parameter in the weight space of each model's weight set based on the weight alignment relationship between the two models. The model fusion unit calculates the fusion ratio of each model's weights at each layer based on the feature contribution of each parameter in each model's weights, and then fuses the weights of the two models based on the fusion ratio.

2. The model fusion system based on interpretable information quantitative evaluation according to claim 1, characterized in that, The weight alignment predictor is pre-trained in the following manner: S1. Obtain the weight parameter set of multiple trained models and construct an alignment transformation matrix for each model; S2. The weight parameter set of each trained model is rearranged using its corresponding alignment transformation matrix, and the rearranged weights are perturbed using a preset enhancement function to obtain the new weight parameter combination for each trained model. The new weight parameter set of each trained model and its original weight parameter set are combined to form a weight parameter set pair, and its corresponding alignment transformation matrix is ​​used as the alignment label to construct a sample. All samples form the training set. S3. Construct the initial weight-aligned predictor; S4. Using the alignment matrix of the input and prediction weight parameters as the output, train the initial weight alignment predictor until convergence using the training set, wherein the parameters of the weight alignment predictor are updated with the goal of minimizing the preset loss function.

3. The model fusion system based on interpretable information quantitative evaluation according to claim 2, characterized in that, The preset loss function is: in, The weighting coefficient for the loss term. Let be the supervision loss between the predicted alignment matrix and the alignment labels, and: in, This represents the alignment matrix corresponding to the j-th sample predicted by the alignment predictor. Let be the alignment label for the j-th sample; The alignment error loss between the weight parameter set after the alignment permutation matrix mapping transformation based on the alignment predictor and the original weight parameter set is: in, Represents the original set of weight parameters. express The new set of weight parameters after rearranging and perturbing their corresponding alignment transformation matrices. express The set of weight parameters after the alignment permutation matrix mapping transformation based on the alignment predictor's prediction. Represents the calculation of the square of the L2 norm; To adapt the loss to the task, and: This is the weighting coefficient, ranging from 0 to 1.

4. The model fusion system based on interpretable information quantitative evaluation according to claim 3, characterized in that, The weight alignment predictor includes: a weight space encoding module, a weight mapping module, a feature relation calculation module, and a permutation matrix generation module, wherein: The weight space encoding module encodes the original weight parameter sets of the two models to be fused to obtain the weight encoding representations corresponding to the two models to be fused respectively; The weight mapping module is used to map the weight encoding representations of the two models to be fused to the activation space to obtain the activation space feature representations of the two models to be fused respectively; The feature relationship calculation module is used to calculate the similarity of the weight parameters of the two models to be fused at the same layer based on the activation space feature representation of the two models to be fused, so as to obtain the feature relationship matrix. The permutation matrix generation module is used to map the feature relationship matrix of two models to be fused to the neuron alignment result to obtain the alignment permutation matrix.

5. The model fusion system based on interpretable information quantitative evaluation according to claim 4, characterized in that, The model evaluation unit evaluates the weight contribution of each parameter in each model weight set in the weight space in the following way: in, Represents the parameters in the model weight set Contribution Represents the set of model weight parameters. It does not include parameters A subset of model weight parameters This indicates adding parameters to the subset of type weight parameters. Post-performance metrics Performance metrics for subsets of representational weight parameters This represents the Shapley weight.

6. The model fusion system based on interpretable information quantitative evaluation according to claim 5, characterized in that, The model fusion unit calculates the fusion ratio of each model's weights for each layer in the following way: in, Indicates the first The fusion ratio of layer weight parameters, Indicates the first The weight contribution of layer weight parameters. This represents the sum of the weight contributions of all layer weight parameters. This represents the total number of layers with weights, where: Indicates the first The total number of parameters in the layer weight parameters; The model fusion unit fuses the weight parameters of two trained models in the following way: in, Indicates the fused first Layer weight parameters, Indicates the model to be fused The Layer weight parameters, Indicates the model to be fused The Layer weight parameters, The first term of the aligned predictor output Layered arrangement matrix, Indicates the first The transpose of the permutation matrix of the layer above the layer.

7. The model fusion system based on interpretable information quantitative evaluation according to claim 6, characterized in that, The system is configured to perform multiple rounds of fusion on a two-model basis until fusion is complete. In each round of fusion, a model is randomly selected from the trained models that have never participated in the fusion and fused with the model after the previous round of fusion.

8. A model fusion method based on interpretable information-based quantitative evaluation, characterized in that, The method includes: T1. Obtain multiple trained models to be fused; T2. The system described in any one of claims 1-7 is used to fuse the multiple models obtained in step T1 to obtain a fused model.

9. A computer device / equipment / system, comprising a memory, a processor, and computer programs / instructions stored in the memory, characterized in that, The processor executes the computer program / instructions to implement the steps of the method of claim 8.

10. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method of claim 8.

Images