A cross-domain fusion general field large model self-adaptive intelligent optimization method and system

By combining the FiLM generator and singular value decomposition with tensor product B-spline interpolation, the problem of insufficient adaptability of generalized large models in different domains is solved. This method achieves efficient cross-domain feature modulation and adaptive optimization in unknown domains, thereby improving the model's cross-domain generalization performance and dynamic adaptability.

CN123070184APending Publication Date: 2026-11-10UNIV OF SCI & TECH BEIJING
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611078657.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-20
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing methods struggle to enable large, cross-domain models to adaptively adjust feature representations, exhibit low cross-domain adaptation efficiency, and lack parameter generation and optimization mechanisms for unknown domains, thus limiting the model's adaptability in open and dynamic environments.

Method used

By using the FiLM generator for conditional encoding, combined with singular value decomposition and tensor product B-spline interpolation, the intermediate features are reduced in order and interpolated in parameter space to generate target projection coefficients. Furthermore, unknown domain encoding parameters are generated using known domain encoding data, thus optimizing the network weights of the FiLM generator.

Benefits of technology

It significantly improves cross-domain adaptation efficiency, enhances the model's ability to autonomously generate and optimize parameters in open and dynamic environments, and improves cross-domain generalization performance and dynamic adaptability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070184A_ABST
    Figure CN123070184A_ABST
Patent Text Reader

Abstract

The application provides a cross-domain fusion general field large model self-adaptive intelligent optimization method and system, and relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining input information of different field target tasks; performing conditional coding on the input information; performing order reduction processing on the intermediate features of the general field large model; determining the channel mapping relationship of the order reduction features in the parameter space; performing parameter space interpolation operation on the channel mapping features to obtain target projection coefficients; performing feature reconstruction on the target projection coefficients; mapping the reconstructed features to the original feature space of the general field large model to generate the output result of the target task; judging whether the target field corresponding to the target task is contained in the preset field set; if yes, taking the encoding data corresponding to the target field as the target encoding parameter; otherwise, generating the target encoding parameter according to the encoding data corresponding to the known field; and optimizing the general field large model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a cross-domain integrated, broad-domain adaptive intelligent optimization method and system for large models. Background Technology

[0002] With the rapid development of artificial intelligence technology, large-scale cross-domain models have demonstrated powerful versatility in downstream tasks across various fields such as computer vision, natural language processing, and speech recognition. In practical applications, data distribution varies significantly across different domains, and target tasks often exhibit highly diverse characteristics. How to efficiently and flexibly enable models to adaptively handle tasks from different domains without altering their pre-training parameters, while maintaining low computational overhead, has become a core challenge in current research on cross-domain generalization and adaptive model optimization.

[0003] Currently, the common approach is to use parameter-efficient fine-tuning methods to adapt large, general-domain models to specific domain tasks. This is achieved by inserting bypass modules into specific layers of the pre-trained model and updating these small-scale parameters on target domain data, thereby achieving adaptation to specific downstream tasks while freezing the pre-trained weights.

[0004] However, existing methods make it difficult for models to adaptively adjust feature representations for different domain tasks, resulting in low cross-domain adaptation efficiency. Furthermore, existing methods lack parameter generation and optimization mechanisms for unknown domains, limiting the model's ability to adapt to unknown domains in open and dynamic environments. Summary of the Invention

[0005] To address the technical challenges of existing methods that make it difficult for models to adaptively adjust feature representations for different domain tasks, resulting in low cross-domain adaptation efficiency, and the lack of parameter generation and optimization mechanisms for unknown domains, which limits the model's adaptive capabilities in open and dynamic environments, this invention provides a cross-domain fusion-based adaptive intelligent optimization method and system for large-scale models.

[0006] The technical solutions provided by the embodiments of the present invention are as follows: The first aspect of this invention provides a cross-domain fusion adaptive intelligent optimization method for a broad-domain large model, comprising: S1: Obtain input information for target tasks in different domains; S2: The input information is conditionally encoded using a FiLM generator to obtain encoded data; S3: Reduce the order of intermediate features in the general domain model to obtain reduced-order feature bases and multiple reduced-order features; S4: Based on the encoded data, determine the channel mapping relationship of the reduced-order features in the parameter space to obtain the channel mapping features; S5: Perform parameter space interpolation on the channel mapping features to obtain the target projection coefficients corresponding to the target task; S6: Based on the reduced-order feature basis, feature reconstruction is performed on the target projection coefficients to obtain the reconstructed features; S7: Map the reconstructed features to the original feature space of the general domain large model to obtain adaptive optimized features, and generate the output results of the target task based on the adaptive optimized features; S8: Based on the input information, determine whether the target domain corresponding to the target task is included in the preset domain set; if so, use the encoded data corresponding to the target domain as the target encoding parameter; otherwise, generate the target encoding parameter based on the encoded data corresponding to the known domain. S9: Optimize large-scale models across a broad domain based on target encoding parameters.

[0007] A second aspect of this invention provides a cross-domain fusion, general-domain large-scale model adaptive intelligent optimization system, comprising: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement the cross-domain fusion pan-domain large model adaptive intelligent optimization method as described in the first aspect.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, by introducing a FiLM generator to conditionally encode the input information and dynamically adjust the reduced-order features, adaptive feature modulation for tasks in different domains is achieved, significantly improving cross-domain adaptation efficiency. At the same time, by determining whether the target domain is included in a preset domain set, if it is an unknown domain, target encoding parameters are generated by combining known domain encoding data, and the FiLM generator is optimized in reverse based on these parameters. This endows the model with the ability to autonomously generate and continuously optimize parameters for unknown domains in an open and dynamic environment, effectively enhancing the cross-domain generalization performance and dynamic adaptability of the generalized large-scale model. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating a cross-domain fusion adaptive intelligent optimization method for a large-scale model with broad domain characteristics, provided in an embodiment of the present invention.

[0011] Figure 2This is a schematic diagram illustrating the adaptive optimization convergence of the FiLM generator under the supervision of target encoding parameters, as provided in an embodiment of the present invention.

[0012] Figure 3 This is a schematic diagram of the structure of a cross-domain fusion general-domain large-model adaptive intelligent optimization system provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0014] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0015] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0016] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0017] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0018] Reference manual attached Figure 1 The diagram illustrates a flowchart of a cross-domain fusion adaptive intelligent optimization method for a large-scale model with broad domain integration, provided by an embodiment of the present invention.

[0019] This invention provides a cross-domain fusion adaptive intelligent optimization method for a broad-domain large model. This method can be implemented by a cross-domain fusion adaptive intelligent optimization device, which can be a terminal or a server. The processing flow of the cross-domain fusion adaptive intelligent optimization method for a broad-domain large model may include the following steps: S1: Obtain input information for target tasks in different domains.

[0020] Optionally, the input information specifically includes: image feature vectors, text feature vectors, speech feature vectors, and task labels.

[0021] Specifically, upon receiving a target task request, the system first determines the data source, task type, and target domain corresponding to the target task. Data sources include, but are not limited to, image acquisition devices, text databases, voice acquisition devices, business system interfaces, publicly available internet data sources, and pre-built training sample libraries. Based on the data type of the target task, input data is obtained from the corresponding data source.

[0022] It should be noted that preprocessing of the input data is included after step S1. Dimensional alignment is performed on the feature vectors of different modalities. For missing modal data, pre-defined zero vectors, learnable mask vectors, or historical domain mean vectors are used to fill in the gaps. For modal features of different dimensions, linear mapping layers, multilayer perceptrons, or attention mapping layers are used to map them to a unified dimensional space. Each modal feature is concatenated with the task label, weighted fusion, or attention fusion is performed. This ensures that input data from different domains, sources, and modalities can be converted into input information in a unified format, thus serving as the data foundation for subsequent FiLM generator conditional encoding, intermediate feature reduction processing, and adaptive intelligent optimization.

[0023] S2: The input information is conditionally encoded using the FiLM generator to obtain encoded data.

[0024] Here, FiLM stands for Feature-Level Linear Modulation, and the FiLM generator is a neural network module used to generate scaling and offset parameters based on input conditions. The scaling and offset parameters are used to perform a linear transformation on each channel of the intermediate features of the neural network.

[0025] Specifically, the preprocessed input information is fed into the FiLM generator. The FiLM generator first maps the input information into a global conditional embedding vector, and then, based on this conditional embedding vector, it independently outputs two key parameters, scaling parameter and offset parameter, for each feature channel in the neural network.

[0026] It should be noted that the scaling parameter is used to control the activation intensity of the corresponding feature channel (increase or decrease), while the offset parameter is used to adjust the baseline level of the channel (increase or decrease the offset).

[0027] Furthermore, the scaling parameter vector, offset parameter vector, and task or domain conditional embedding vector generated for each channel are concatenated to form complete encoded data. The conditional embedding vector is used to represent the basic feature representation of the current input information in the global semantic space.

[0028] In this embodiment of the invention, the FiLM generator maps the multimodal task vector to independent scaling and offset parameters for each feature channel, enabling the model to dynamically adjust the intermediate features of the neural network based on the input information. Complex feature modulation can be achieved with only two parameters per channel, resulting in low computational overhead.

[0029] S3: Reduce the order of intermediate features in the general domain model to obtain reduced-order feature bases and multiple reduced-order features.

[0030] Among them, a broad-domain large model refers to a large-scale pre-trained model capable of handling multiple different domains and modalities. The core feature is "broad-domain," meaning that the model's training data covers a wide range of knowledge domains and diverse task types, enabling the model to possess cross-domain general knowledge representation capabilities and transfer learning capabilities.

[0031] It should be noted that a general-domain large model can be a multimodal pre-trained large model, such as a cross-modal model that processes images, text and speech simultaneously, or a general-purpose large language model that covers a variety of downstream natural language processing tasks.

[0032] In one possible implementation, S3 specifically includes sub-steps S301 to S303: S301: Construct a training feature matrix based on the intermediate features of the general domain large model.

[0033] Specifically, intermediate features (such as activation values ​​of each layer) are extracted from the training data of the general domain large model. For each training sample, its intermediate features are flattened into vectors, arranged into a matrix according to the samples, to obtain the training feature matrix. ,in, Representing feature dimension, Indicates the number of samples.

[0034] S302: Perform singular value decomposition on the training feature matrix to obtain the singular value decomposition result, which specifically includes the reduced-order feature basis, the singular value matrix, and the right singular vector.

[0035] Singular Value Decomposition (SVD) is a matrix factorization technique that decomposes an arbitrary real matrix into the product of three specific matrices.

[0036] Specifically, the singular value decomposition (SVD) of the feature matrix is ​​calculated using the following formula:

[0037] in, Represents the training feature matrix, Describes a left singular vector matrix. Represents a singular value matrix. Represents a right singular vector matrix, with subscripts... This indicates the transpose operation.

[0038] It should be noted that after performing singular value decomposition on the training feature matrix, a set of left singular vectors arranged in descending order of their corresponding singular values ​​is obtained. The magnitude of each singular value represents the energy proportion of that direction in the original data. To preserve the main information while reducing dimensionality, an energy retention threshold needs to be set, and then left singular vectors whose sum of the first few singular values ​​reaches the energy retention threshold are selected. These selected left singular vectors together constitute the reduced-order feature basis.

[0039] It should be noted that those skilled in the art can set the energy retention threshold according to actual needs, and this invention does not limit it.

[0040] In this embodiment of the invention, singular value decomposition can decompose the high-dimensional training feature matrix into orthogonal bases arranged in descending order of energy proportion. By retaining the left singular vectors corresponding to the first few largest singular values, the optimal low-rank reduced-order feature base is constructed. This significantly reduces the feature dimension while preserving the main information of the original data to the maximum extent, thereby reducing subsequent computational overhead.

[0041] S303: Perform order reduction processing on the singular value decomposition results to obtain multiple reduced-order features.

[0042] Specifically, the original intermediate feature vector of each training sample in the training feature matrix is ​​projected onto the reduced-order feature basis. The inner product of the original feature vector and each basis vector is then calculated to obtain a feature base of length [missing information]. r The reduced-order eigenvectors, where, r This represents the number of basis vectors contained in the reduced-order feature basis, which is the feature dimension retained after dimensionality reduction. Finally, the reduced-order feature vectors of all samples are arranged in order to form the reduced-order feature matrix, obtaining each reduced-order feature, thus completing the transformation from high-dimensional to low-dimensional and achieving feature dimensionality reduction.

[0043] In this embodiment of the invention, by retaining the top principal components with the largest energy proportions, the feature dimensionality is significantly reduced while preserving the main information of the original data to the maximum extent. This effectively reduces the computational overhead and storage requirements of subsequent FiLM adjustment and interpolation prediction. Simultaneously, the reduction process removes redundant and noisy information from the original features, resulting in better orthogonality and compactness of the reduced features. This provides a more stable and discriminative feature representation for the channel selection and mapping in the subsequent S4 step, thereby improving the overall model's operating efficiency and generalization ability.

[0044] S4: Based on the encoded data, determine the channel mapping relationship of the reduced-order features in the parameter space to obtain the channel mapping features.

[0045] In one possible implementation, S4 specifically includes sub-steps S401 and S402: S401: Determine the target active feature channel index based on the scaling and offset parameters in the encoded data.

[0046] Specifically, the scaling parameter vector and offset parameter vector output by the FiLM generator in step S2 are extracted. For each feature channel in the reduced-order feature, the combined significance score of its corresponding scaling parameter and offset parameter is calculated.

[0047] For example, the overall saliency score can be obtained by summing the absolute values ​​of the scaling parameter and the offset parameter. Alternatively, a weighted summation method can be used. After calculating the saliency score for all channels, it is compared with a preset saliency threshold. Channels with saliency scores below the threshold (these channels will be suppressed or turned off under the current target task) are filtered out, while channels with saliency scores above the threshold are retained. The set of numbers corresponding to the retained channels is determined as the target activation feature channel index.

[0048] S402: Based on the target activation feature channel index, channel filtering and recombination are performed on the reduced-order features to obtain channel mapping features.

[0049] Specifically, based on the target activated feature channel index, the reduced-order features generated in step S3 are sliced ​​and filtered along the channel dimension. Along the feature channel dimension, only the feature channel components that match the target activated feature channel index are extracted from the reduced-order features, discarding inactive redundant channel components. Subsequently, the extracted effective feature channels are concatenated and reorganized according to a preset topological order to generate a more concise channel mapping feature that contains only features strongly correlated with the current target task. This feature is then used as input data for parameter space interpolation in the subsequent step S5.

[0050] In this embodiment of the invention, "on-demand extraction" of the feature space is achieved by filtering and reorganizing channels based on the saliency of the FiLM encoding parameters. Since the scaling and offset parameters of FiLM directly reflect the importance of each feature channel to the current task, using them as index priors can effectively eliminate noisy channels and redundant information irrelevant to the current task in the reduced-order features. This not only further reduces the computational and storage overhead of subsequent parameter space interpolation operations, but also allows the interpolation process to focus on the most discriminative key feature subspace, thereby improving the accuracy of the target projection coefficient vector and the overall running efficiency of the model.

[0051] S5: Perform parameter space interpolation on the channel mapping features to obtain the target projection coefficients corresponding to the target task.

[0052] In one possible implementation, S5 specifically includes sub-steps S501 to S503: S501: Determine the target position of the channel mapping feature in the preset parameter space based on the target task parameters of the channel mapping feature.

[0053] Among them, the target task parameter refers to the quantitative indicator used to uniquely identify and describe the position of a specific task in its parameter space.

[0054] Specifically, by mapping the input continuous target task parameter values ​​to a preset parameter space grid coordinate system, the specific coordinate point of the task parameter in space can be determined as the target location.

[0055] It should be noted that, since the preset parameter space is usually divided into multiple subdomains based on the nonlinear characteristics of the model or resolution requirements, determining the target location also includes determining which specific subdomain the parameter coordinates belong to. This is achieved by comparing the target task parameters with the boundaries of the preset subdomains, thereby determining the parameter space patch in which the point lies, so that interpolation operations can be performed in the mesh of that specific subdomain subsequently.

[0056] S502: Generate projection coefficients of multiple training samples that are close to the target location using a pre-built parameter space interpolation model.

[0057] Among them, the parameter space interpolation model refers to the tensor product spline interpolation model (TPI) constructed based on the Cartesian product grid of physical parameters, which uses the cubic B-spline basis function for local smooth approximation.

[0058] Among them, the B-spline basis function is a piecewise polynomial function used to construct smooth curves and surfaces, and it is the basic building block in spline interpolation methods.

[0059] Among them, the training sample projection coefficients refer to the weight vectors obtained by projecting the high-dimensional feature data of the training samples onto the reduced-order feature basis obtained through singular value decomposition.

[0060] Specifically, based on the target location, the parameter space interpolation model locates the cells containing the target parameters in the grid, identifies the neighboring training sample grid nodes that have local support for the target location, and then extracts the training sample projection coefficients corresponding to these neighboring nodes from the pre-computed and stored three-dimensional coefficient tensor, which serves as the data basis for the weighted calculation of the target projection coefficients in subsequent steps.

[0061] S503: By using the basis function weights corresponding to the projection coefficients of multiple training samples, the projection coefficients of each training sample are weighted and calculated to obtain the target projection coefficients corresponding to the target task parameters.

[0062] Specifically, B-spline basis functions are used as weight functions to perform weighted interpolation on the projection coefficients of neighboring nodes. The calculation formula is as follows:

[0063] in, The tensor product operation representing one-dimensional basis functions. Indicated by control coefficient The constructed tensor product interpolation function, , as well as These represent the first, second, and third core condition parameters of the target task, respectively. j , k as well as l These represent the three core condition parameters in the three-dimensional parameter space. The index of the grid node within the local support domain in the direction of ) , as well as These represent the corresponding grid points in the three parameter dimensions. Represents parameter grid points The projection coefficients of the training samples are the control coefficients of the tensor product interpolation function. Indicates the first j One B-spline basis function, Indicates the first k One B-spline basis function, Indicates the first l One B-spline basis function, Let the tensor product of three one-dimensional B-spline basis functions form a combined basis function in three-dimensional space.

[0064] It should be noted that the above summation operation only applies to locations including the target position. Grid points corresponding to the local support domain This is done by utilizing the local support properties of B-spline basis functions, thus ensuring that adjustments to local parameters do not affect the global prediction results.

[0065] For example, the first, second, and third core conditional parameters of a target task can be represented as target category parameters, spatial scale parameters, and viewpoint parameters in a visual reasoning task. In a specific physical signal generation task, they can be specifically represented as mass ratio parameters, first spin parameters, and second spin parameters.

[0066] In this embodiment of the invention, tensor product B-spline interpolation is employed. Its advantage lies in its ability to achieve smooth, locally controllable, and efficient prediction of target projection coefficients in a three-dimensional parameter space. Through tensor product operations, the multiplication of three one-dimensional basis functions simultaneously captures the cross-coupling effects of multiple core conditional parameters. The local support characteristics of the B-spline basis functions ensure that adjusting local data does not affect the global result, avoiding oscillations. Ultimately, this method guarantees that the prediction results change smoothly and continuously with the parameters, exhibiting high computational efficiency and suitability for real-time prediction scenarios.

[0067] Specifically, the target projection coefficient vector obtained in step S503 is saved as a low-dimensional projection representation corresponding to the target task, and used as an input parameter for the subsequent high-dimensional feature reconstruction process.

[0068] In this embodiment, the target projection coefficients are obtained through a tensor product B-spline interpolation model, enabling rapid response to unknown task parameter points without retraining a large-scale generalized model. Since the interpolation process considers the coupling relationships between multiple core condition parameters simultaneously, it effectively improves the generalization ability across different domains and real-time inference efficiency.

[0069] It should be noted that the tensor product B-spline interpolation in this embodiment does not directly predict the semantic features of the original input sample, but rather establishes a mapping relationship between task parameters and feature representations in the reduced-order feature space. Since the reduced-order feature space preserves the main semantic structures shared by different tasks, the reduced-order features corresponding to similar tasks exhibit continuous variation characteristics in the low-dimensional manifold space. The tensor product B-spline interpolation model utilizes the distribution law of the reduced-order features corresponding to the known task parameters to estimate the position of the target task in the low-dimensional feature space, thereby obtaining the target projection coefficient vector representation corresponding to the target task.

[0070] Furthermore, the features predicted in this embodiment are task-related low-dimensional feature representations, reflecting the feature evolution trend of the target task relative to known tasks, rather than being directly generated from the specific input sample content. The actual semantic information of the current input sample still originates from intermediate features extracted by a large-scale model across the general domain. Tensor product B-spline interpolation is only used to guide task adaptation and domain transfer of these intermediate features. The reduced-order feature space is constructed by performing singular value decomposition on the intermediate features of multiple known tasks, which can represent the dominant semantic structure shared between different tasks. Based on this shared feature space, the feature representations corresponding to the parameters of adjacent tasks satisfy the assumption of local continuity. Therefore, tensor product B-splines can be used to approximate the feature representation corresponding to the unknown task.

[0071] S6: Based on the reduced-order feature basis, feature reconstruction is performed on the target projection coefficients to obtain the reconstructed features.

[0072] In one possible implementation, S6 specifically includes sub-steps S601 and S602: S601: Based on the mapping relationship between the reduced-order feature basis and the target projection coefficients, the target projection coefficients are back-mapped to obtain the high-dimensional feature information corresponding to the target projection coefficients.

[0073] Specifically, based on the inverse mapping relationship established by singular value decomposition, the low-dimensional target projection coefficient vector obtained in step S503 is multiplied with the reduced-order feature basis by matrix multiplication, and then mapped back to the high-dimensional feature space to recover the corresponding high-dimensional feature information.

[0074] S602: Combine and reconstruct the various high-dimensional feature information to obtain the reconstructed features.

[0075] Specifically, when constructing a reduced-order model, the feature data may be decomposed into multiple decoupled reference feature components, or feature amplitude components and feature phase components, and then processed separately for reduced-order. Based on the original structural logic of the signal, these independently recovered high-dimensional feature information are processed, such as combining the real and imaginary parts, or fusing and splicing the feature amplitude and feature phase, and performing mathematical operations and combination splicing to generate structurally complete reconstructed features for use by subsequent network layers.

[0076] In this embodiment of the invention, the target projection coefficient vector is reverse-mapped and reconstructed using a reduced-order feature basis. Utilizing the reverse mapping relationship established by singular value decomposition, the low-dimensional target projection coefficient vector is efficiently and accurately restored to high-dimensional feature information, achieving an inverse transformation from compressed representation to complete features. Simultaneously, by combining and splicing multiple independent physical components according to their original structural logic, a structurally complete reconstructed feature is generated, ensuring the consistency and integrity of the reconstructed signal. This preserves the dimensionality compression advantage brought by the reduced-order processing while simultaneously enabling the recovery of high-dimensional features as needed, providing a data foundation for subsequent channel adjustment and task output.

[0077] S7: Map the reconstructed features to the original feature space of the general domain large model to obtain adaptive optimized features, and generate the output results of the target task based on the adaptive optimized features.

[0078] In one possible implementation, S7 specifically includes sub-steps S701 to S703: S701: Based on the feature adjustment parameters in the encoded data, perform feature channel-level affine adjustment on the reconstructed features.

[0079] Specifically, the adjustment formula is as follows:

[0080] in, Represents the characteristic-level linear modulation function. Indicates the first i The sample or task in the first Scaling parameters on each feature channel Indicates the first i The sample or task in the first Offset parameters on each feature channel Indicates the first i The sample or task in the first The reconstructed feature values ​​in each feature channel, that is, the reconstructed features generated in step S6 in each channel. The portion on top.

[0081] Furthermore, the scaling parameter is multiplied by the corresponding feature channel of the reconstructed feature using a Hadamard product (element-wise multiplication) to adjust its activation intensity. Then, the offset parameter is superimposed to adjust its baseline offset, thereby injecting the conditional information of the target task (such as the semantics of the input text or speech) into the reconstructed feature, achieving adaptive dynamic adjustment of the feature representation.

[0082] S702: Map the adjusted reconstructed features to the original feature space of the generalized large model to obtain adaptive optimized features.

[0083] Specifically, by utilizing the mapping relationship between the reduced-order feature base and the original feature space, the reconstructed features after feature adjustment are mapped back to the original feature space of the corresponding layer of the generalized domain model, so that the reconstructed features are restored to the feature dimension and representation form consistent with the original intermediate features.

[0084] Furthermore, during the mapping process, the structural constraints and semantic distribution characteristics of the original feature space are maintained, so that the adjusted reconstructed features maintain the same representation scale and feature structure as the original feature space of the general domain large model, thereby obtaining adaptive optimized features suitable for the target task and providing optimized feature inputs for subsequent inference computation.

[0085] S703: Input the adaptive optimization features into the task processing layer corresponding to the general domain large model, perform inference calculation on the adaptive optimization features, and obtain the output results.

[0086] Among them, the task processing layer refers to the network module in the general domain large model that is specifically used to receive adaptive optimization features, perform final inference calculations, and output the target task results.

[0087] Specifically, the output is the final task prediction or decision information obtained by inference calculation after the adaptive optimization features are input into the task processing layer of the general domain model.

[0088] Specifically, the adaptively optimized features, after conditional adjustment and spatial mapping, are input into a dedicated task processing layer downstream of the general domain model. Further nonlinear transformations and distribution mappings are performed on the network weights within the task processing layer (e.g., calculating the predicted probability distribution through a multilayer perceptron or convolutional layers), ultimately leading to the inference and output of the target task's output.

[0089] In this embodiment of the invention, the scaling and offset parameters of the FiLM generator are used to perform channel-by-channel affine adjustment of the reconstructed features, achieving precise injection of conditional information into the feature representation, enhancing task-related features and suppressing interference. Simultaneously, the inference output from the task processing layer achieves end-to-end computation from conditional encoding to final prediction, significantly improving the output accuracy and model practicality in multimodal tasks.

[0090] S8: Based on the input information, determine whether the target domain corresponding to the target task is included in the preset domain set. If so, use the encoded data corresponding to the target domain as the target encoding parameter. Otherwise, generate the target encoding parameter based on the encoded data corresponding to the known domain.

[0091] It should be noted that if the target domain corresponding to the target task is included in the preset domain set, it indicates that the target domain has direct supervision signals. In this case, the encoded data corresponding to the target domain (i.e., the scaling and offset parameters generated by the FiLM generator) is directly retrieved and extracted from the preset parameter library as the target encoded parameters. If the target domain corresponding to the target task is not in the preset domain set, it indicates that the target domain belongs to a completely new domain that the model has not seen (i.e., a zero-sample scenario). In this case, the target encoded parameters are generated based on the encoded data corresponding to the known domain using concept decomposition and linear combination.

[0092] Those skilled in the art can set preset domain sets and preset parameter libraries according to actual needs, and this invention does not limit them.

[0093] In one possible implementation, generating the target encoding parameters in step S8 based on the encoded data corresponding to the known domain specifically includes: From the known domains, identify a first domain with a first correlation to the target domain, a second domain with a second correlation, and a third domain with a benchmark correlation.

[0094] Specifically, from the set of known domains, the known domains that share the first type of attribute features with the target domain are identified as the first domain, the known domains that share the second type of attribute features with the target domain are identified as the second domain, and the baseline domain that lacks both the first and second type of attribute features is identified as the third domain. Here, the first type of attribute feature can be a specific color attribute, and the second type of attribute feature can be a specific shape attribute.

[0095] The encoded data corresponding to the first domain is superimposed with the encoded data corresponding to the second domain to obtain the superposition result.

[0096] Specifically, extract the encoded data corresponding to the first domain. The encoded data corresponding to the second domain And the coded data corresponding to the third domain ,in, , and These represent the scaling parameters corresponding to the first, second, and third domains, respectively. , and These represent the offset parameters corresponding to the first, second, and third domains, respectively. The encoded data from the first and second domains are then superimposed. and This is to aggregate the two attribute features required by the target.

[0097] Subtract the encoded data corresponding to the third domain from the superposition result to generate the target encoding parameters.

[0098] Furthermore, by subtracting the encoded data of the third domain (i.e., ...) from the superposition result and This process involves stripping away redundant, generic attribute features from the baseline domain. Through this analogical reasoning-based linear combination computation, the generated encoded data can accurately represent the feature combinations unique to the target domain, thereby generating the target encoding parameters corresponding to the target domain.

[0099] In this embodiment of the invention, the dependence of traditional large models on labeled data from new domains is overcome. By combining known capabilities with semantics, the model is endowed with zero-shot generalization ability to handle unknown combinations of concepts. Simultaneously, the characteristics of semantic concept decoupling and approximately linear relationships in the FiLM parameter space are fully utilized, transforming the complex domain adaptation problem into efficient parameter space algebraic operations. Parameters generated in this way can be used to back-optimize a lightweight FiLM generator at low cost, without modifying the massive underlying feature network. This enables dynamic adaptation and continuous self-evolution of broad-domain large models to new domains, effectively solving the zero-shot generalization problem, significantly expanding the model's application scope, and reducing adaptation costs.

[0100] S9: Optimize large-scale models across a broad domain based on target encoding parameters.

[0101] In one possible implementation, S9 specifically includes sub-steps S901 to S904: S901: Input the target task's input information into the FiLM generator to obtain the prediction coding parameters.

[0102] Specifically, the final target encoding parameters determined in step S8 are obtained. For parameters in the known domain, these are directly extracted parameters; for parameters in the unknown domain, they are linearly combined parameters, which are used as pseudo-labels for supervised learning. Simultaneously, the input information from the current target domain is re-inputted into the FiLM generator, and forward propagation yields the predicted encoding parameters currently output by the FiLM generator, i.e., the predicted scaling parameter vector. and offset parameter vector .

[0103] S902: Construct parameter reconstruction loss based on target coding parameters and predicted coding parameters.

[0104] Specifically, the distance between the predicted coding parameters output by the FiLM generator and the target coding parameters used as pseudo-labels is calculated to construct a parameter reconstruction loss function. The loss function can be the mean squared error loss (MSE), calculated as follows:

[0105] in, This represents the parameter reconstruction loss. and These represent the prediction scaling parameter and prediction offset parameter output by the FiLM generator, respectively. and These represent the target scaling parameter and the target offset parameter in the target encoding parameters, respectively. This represents the square of the L2 norm of a vector.

[0106] S903: Based on the parameter reconstruction loss, the network weights of the FiLM generator are updated by backpropagation to obtain the optimized FiLM generator.

[0107] Specifically, the reconstruction loss is based on the calculated parameters. The gradient descent algorithm is used to calculate the gradient of the loss function with respect to the weights of the FiLM generator network, and the parameters of the FiLM generator are updated by backpropagation. The optimization goal is to enable the FiLM generator to directly output modulation parameters consistent with the target coding parameters when it receives input information from the target domain, thereby internalizing the ability of externally calculated or combined parameters into the generator itself.

[0108] S904: Integrate the network weights of the optimized FiLM generator into the general domain large model to obtain the optimized general domain large model.

[0109] Specifically, the network weights of the FiLM generator, after weight updates, are loaded into a large-scale, cross-domain model to construct an end-to-end, independently operable inference system. This allows the model to dynamically generate precise modulation parameters when faced with target domain inputs, without relying on external parameter retrieval or algebraic combinations, thereby adaptively controlling the selective activation and suppression of the feature network.

[0110] In this embodiment of the invention, by using the target encoding parameters as pseudo-labels to supervise the training of the FiLM generator, the ability to transform from external parameter computation to internal parameter generation is realized. This process no longer involves repeated modulation of intermediate features, but instead performs supervised regression directly in the parameter space, avoiding logical conflicts between the inference and training processes. Simultaneously, this optimization process solidifies zero-shot generalization capability into the generator's intrinsic weights, endowing large models with the ability to handle combinatorial concepts not seen during training, and enabling flexible expansion of shared feature networks to new domains with extremely low parameter costs. Ultimately, it ensures the consistency of the model architecture and weights, allowing for direct and efficient deployment in multi-domain tasks without additional runtime assembly.

[0111] Reference manual attached Figure 2 This diagram illustrates the adaptive optimization convergence of the FiLM generator under the supervision of target encoding parameters provided in an embodiment of the present invention.

[0112] Specifically, in the adaptive optimization scenario of an unknown composite target domain, comparing the non-combinatorial target encoding and task-level fine-tuning only, the figures show that when using target encoding pseudo-label supervision, the MSE of the encoding parameter reconstruction decreases from approximately 0.142 initially to approximately 0.019 at the 40th iteration; while the non-combinatorial target encoding scheme decreases to approximately 0.086, and the task-level fine-tuning scheme remains at approximately 0.122. This result indicates that the target encoding parameters can provide a more explicit parameter space supervision signal for the FiLM generator because the optimization process directly constrains the prediction scaling parameters and offset parameters to approximate the unknown domain target encoding, rather than relying solely on indirect updates based on downstream output errors. This demonstrates that this mechanism can gradually internalize the unknown domain adaptation capability obtained from external combinations into generator weights, thereby enhancing the model's continuous adaptive capability in open dynamic domains.

[0113] Reference manual attached Figure 3 The diagram shows a schematic of the structure of a cross-domain fusion, general-domain large-model adaptive intelligent optimization system provided by the present invention.

[0114] This invention also provides a cross-domain fusion general-domain large-scale model adaptive intelligent optimization system 20, applied to the above-mentioned cross-domain fusion general-domain large-scale model adaptive intelligent optimization method, comprising: Processor 201.

[0115] The memory 202 stores computer-readable instructions. When the computer-readable instructions are executed by the processor 201, they implement the cross-domain fusion adaptive intelligent optimization method for a large-scale model with broad domain characteristics as described in the method embodiment.

[0116] The cross-domain fusion general-domain large model adaptive intelligent optimization system 20 provided by the present invention can execute the above-mentioned cross-domain fusion general-domain large model adaptive intelligent optimization method and achieve the same or similar technical effects. To avoid repetition, the present invention will not elaborate further.

[0117] It should be understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0118] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0119] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0120] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0121] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0122] It should be understood that, in various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0123] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0124] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0125] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0126] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0127] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0128] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0129] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0130] The following points need to be explained: (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0131] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0132] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0133] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cross-domain fusion adaptive intelligent optimization method for large-scale, pan-domain models, characterized in that, include: S1: Obtain input information for target tasks in different domains; S2: The input information is conditionally encoded using a FiLM generator to obtain encoded data; S3: Reduce the order of intermediate features in the general domain model to obtain reduced-order feature bases and multiple reduced-order features; S4: Based on the encoded data, determine the channel mapping relationship of the reduced-order feature in the parameter space to obtain the channel mapping feature; S5: Perform parameter space interpolation on the channel mapping features to obtain the target projection coefficients corresponding to the target task; S6: Based on the reduced-order feature basis, perform feature reconstruction on the target projection coefficients to obtain the reconstructed features; S7: Map the reconstructed features to the original feature space of the general domain large model to obtain adaptive optimized features, and generate the output result of the target task based on the adaptive optimized features; S8: Based on the input information, determine whether the target domain corresponding to the target task is included in the preset domain set; If so, the encoded data corresponding to the target domain shall be used as the target encoding parameter; Otherwise, the target encoding parameters are generated based on the encoding data corresponding to the known domain. S9: Optimize the general domain large model based on the target encoding parameters.

2. The cross-domain fusion adaptive intelligent optimization method for large-scale models with broad domain integration according to claim 1, characterized in that, The input information specifically includes: image feature vectors, text feature vectors, speech feature vectors, and task labels.

3. The cross-domain fusion adaptive intelligent optimization method for large-scale models with broad domain integration according to claim 1, characterized in that, S3 specifically includes: S301: Construct a training feature matrix based on the intermediate features of the general domain large model; S302: Perform singular value decomposition on the training feature matrix to obtain the singular value decomposition result, wherein the singular value decomposition result specifically includes the reduced-order feature basis, the singular value matrix, and the right singular vector; S303: The singular value decomposition result is reduced in order to obtain multiple reduced-order features.

4. The cross-domain fusion adaptive intelligent optimization method for large-scale models with broad domain integration according to claim 1, characterized in that, S4 specifically includes: S401: Determine the target active feature channel index based on the scaling parameter and offset parameter in the encoded data; S402: Based on the target activation feature channel index, perform channel filtering and recombination on the reduced-order feature to obtain the channel mapping feature.

5. The cross-domain fusion adaptive intelligent optimization method for large-scale models with broad domain integration according to claim 1, characterized in that, S5 specifically includes: S501: Determine the target position of the channel mapping feature in the preset parameter space based on the target task parameters of the channel mapping feature; S502: Generate projection coefficients of multiple training samples that are adjacent to the target location using a pre-built parameter space interpolation model; S503: By using the basis function weights corresponding to the projection coefficients of multiple training samples, the projection coefficients of each training sample are weighted and calculated to obtain the target projection coefficients corresponding to the target task parameters.

6. The cross-domain fusion adaptive intelligent optimization method for large-scale models with broad domain integration according to claim 1, characterized in that, S6 specifically includes: S601: Based on the mapping relationship between the reduced-order feature basis and the target projection coefficient, perform reverse mapping calculation on the target projection coefficient to obtain the high-dimensional feature information corresponding to the target projection coefficient; S602: Combine and reconstruct the various high-dimensional feature information to obtain the reconstructed features.

7. The cross-domain fusion adaptive intelligent optimization method for large-scale models with broad domain integration according to claim 1, characterized in that, Specifically, S7 includes: S701: Perform feature channel-level affine adjustment on the reconstructed features according to the feature adjustment parameters in the encoded data; S702: Map the adjusted reconstructed features to the original feature space of the generalized large model to obtain the adaptive optimized features; S703: Input the adaptive optimization features into the task processing layer corresponding to the general domain large model, perform inference calculation on the adaptive optimization features, and obtain the output result.

8. The cross-domain fusion adaptive intelligent optimization method for large-scale models with broad domain integration according to claim 1, characterized in that, In step S8, generating the target encoding parameters based on the known domain-specific encoding data specifically includes: From the known domains, determine a first domain that has a first association with the target domain, a second domain that has a second association with the target domain, and a third domain that has a benchmark association with the target domain; The encoded data corresponding to the first domain is superimposed with the encoded data corresponding to the second domain to obtain the superposition result; The target encoding parameters are generated by subtracting the encoded data corresponding to the third domain from the superposition result.

9. The cross-domain fusion adaptive intelligent optimization method for a large-scale, cross-domain model according to claim 1, characterized in that, S9 specifically includes: S901: Input the input information of the target task into the FiLM generator to obtain the prediction coding parameters; S902: Construct a parameter reconstruction loss based on the target coding parameters and the predicted coding parameters; S903: Based on the parameters, reconstruct the loss and perform backpropagation to update the network weights of the FiLM generator to obtain the optimized FiLM generator; S904: Integrate the network weights of the optimized FiLM generator into the general domain large model to obtain the optimized general domain large model.

10. A cross-domain integrated, pan-domain large-scale model adaptive intelligent optimization system, characterized in that, include: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement the cross-domain fusion adaptive intelligent optimization method for a broad-domain large model as described in any one of claims 1 to 9.