Intelligent parameter sharing method and system

Abstract

The invention discloses an intelligent parameter sharing method and system, and the method comprises the following steps: 1, inputting data, and carrying out the multi-modal feature extraction, wherein the multi-modal features comprise an image feature, a text description feature and a context feature; 2, mapping the extracted image features, text description features and context features into feature vectors through a cross-modal embedding model; 3, generating a parameter sharing decision in real time through a dynamic adjustment strategy based on the feature vector; 4, according to the parameter sharing decision, the shared parameter pools among the multiple model modules are coordinated through the unified management node, through the implementation of the method, the sharing strategy is automatically adjusted, it is ensured that the shared parameters of each module meet the current task requirement, meanwhile, manual intervention is reduced, and the efficiency is improved. Developers can save a large amount of development time and debugging cost, and certain use value and popularization value are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of parameter sharing, and particularly to an intelligent parameter sharing method and system. Background Art

[0002] With the rapid development of artificial intelligence technology, deep learning models have been widely applied in many fields such as computer vision and natural language processing. In order to further improve the model performance, it has become a trend to design highly complex network structures and a large number of parameters. However, the increase in model complexity and parameter scale will lead to a significant increase in computational costs and storage requirements, which to a certain extent limits the deployment and application of models in resource-constrained environments.

[0003] In this context, parameter sharing technology has emerged. Parameter sharing can not only effectively reduce the parameter scale of the model, but also transfer knowledge between different tasks, thereby enhancing the generalization ability of the model. For example, in a convolutional neural network (CNN), weight sharing is widely used in the design of convolutional kernels, significantly reducing the number of parameters that need to be trained. However, traditional parameter sharing mechanisms are usually static, that is, the scope and method of parameter sharing are fixed at the model design stage, making it difficult to adapt to diverse and dynamically changing task requirements.

[0004] To solve the above problems, this study proposes a mechanism for intelligent shared parameter nodes. This node can dynamically adjust the scope of parameter sharing based on the characteristics of the input data (such as text description, image style, etc.) or the context environment. By dynamically managing shared parameters, a balance between model complexity and computational efficiency is achieved, thereby enhancing the adaptability of the model in multi-task and multi-scenario situations.

[0005] Currently, existing research and applications mostly focus on fixed-mode parameter sharing, lacking flexibility and intelligence in dynamic scenarios. The proposed intelligent shared parameter node not only fills this technical gap, but also provides a new direction for model optimization and expansion, having important theoretical value and practical application significance.

[0006] In the current technical field, parameter sharing mechanisms generally adopt static configuration methods, that is, at the initial stage of model construction or training, it is pre-set which parameters should be shared and which parameters should be independent. Such fixed designs lack the ability to dynamically adjust according to the characteristics of the input data, resulting in limited flexibility of parameter sharing and difficulty in adapting to various different task requirements.

[0007] The management strategy of parameter sharing is relatively simple, mainly relying on manual rules or fixed-level structure sharing, and it is difficult to make intelligent adjustments based on the context features of the input data, task requirements, or optimization goals. This lack of context-aware sharing strategy may lead to problems such as parameter redundancy or insufficient sharing.

[0008] In summary, there is a need for an intelligent parameter sharing method and system to address the deficiencies in the prior art. Summary of the Invention

[0009] In view of the deficiencies of the prior art, the present invention provides an intelligent parameter sharing method and system, aiming to solve the above problems.

[0010] To achieve the above object, the present invention provides the following technical solution: An intelligent parameter sharing method, comprising the following steps: Step 1: Input data and perform multi-modal feature extraction. The multi-modal features include image features, text description features, and context features; Step 2: Map the extracted image features, text description features, and context features into feature vectors through a cross-modal embedding model; Step 3: Based on the feature vectors, generate a parameter sharing decision in real time through a dynamic adjustment strategy; Step 4: According to the parameter sharing decision, coordinate the shared parameter pool among multiple model modules through a unified management node, and realize the dynamic loading and collaborative optimization of cross-module parameters based on the parameter dependency graph; Step 5: During model training or inference, dynamically adjust the parameter sharing scope and computing resource allocation according to real-time task requirements and resource status to balance model performance and computing efficiency, and achieve global resource optimization. Through intelligent sharing management, the system can automatically adjust the sharing strategy to ensure that the shared parameters of each module meet the current task requirements. At the same time, it reduces manual intervention, and developers can save a large amount of development time and debugging costs, especially in multi-task processing and large-scale model training, with remarkable effects.

[0011] Further, in the above Step 1, inputting data and performing multi-modal feature extraction specifically includes the following steps: Step 1.1: Extract the color distribution, texture information, and edge features of the image through a convolutional neural network; Step 1.2: Semantically encode the text description through a cross-modal embedding model to generate a text feature vector aligned with the image features; Step 1.3: Combine the context environment features, including task type, device resource status, and user requirements, to generate a comprehensive feature vector to guide the parameter sharing decision.

[0012] Further, in the above Step 3, generating a parameter sharing decision in real time through a dynamic adjustment strategy, the implementation of the dynamic adjustment strategy includes a reinforcement learning strategy or a rule system strategy based on cross-modal feature alignment.

[0013] Further, the parameter sharing decision for generating the reinforcement learning policy specifically includes the following steps: Step 3.1.1: Construct a policy network with the input being the feature vector and the output being the shared weights or independent loading identifiers of each model parameter; Step 3.1.2: Train the policy network through a deep reinforcement learning algorithm, where the reward function is based on the quality metric and resource consumption metric of the generated result; Step 3.1.3: In the inference stage, dynamically optimize the shared policy according to the input features in real time to minimize the computational cost and maximize the task performance; The parameter sharing decision for generating the rule system policy specifically includes the following steps: Step 3.2.1: Preset a shared rule library, which contains parameter sharing conditions based on color distribution, semantic sentiment, and task type; Step 3.2.2: Match the feature vector with the rule library through a rule engine to trigger the corresponding parameter sharing policy.

[0014] Further, in step 4, a unified management node coordinates the shared parameter pool among multiple model modules. The unified management node organizes the shared parameter pool using a circular matrix structure and realizes cross-module collaboration through the following steps: Step 4.1: Construct a parameter dependency graph between modules, mark the call paths of shared parameters, and ensure the consistent call of shared parameters in multiple tasks; Step 4.2: Coordinate the parallel computing requests of multiple modules through memory sharing technology and an asynchronous task scheduling system to reduce duplicate calculations and improve resource utilization efficiency; Step 4.3: Through the dynamic allocation policy of the parameter pool, adjust the parameter loading order according to the module priority and task urgency.

[0015] Further, in step 4.1, construct a parameter dependency graph between modules: , where, represents parameter depending on parameter ; In step 4.2, the goal of the memory sharing technology is to minimize the amount of duplicate calculations: , where P is the parameter loading position allocation vector.

[0016] Further, in step 5, dynamically adjust the parameter sharing range and computing resource allocation according to the real-time task requirements and resource status. The dynamic adjustment specifically includes the following steps: Step 5.1: In a resource-constrained scenario, load task-specific layer parameters independently and share the parameters of the basic network layer; Step 5.2: According to the feedback of real-time inference speed and memory occupancy, adaptively adjust the sharing ratio through a closed-loop control algorithm to optimize the end-to-end latency; Step 5.3: When multiple tasks are processed in parallel, allocate shared computing resources through a task scheduling system to avoid redundant calculations.

[0017] Furthermore, in the said Step 5.2, the sharing ratio is adaptively adjusted through a closed-loop control algorithm, and the specific formula is: , where, is the sharing ratio, is the target latency, is the current latency, is the learning rate.

[0018] An intelligent parameter sharing system, comprising: A feature extraction module, configured to extract multimodal features through a convolutional neural network and a cross-modal embedding model; A dynamic adjustment strategy module, configured to generate parameter sharing decisions according to multimodal features, and support the switching between reinforcement learning strategies and rule system strategies; A parameter pool and a unified management node, configured to store shared parameters and coordinate cross-module collaboration; A task scheduling module, configured to dynamically allocate computing resources according to the resource status, and optimize the training and inference processes through asynchronous scheduling; Wherein, the output of the feature extraction module is connected to the dynamic adjustment strategy module, the output of the dynamic adjustment strategy module is connected to the unified management node, and the unified management node communicates bidirectionally with the task scheduling module.

[0019] Furthermore, the parameter pool and the unified management node adopt a distributed memory sharing technology to coordinate parallel requests of multiple model modules through an asynchronous task scheduling system; The task scheduling module has a built-in priority queue, dynamically allocates access rights to the shared parameter pool according to the task urgency and module computing load, and adjusts the parameter loading order based on the resource occupancy feedback.

[0020] The substantial effects of the present invention: 1. In the present invention, feature evaluation and sharing strategy optimization, in diverse task scenarios, by evaluating input features and optimizing the sharing strategy, significantly enhance the adaptability of the model, improve its generalization ability, and can automatically adjust in complex tasks, thus avoiding manual intervention.

[0021] 2. In the present invention, through intelligent sharing management, the system can automatically adjust the sharing strategy to ensure that the sharing parameters of each module meet the current task requirements. At the same time, manual intervention is reduced, and developers can save a large amount of development time and debugging costs, especially in multi-task processing and large-scale model training, which has remarkable effects.

[0022] 3. In the present invention, in a complex deep learning system, multiple modules share calculation results, optimize parameters, and save resources through collaborative work. This cross-module collaboration mechanism significantly improves the training efficiency and reduces the calculation cost, especially in large-scale systems, the effect is particularly obvious. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a comparative schematic diagram of the experimental data verification effects under different tasks in Example 4.

[0024] Figure 2 It is a comparative schematic diagram of the number of iterations and GPU video memory occupancy in Example 4.

[0025] Figure 3 It is a comparative schematic diagram of the training duration varying with the number of iterations in Example 4.

[0026] Figure 4 It is a comparative schematic diagram of the improvement of the accuracy index in Example 4. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] For the convenience of understanding the present invention, the present invention will be described in more detail below in conjunction with the drawings and specific embodiments. It should be noted that when an element is expressed as "fixed to" another element, it can be directly on the other element, or there can be one or more intermediate elements therebetween. When an element is expressed as "connected to" another element, it can be directly connected to the other element, or there can be one or more intermediate elements therebetween. The terms "vertical", "horizontal", "left", "right" and similar expressions used in this specification are only for the purpose of illustration.

[0028] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not used to limit the present invention. The term "and / or" used in this specification includes any and all combinations of one or more of the related listed items. Example 1:

[0029] An intelligent parameter sharing method includes the following steps: Step 1: Input data and perform multi-modal feature extraction. The multi-modal features include image features, text description features, and context features. Image features are used to extract color distribution, texture information, and edge features through convolutional neural networks (such as ResNet, Vision Transformer). Text description features are semantically encoded through cross-modal embedding models (such as BERT, CLIP text encoder). Context features include task context: encoding of the type / target / constraint conditions of the current task, environmental context: hardware resource status / user device parameters, temporal context: feedback loop features of historical generation results, and cross-modal association: joint embedding representation after alignment of image and text features. Step 2: Map the extracted image features, text description features, and context features into feature vectors through a cross-modal embedding model. Step 3: Based on the feature vectors, generate parameter-sharing decisions in real time through a dynamic adjustment strategy. Step 4: According to the parameter-sharing decisions, coordinate the shared parameter pool among multiple model modules through a unified management node, and realize the dynamic loading and collaborative optimization of cross-module parameters based on the parameter dependency graph. Step 5: During model training or inference, dynamically adjust the parameter-sharing scope and computing resource allocation according to real-time task requirements and resource status to balance model performance and computing efficiency and achieve global resource optimization.

[0030] As an implementation, in Step 1, inputting data and performing multi-modal feature extraction specifically includes the following steps: Step 1.1: Extract the color distribution, texture information, and edge features of the image through a convolutional neural network. Step 1.2: Semantically encode the text description through a cross-modal embedding model to generate a text feature vector aligned with the image features. Step 1.3: Combine context features, including task type, device resource status, and user requirements, to generate a comprehensive feature vector to guide parameter-sharing decisions.

[0031] As an implementation, in Step 3, generate parameter-sharing decisions in real time through a dynamic adjustment strategy. The implementation of the dynamic adjustment strategy includes a reinforcement learning strategy or a rule system strategy based on cross-modal feature alignment, and specifically includes the following steps: Step 3.0: Calculate the similarity score between the image feature vector and the text description feature vector through a cross-modal alignment module to generate a cross-modal alignment weight matrix. Step 3.1: If a reinforcement learning strategy is adopted, add a cross-modal alignment index to the reward function, specifically: , where For the quality metric of the generated result (such as PSNR or accuracy), For the resource consumption metric (such as computation time or memory usage), For the cross-modal alignment score, calculated by the semantic consistency between the generated result and the input text after sharing parameters, For the weight coefficient; Step 3.2: If the rule-based system strategy is adopted, preset the cross-modal alignment conditions in the rule base, including: When the semantic similarity between the image and the text exceeds the threshold (>85%), force the sharing of the semantic encoding layer parameters, otherwise load independently. Cross-modal alignment drive: The parameter sharing decision not only depends on the task requirements, but also on the collaborative alignment of the image and text features to ensure that the shared parameters can simultaneously adapt to the semantic consistency of the multi-modal input; Dynamic adjustment mechanism: In reinforcement learning, the cross-modal alignment score directly affects the optimization direction of the policy network; in the rule-based system, the alignment condition becomes a hard constraint.

[0032] As an implementation, the reinforcement learning strategy (if the task scenario requires dynamic learning ability (such as complex and changing input data), then reinforcement learning is preferred) to generate the parameter sharing decision specifically includes the following steps: Step 3.1.1: Construct a policy network, with the input being the feature vector and the output being the sharing weight or independent loading flag of each model parameter; Policy network: Input feature vector , Output sharing weight , where is the policy network, represents the sharing probability of each parameter; Step 3.1.2: Train the policy network through a deep reinforcement learning algorithm (Proximal Policy Optimization), and the reward function is based on the quality metric and resource consumption metric of the generated result; Reward function: , , where is the quality metric of the generated result (such as PSNR or accuracy), is the resource consumption metric (such as computation time or memory usage), is the weight coefficient, Step 3.1.3: In the inference stage, dynamically optimize the sharing policy according to the input features in real time to minimize the computational cost and maximize the task performance; Policy optimization: Update the network parameters through the policy gradient method θ: , , where Denotes the maximized expected return value, Denotes the partial derivative of the parameter θ, Denotes the expected value under the policy where is the action, Denotes the immediate reward obtained by executing the action and Denotes the partial derivative of the parameter θ, where is the probability of selecting the action under the given feature .

[0033] The rule system policy (if the task rules are clear and deterministic control is required (such as fixed-style classification), then select the rule engine) to generate parameter sharing decisions specifically includes the following steps: Step 3.2.1: Preset a shared rule library, and the rule library contains parameter sharing conditions based on color distribution, semantic sentiment, and task type; Define the rule library , where is the rule condition (such as the color distribution threshold ), is the corresponding shared policy; Step 3.2.2: Match the feature vector with the rule library through the rule engine to trigger the corresponding parameter sharing policy; Matching process: , where is the indicator function, and when the condition is satisfied, its value is 1; otherwise it is 0; First, traverse all policies : For each policy , calculate the number of features that satisfy the condition , and then calculate the number of features that satisfy the condition: For each feature , check whether it satisfies the condition , and if it is satisfied, increment the counter by 1; finally, select the optimal policy : Find the policy with the largest number of features that satisfy the condition .

[0034] As an implementation method, in step 4, the unified management node coordinates the shared parameter pool among multiple model modules. The unified management node organizes the shared parameter pool using a circular matrix structure (such as mapping parameter addresses through a hash table) and realizes cross-module collaboration through the following steps: Step 4.1: Construct a parameter dependency graph between modules, mark the call paths of shared parameters, and ensure the consistent call of shared parameters in multiple tasks; Construct a parameter dependency graph between building blocks: , Among them, the adjacency matrix is a matrix of size , and the element values are 0 or 1, indicating that parameter depends on parameter ; indicating that parameter does not depend on parameter ; Step 4.2: Coordinate the parallel computing requests of multiple modules through memory sharing technology and an asynchronous task scheduling system to reduce redundant calculations and improve resource utilization efficiency; The goal of memory sharing technology is to minimize the amount of redundant calculations: , Among them, is the vector for allocating parameter loading positions; Step 4.3: Adjust the parameter loading order according to the module priority and task urgency through the dynamic allocation strategy of the parameter pool; According to the module priority and the task urgency , allocate the parameter loading order: The priority score is equal to , is the element-wise multiplication. By marking the call paths of shared parameters, ensure that when sharing parameters between different modules, the dependency relationship is clear and consistent; by analyzing parameter dependencies, reduce redundant calculations (for example, multiple tasks depending on the same parameter can share the calculation results).

[0035] As an implementation, in Step 5, dynamically adjust the parameter sharing range and computing resource allocation according to real-time task requirements and resource status. The dynamic adjustment specifically includes the following steps: Step 5.1: In a resource-constrained scenario, independently load the parameters of the task-specific layer and share the parameters of the basic network layer; Step 5.2: According to the real-time inference speed and memory occupancy feedback, adaptively adjust the sharing ratio through a closed-loop control algorithm to optimize the end-to-end latency; Adaptively adjust the sharing ratio through a closed-loop control algorithm: , Among them, is the sharing ratio, is the target latency, is the current latency, is the learning rate; if > It indicates that the current latency is too high. The algorithm increases the sharing ratio to share more parameters, reduce the computational amount, and thus reduce the latency. It is the benchmark value for system performance optimization, and dynamically balances the computational efficiency (latency) and task quality (such as the accuracy of the generated result) through closed-loop feedback.

[0036] Step 5.3: When performing multi-task parallel processing, allocate shared computing resources through the task scheduling system to avoid redundant calculations; Task scheduling optimization goal: , where, is the amount of resources allocated to task , is the weight, is the total resource budget. Example 2:

[0037] This example is basically the same as Example 1. The difference is that in this example, the multi-modal feature extraction and cross-modal alignment steps adopt extracting local features (such as sole texture, breathable mesh structure) of the commodity picture through EfficientNet to generate image feature vectors ; semantically encoding the text input by the user through the BERT model to generate text feature vectors ; calculating the semantic similarity score between the image and the text: , and generating an alignment weight matrix , marking the association strength between the image region and the text keywords (for example, "breathable" corresponds to the mesh area of the shoe upper); Dynamic parameter sharing strategy: Policy network input: fused feature vector ; Shared decision output: output shared weight , controlling the parameter sharing ratio of 100 candidate commodity recommendation models; Reward function: , where the cross-modal alignment score directly correlates with the consistency between the recommended commodity and the user's intention (such as whether the recommendation result matches both the text "breathable" and the mesh design in the picture); Policy optimization: If > 0.9; Rule system strategy: Preset rule library: Rule 1: If > 0.85, Rule 2: If the proportion of "sports shoes" in the user's historical click records > 70%, load the user preference model parameters independently; Rule trigger: When the user uploads a picture and the text contains "summer" and "breathable", trigger Rule 1 and reuse the underlying feature extraction parameters of the image classifier; Parameter sharing and resource optimization: Parameter dependency graph construction, marking the parameter dependency paths of the product recommendation model (such as "image classifier → attribute encoder → recommendation sorter") to ensure consistent invocation of shared parameters; Dynamic resource allocation, during peak periods (such as "Double 11"), adjust the sharing ratio through a closed-loop control algorithm: , Stabilize the end-to-end response latency below 200ms; The asynchronous task scheduling system preferentially processes requests with a high alignment score ( > 0.8) and allocates more GPU resources.

[0038] The substantial technical effects of this embodiment: Improved recommendation accuracy: The cross-modal alignment mechanism increases the matching degree (F1 Score) between the recommended products and the user's intention from 78% to 89%; The user click-through rate increases by 15%, especially in the mixed input (text + picture) scenario; Optimized resource efficiency: The parameter sharing strategy reduces 30% of the repeated calculations, and the peak value of the server CPU load drops by 25%; The response latency drops from an average of 350ms to 180ms, supporting tens of thousands of concurrent requests per second; Enhanced scenario adaptability: The reinforcement learning strategy dynamically adjusts the sharing ratio during the promotion season to cope with traffic fluctuations (such as the peak period of "Double 11") and ensures that the response latency is stabilized below 200ms; The rule system provides deterministic guarantees in fixed categories (such as "sports shoes") to avoid unexpected performance degradation. Embodiment 3:

[0039] An intelligent parameter sharing system, including: A feature extraction module for extracting multi-modal features through a convolutional neural network and a cross-modal embedding model; A dynamic adjustment strategy module for generating parameter sharing decisions based on multi-modal features, supporting the switching between reinforcement learning strategies and rule system strategies; A parameter pool and a unified management node for storing shared parameters and coordinating cross-module collaboration; A task scheduling module for dynamically allocating computing resources according to the resource status and optimizing the training and inference processes through asynchronous scheduling; Among them, the output of the feature extraction module is connected to the dynamic adjustment strategy module, the output of the dynamic adjustment strategy module is connected to the unified management node, and the unified management node communicates bidirectionally with the task scheduling module.

[0040] As an implementation, the parameter pool and the unified management node adopt the distributed memory sharing technology, and coordinate the parallel requests of multiple model modules through the asynchronous task scheduling system; The task scheduling module has a built-in priority queue, dynamically allocates access rights to the shared parameter pool according to the urgency of tasks and the computing load of modules, and adjusts the parameter loading order based on the feedback of the resource occupancy rate. Example 4:

[0041] This example is basically the same as Example 1, except that in this example, experimental verification has been carried out in multiple typical deep learning tasks (such as image classification, text generation, etc.), and remarkable results have been obtained. The specific experimental data is as follows: Experimental settings: Datasets: A variety of standard datasets were used for verification, including ImageNet, CIFAR-10, COCO (for image classification) and OpenWebText (for text generation).

[0042] Evaluation metrics: Multiple metrics such as accuracy, F1 score, training time, inference speed, memory usage, etc. were used for evaluation.

[0043] Comparison baseline: Compared with traditional static parameter sharing models (such as traditional VGG, ResNet, LSTM, etc.), the advantages of this technical solution in resource optimization and task adaptability were evaluated.

[0044] Experimental results: Referring to the accompanying drawings of the specification Figures 1 - 4 It can be seen that: Reduction in the number of parameters: In the image classification task, under the optimization of the sharing strategy, the number of model parameters of this technical solution has been reduced by more than 20%. This is mainly due to the dynamic adjustment of the sharing range, which avoids the loading of redundant model parameters; Performance improvement: In multiple tasks, the performance (such as accuracy) of this solution has been improved by 5%-10% on average. Especially in the image classification task, by sharing the parameters of the convolutional layer, the model significantly reduces the computing load without sacrificing accuracy; Shorter training time: During the training process, due to the automatic management of the intelligent sharing node, the training time has been reduced by 15%-20%. Sharing computing resources and optimizing memory usage have significantly improved the overall training efficiency; Improved inference speed: Due to the reduction in the number of model parameters and the optimization of cross-module collaboration, the inference speed has been increased by 10%-15%, ensuring the efficient execution of real-time tasks.

[0045] Compared with traditional technologies, this solution demonstrates lower computational resource consumption, higher model adaptability, and significant performance improvement in multiple tasks. In traditional methods, parameter sharing is usually statically configured and cannot be flexibly adjusted according to task requirements. However, the context awareness and dynamic adjustment strategy of this solution enable the model to select the optimal parameter sharing strategy in real time.

[0046] Facing future challenges, this technical solution has good scalability and can adapt to larger-scale datasets and more complex task requirements. For example, for future multi-modal data (such as mixed input of text, images, and audio) and more complex generation tasks (such as generating videos, 3D scenes, etc.), this solution can further optimize the sharing strategy to ensure the best balance between resource utilization and computational efficiency of the system.

[0047] It should be noted that the description and drawings of the present invention provide preferred embodiments of the present invention. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. These embodiments are not additional limitations to the content of the present invention. The purpose of providing these embodiments is to make the understanding of the disclosed content of the present invention more thorough and comprehensive. Moreover, the above technical features continue to be combined with each other to form various embodiments not listed above, which are all regarded as the scope described in the specification of the present invention; further, for those of ordinary skill in the art, improvements or transformations can be made according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. An intelligent parameter sharing method, characterized in that: The following steps are involved: Step 1: Input data and extract multimodal features, which include image features, text description features, and context features; Step 2: Map the extracted image features, text description features, and context features into feature vectors through a cross-modal embedding model; Step 3: Based on the feature vector, a parameter sharing decision is generated in real time by dynamically adjusting the strategy; Step 4: Based on the parameter sharing decision, coordinate the shared parameter pools among multiple model modules through a unified management node, and implement dynamic loading and collaborative optimization of cross-module parameters based on the parameter dependency graph; Step 5: During model training or inference, dynamically adjust the parameter sharing range and computing resource allocation according to real-time task requirements and resource status to balance model performance and computing efficiency and achieve global resource optimization.

2. According to claim 1, a smart parameter sharing method is characterized in that: In step 1, inputting data and performing multimodal feature extraction specifically includes the following steps: Step 1.1: Extract the color distribution, texture information and edge features of the image through convolutional neural network; Step 1.2: Semantically encode the text description through a cross-modal embedding model to generate a text feature vector aligned with the image features; Step 1.3: Combine the contextual environment features, including task type, device resource status, and user needs, to generate a comprehensive feature vector to guide parameter sharing decisions.

3. According to claim 1, the intelligent parameter sharing method is characterized in that: In step 3, a parameter sharing decision is generated in real time through a dynamic adjustment strategy, and the implementation of the dynamic adjustment strategy includes a reinforcement learning strategy or a rule system strategy based on cross-modal feature alignment.

4. According to claim 3, a smart parameter sharing method is characterized in that: The reinforcement learning strategy generates parameter sharing decisions, which specifically include the following steps: Step 3.1.1: constructing a strategy network, with the input being the feature vector and the output being the shared weights or independent loading identifiers of the model parameters; Step 3.1.2: Train the policy network using a deep reinforcement learning algorithm, where the reward function is based on the quality index of the generated results and the resource consumption index; Step 3.1.3: In the inference phase, dynamically optimize the sharing strategy based on the input features in real time to minimize the computational cost and maximize the task performance; The rule system strategy generates parameter sharing decisions, which specifically include the following steps: Step 3.2.1: Preset a shared rule base, wherein the rule base includes parameter sharing conditions based on color distribution, semantic emotion, and task type; Step 3.2.2: Match the feature vector with the rule base through the rule engine to trigger the corresponding parameter sharing strategy.

5. The intelligent parameter sharing method according to claim 1, characterized in that: In step 4, a shared parameter pool among multiple model modules is coordinated through a unified management node, and the unified management node organizes the shared parameter pool in a ring matrix structure, and implements cross-module collaboration through the following steps: Step 4.1: Build a parameter dependency graph between modules, mark the calling paths of shared parameters, and ensure the consistent calling of shared parameters in multiple tasks; Step 4.2: Coordinate the parallel computing requests of multiple modules through memory sharing technology and asynchronous task scheduling system to reduce repeated computing and improve resource utilization efficiency; Step 4.3: Adjust the parameter loading order according to the module priority and task urgency through the dynamic allocation strategy of the parameter pool.

6. The intelligent parameter sharing method according to claim 5, characterized in that: In step 4.1, the parameter dependency diagram between the modules is constructed: ， in, Representation parameters Dependent parameters ; In step 4.2, the goal of the memory sharing technique is to minimize the amount of repeated calculations: ， Among them, P is the parameter loading position allocation vector.

7. The intelligent parameter sharing method according to claim 1, characterized in that: In step 5, the parameter sharing range and computing resource allocation are dynamically adjusted according to the real-time task requirements and resource status, and the dynamic adjustment specifically includes the following steps: Step 5.1: In resource-constrained scenarios, load task-specific layer parameters independently and share the parameters of the base network layer; Step 5.2: Based on the real-time inference speed and memory usage feedback, the sharing ratio is adaptively adjusted through a closed-loop control algorithm to optimize the end-to-end latency. Step 5.3: When multiple tasks are processed in parallel, the task scheduling system is used to allocate shared computing resources to avoid redundant computing.

8. The intelligent parameter sharing method according to claim 7, characterized in that: In step 5.2, the sharing ratio is adaptively adjusted through a closed-loop control algorithm, and the specific formula is: ， in, is the sharing ratio, For the target delay, is the current delay, is the learning rate.

9. An intelligent parameter sharing system, characterized in that: include: Feature extraction module, which is used to extract multimodal features through convolutional neural networks and cross-modal embedding models; Dynamic adjustment strategy module, used to generate parameter sharing decisions based on multimodal features, and supports switching between reinforcement learning strategies and rule system strategies; Parameter pool and unified management node, used to store shared parameters and coordinate cross-module collaboration; The task scheduling module is used to dynamically allocate computing resources according to resource status and optimize the training and reasoning processes through asynchronous scheduling; The output of the feature extraction module is connected to the dynamic adjustment strategy module, the output of the dynamic adjustment strategy module is connected to the unified management node, and the unified management node communicates with the task scheduling module in a two-way manner.

10. The intelligent parameter sharing system according to claim 9, characterized in that: The parameter pool and the unified management node adopt distributed memory sharing technology and coordinate parallel requests of multiple model modules through an asynchronous task scheduling system; The task scheduling module has a built-in priority queue, which dynamically allocates access rights to the shared parameter pool according to the task urgency and module computing load, and adjusts the parameter loading order based on resource occupancy feedback.

Images