Intelligent parameter sharing method and system

Abstract

The present invention discloses an intelligent parameter sharing method and system, which includes the following steps: Step 1: input data and perform multimodal feature extraction, where the multimodal 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 parameter sharing decisions in real time through a dynamic adjustment strategy; Step 4: according to the parameter sharing decision, coordinate the shared parameter pools among multiple model modules through a unified management node. Through the implementation of the present invention, automatic adjustment of the sharing strategy is realized to ensure that the shared parameters of each module meet the current task requirements. At the same time, manual intervention is reduced, and developers can save a lot of development time and debugging costs. The system has certain use value and promotion value.

Description

Technical Field

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

[0002] With the rapid development of artificial intelligence (AI), deep learning models have been widely used in fields such as computer vision and natural language processing. To further improve model performance, designing highly complex network structures and large numbers of parameters has become a trend. However, the increase in model complexity and parameter size leads to significant increases in computational costs and storage requirements, which, to a certain extent, limits the deployment and application of models in resource-constrained environments.

[0003] Against this backdrop, parameter sharing technology has emerged. Parameter sharing not only effectively reduces the model's parameter size but also transfers knowledge between different tasks, thereby enhancing the model's generalization capabilities. For example, in convolutional neural networks (CNNs), weight sharing is widely used in the design of convolution kernels, significantly reducing the number of parameters required for training. However, traditional parameter sharing mechanisms are typically static, meaning the scope and method of parameter sharing are fixed during the model design phase. This makes them difficult to adapt to diverse and dynamically changing task requirements.

[0004] To address these issues, this study proposes a mechanism for intelligent parameter sharing nodes. This node dynamically adjusts the scope of parameter sharing based on input data characteristics (such as text descriptions and image style) or context. By dynamically managing shared parameters, we achieve a balance between model complexity and computational efficiency, thereby enhancing the model's adaptability across multiple tasks and scenarios.

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

[0006] In current technology, parameter sharing mechanisms generally employ static configuration methods. This involves presetting which parameters should be shared and which should be independent during the initial stages of model building or training. This fixed design lacks the ability to dynamically adjust based on input data characteristics, limiting the flexibility of parameter sharing and making it difficult to adapt to diverse task requirements.

[0007] Parameter sharing management strategies are relatively simple, relying primarily on manual rules or fixed-level structured sharing. These strategies struggle to intelligently adjust parameters based on the contextual characteristics of input data, task requirements, or optimization objectives. This lack of contextual awareness can lead to redundant or insufficient parameter sharing.

[0008] In summary, an intelligent parameter sharing method and system are needed to address the deficiencies in the existing technology. Summary of the Invention

[0009] In view of the deficiencies in 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:

[0011] Step 1: Input data and perform multimodal feature extraction. Multimodal features include image features, text description features, and context features.

[0012] Step 2: Map the extracted image features, text description features, and context features into feature vectors through a cross-modal embedding model;

[0013] Step 3: Based on the feature vector, generate parameter sharing decisions in real time by dynamically adjusting the strategy;

[0014] Step 4: Based on the parameter sharing decision, the shared parameter pool among multiple model modules is coordinated through a unified management node, and dynamic loading and collaborative optimization of cross-module parameters are achieved based on the parameter dependency graph;

[0015] Step 5: During model training or inference, parameter sharing scope and computing resource allocation are dynamically adjusted based on real-time task requirements and resource status to balance model performance and computational efficiency, achieving global resource optimization. Through intelligent sharing management, the system automatically adjusts sharing strategies to ensure that each module's shared parameters meet the current task requirements. This reduces manual intervention, saving developers significant development time and debugging costs. This is particularly effective in multi-tasking and large-scale model training.

[0016] Furthermore, in step 1, inputting data and performing multimodal feature extraction specifically includes the following steps:

[0017] Step 1.1: Extract the color distribution, texture information and edge features of the image through convolutional neural network;

[0018] 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;

[0019] Step 1.3: Combine contextual features, including task type, device resource status, and user needs, to generate a comprehensive feature vector to guide parameter sharing decisions.

[0020] Furthermore, 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.

[0021] Furthermore, the reinforcement learning strategy generates a parameter sharing decision, which specifically includes the following steps:

[0022] Step 3.1.1: Construct a policy network with the feature vector as input and the shared weights or independent loading flags of each model parameter as output;

[0023] Step 3.1.2: Train the policy network using a deep reinforcement learning algorithm, with a reward function based on the quality of the generated results and the resource consumption indicator;

[0024] Step 3.1.3: During the inference phase, dynamically optimize the sharing strategy based on the input features in real time to minimize computational cost and maximize task performance.

[0025] The rule system strategy generates parameter sharing decisions, which specifically include the following steps:

[0026] 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;

[0027] Step 3.2.2: Match the feature vector with the rule base through the rule engine to trigger the corresponding parameter sharing strategy.

[0028] Furthermore, in step 4, a shared parameter pool among multiple model modules is coordinated by a unified management node. The unified management node organizes the shared parameter pool in a ring matrix structure and implements cross-module collaboration through the following steps:

[0029] Step 4.1: Build a parameter dependency graph between modules, mark the call paths of shared parameters, and ensure that shared parameters are called consistently across multiple tasks;

[0030] Step 4.2: Use memory sharing technology and an asynchronous task scheduling system to coordinate parallel computing requests from multiple modules, reducing duplicate computing and improving resource utilization efficiency.

[0031] Step 4.3: Adjust the parameter loading order according to module priority and task urgency through the dynamic allocation strategy of the parameter pool.

[0032] Furthermore, in step 4.1, a parameter dependency diagram between modules is constructed:

[0033] ,

[0034] in, Representation parameters Dependent parameters ;

[0035] In step 4.2, the goal of the memory sharing technique is to minimize the amount of repeated computation:

[0036] ,

[0037] Among them, P is the parameter loading position allocation vector.

[0038] Furthermore, 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:

[0039] Step 5.1: In resource-constrained scenarios, load the task-specific layer parameters independently and share the parameters of the base network layer;

[0040] Step 5.2: Based on the real-time inference speed and memory usage feedback, the closed-loop control algorithm adaptively adjusts the sharing ratio to optimize the end-to-end latency.

[0041] 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.

[0042] Furthermore, in step 5.2, the sharing ratio is adaptively adjusted through a closed-loop control algorithm, and the specific formula is:

[0043] ,

[0044] in, is the sharing ratio, For target delay, is the current delay, is the learning rate.

[0045] An intelligent parameter sharing system, comprising:

[0046] Feature extraction module, which is used to extract multimodal features through convolutional neural networks and cross-modal embedding models;

[0047] Dynamic adjustment strategy module, used to generate parameter sharing decisions based on multimodal features, supporting switching between reinforcement learning strategies and rule system strategies;

[0048] Parameter pool and unified management node, used to store shared parameters and coordinate cross-module collaboration;

[0049] The task scheduling module is used to dynamically allocate computing resources based on resource status and optimize the training and inference processes through asynchronous scheduling;

[0050] 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.

[0051] Furthermore, the parameter pool and the unified management node adopt distributed memory sharing technology to coordinate the parallel requests of multiple model modules through an asynchronous task scheduling system;

[0052] 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.

[0053] Substantial effects of the present invention:

[0054] 1. In this invention, feature evaluation and sharing strategy optimization significantly enhance the adaptability of the model and improve its generalization ability in diverse task scenarios by evaluating input features and optimizing sharing strategies. It can also automatically adjust to complex tasks, thus avoiding manual intervention.

[0055] 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, it reduces manual intervention and developers can save a lot of development time and debugging costs, especially in multi-tasking processing and large-scale model training.

[0056] 3. In the present invention, in a complex deep learning system, multiple modules work together to share computing results, optimize parameters, and save resources. This cross-module collaboration mechanism significantly improves training efficiency and reduces computing costs, especially in large-scale systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram comparing the experimental data verification effects under different tasks in Example 4.

[0058] Figure 2 Schematic diagram comparing the number of iterations and GPU memory usage in Example 4.

[0059] Figure 3 Schematic diagram showing the comparison of the training duration of Example 4 with the number of iterations.

[0060] Figure 4 This is a comparative diagram showing the improvement in accuracy of Example 4. DETAILED DESCRIPTION

[0061] For ease of understanding of the present invention, the present invention will be described in more detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that when an element is described as "fixed to" another element, it can be directly on the other element, or there can be one or more centered elements therebetween. When an element is described as "connected to" another element, it can be directly connected to the other element, or there can be one or more centered elements therebetween. The terms "vertical", "horizontal", "left", "right" and similar expressions used in this specification are for illustrative purposes only.

[0062] Unless otherwise defined, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification are intended solely for the purpose of describing specific embodiments and are not intended to limit the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0063] Example 1:

[0064] An intelligent parameter sharing method comprises the following steps:

[0065] Step 1: Input data and extract multimodal features. Multimodal features include image features, text description features, and contextual features. Image features are extracted using convolutional neural networks (such as ResNet and Vision Transformer) to extract color distribution, texture information, and edge features. Text description features are semantically encoded using cross-modal embedding models (such as BERT and CLIP text encoders). Contextual features include task context (encoding the type / goal / constraints of the current task), environmental context (encoding hardware resource status and 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).

[0066] Step 2: Map the extracted image features, text description features, and context features into feature vectors through a cross-modal embedding model;

[0067] Step 3: Generate parameter sharing decisions in real time by dynamically adjusting the strategy based on the feature vector;

[0068] Step 4: Based on the parameter sharing decision, the shared parameter pool among multiple model modules is coordinated through a unified management node, and dynamic loading and collaborative optimization of cross-module parameters are achieved based on the parameter dependency graph;

[0069] Step 5: During model training or inference, dynamically adjust the parameter sharing range and computing resource allocation based on real-time task requirements and resource status to balance model performance and computing efficiency and achieve global resource optimization.

[0070] As an implementation method, in step 1, inputting data and performing multimodal feature extraction specifically includes the following steps:

[0071] Step 1.1: Extract the color distribution, texture information and edge features of the image through convolutional neural network;

[0072] 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;

[0073] Step 1.3: Combine contextual features, including task type, device resource status, and user needs, to generate a comprehensive feature vector to guide parameter sharing decisions.

[0074] As an implementation method, in step 3, a parameter sharing decision is generated in real time by 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, specifically including the following steps:

[0075] Step 3.0: Use the cross-modal alignment module to calculate the similarity score between the image feature vector and the text description feature vector, and generate a cross-modal alignment weight matrix.

[0076] Step 3.1: If a reinforcement learning strategy is used, add a cross-modal alignment indicator to the reward function, specifically:

[0077] ,

[0078] in, To generate the quality indicators of the results (such as PSNR or accuracy), is a resource consumption indicator (such as computing time or memory usage), The cross-modal alignment score is calculated by sharing parameters to generate the semantic consistency between the result and the input text. is the weight coefficient;

[0079] Step 3.2: If a rule system strategy is adopted, preset cross-modal alignment conditions in the rule base, including:

[0080] When the semantic similarity between image and text exceeds a threshold (>85%), the semantic encoding layer parameters are forced to be shared; otherwise, they are loaded independently. Cross-modal alignment drives parameter sharing: Parameter sharing decisions are driven not only by task requirements but also by the co-alignment of image and text features, ensuring that shared parameters can simultaneously adapt to the semantic consistency of multimodal inputs. Dynamic adjustment mechanism: In reinforcement learning, the cross-modal alignment score directly influences the optimization direction of the policy network; in rule-based systems, alignment conditions become hard constraints.

[0081] As an implementation method, a reinforcement learning strategy (preferred if the task scenario requires dynamic learning capabilities (e.g., complex and changing input data) generates parameter sharing decisions by specifically including the following steps:

[0082] Step 3.1.1: Construct a policy network with the input as the feature vector and the output as the shared weights or independent loading flags of each model parameter;

[0083] Policy network: input feature vector , output shared weights ,in For the strategy network, represents the shared probability of each parameter;

[0084] Step 3.1.2: Train the policy network using a deep reinforcement learning algorithm (proximal policy optimization), with a reward function based on the quality of the generated results and resource consumption indicators;

[0085] Reward function: ,

[0086] in, To generate the quality indicators of the results (such as PSNR or accuracy), is a resource consumption indicator (such as computing time or memory usage), is the weight coefficient,

[0087] Step 3.1.3: During the inference phase, dynamically optimize the sharing strategy based on the input features in real time to minimize computational cost and maximize task performance.

[0088] Policy optimization: Update network parameters θ through the policy gradient method: ,

[0089] in, represents the expected return value to be maximized, represents the partial derivative of the parameter θ, Indicates that in the strategy The expected value under It's action. Indicates execution of an action Instant rewards received, represents the partial derivative with respect to the parameter θ, where In a given feature Select Action probability.

[0090] The rule system strategy (if the task rules are clear and deterministic control is required (such as fixed style classification), choose the rule engine) to generate parameter sharing decisions. The specific steps include:

[0091] Step 3.2.1: Preset a shared rule base, which contains parameter sharing conditions based on color distribution, semantic emotion, and task type;

[0092] Defining the rule base ,in For rule conditions (such as color distribution threshold ), For the corresponding sharing strategy;

[0093] Step 3.2.2: Use the rule engine to match the feature vector with the rule base and trigger the corresponding parameter sharing strategy;

[0094] Matching process: ,

[0095] in, is an indicator function, when the condition If satisfied, its value is 1; otherwise, it is 0;

[0096] First, iterate over all strategies : For each strategy , calculate the conditions that meet Features Then calculate the number of features that meet the conditions: For each feature , check whether the conditions are met , if satisfied, the counter increases by 1; finally, the optimal strategy is selected : Find the strategy with the largest number of features that meet the conditions .

[0097] As an implementation method, in step 4, a shared parameter pool among multiple model modules is coordinated by a unified management node. The unified management node organizes the shared parameter pool using a ring matrix structure (e.g., mapping parameter addresses through a hash table) and implements cross-module collaboration through the following steps:

[0098] Step 4.1: Build a parameter dependency graph between modules, mark the call paths of shared parameters, and ensure that shared parameters are called consistently across multiple tasks;

[0099] Graph of parameter dependencies between building blocks:

[0100] ,

[0101] Among them, the adjacency matrix It is the size of The matrix, the element value is 0 or 1, Representation parameters Dependent parameters ; Representation parameters Not dependent on parameters ;

[0102] Step 4.2: Use memory sharing technology and an asynchronous task scheduling system to coordinate parallel computing requests from multiple modules, reducing duplicate computing and improving resource utilization efficiency.

[0103] The goal of memory sharing techniques is to minimize repeated computations:

[0104] ,

[0105] in, Assign vectors to parameter loading locations;

[0106] Step 4.3: Adjust the parameter loading order according to module priority and task urgency through the dynamic allocation strategy of the parameter pool;

[0107] Based on module priority and task urgency , assign parameters loading order:

[0108] Priority score equals , This is element-by-element multiplication. By marking the call paths of shared parameters, we ensure that when different modules share parameters, the dependencies are clear and consistent. By analyzing parameter dependencies, we can reduce duplicate computations (for example, multiple tasks that rely on the same parameter can share computation results).

[0109] As an implementation method, in step 5, the parameter sharing range and computing resource allocation are dynamically adjusted according to the real-time task requirements and resource status. The dynamic adjustment specifically includes the following steps:

[0110] Step 5.1: In resource-constrained scenarios, load the task-specific layer parameters independently and share the parameters of the base network layer;

[0111] Step 5.2: Based on the real-time inference speed and memory usage feedback, the closed-loop control algorithm adaptively adjusts the sharing ratio to optimize the end-to-end latency.

[0112] Adaptively adjust the sharing ratio through a closed-loop control algorithm:

[0113] ,

[0114] in, is the sharing ratio, For target delay, is the current delay, is the learning rate; if > This indicates that the current delay is too high, and the algorithm increases the sharing ratio. This allows for sharing more parameters, reducing computational effort and thus lowering latency. It is the benchmark value for system performance optimization, dynamically balancing computing efficiency (latency) and task quality (such as the accuracy of generated results) through closed-loop feedback.

[0115] 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.

[0116] Task scheduling optimization goals:

[0117] ,

[0118] in, For the task The amount of resources allocated, is the weight, The total resource budget.

[0119] Example 2:

[0120] This embodiment is basically the same as embodiment 1, except that the multimodal feature extraction and cross-modal alignment steps used in this embodiment extract local features of the product image (such as the sole texture and breathable mesh structure) through EfficientNet to generate image feature vectors. ;

[0121] The BERT model is used to semantically encode the text input by the user and generate a text feature vector ;

[0122] Calculate the semantic similarity score between image and text:

[0123] ,

[0124] And generate the alignment weight matrix , the association strength between the labeled image area and the text keywords (e.g., “breathable” corresponds to the mesh area of ​​the shoe upper);

[0125] Dynamic parameter sharing strategy:

[0126] Policy network input: fused feature vector ;

[0127] Shared decision output: output shared weights , controls the parameter sharing ratio of 100 candidate product recommendation models;

[0128] Reward function:

[0129] ,

[0130] The cross-modal alignment score Directly correlate the consistency between recommended products and user intent (e.g., whether the recommendation matches both the text "breathable" and the mesh design in the image);

[0131] Strategy optimization: If >0.9;

[0132] Rule system strategy: Preset rule base: Rule 1: If >0.85, Rule 2: If the proportion of "sports shoes" in the user's historical click records is greater than 70%, the user preference model parameters are independently loaded; Rule triggering: When the user uploads an image and the text contains "summer" and "breathable", Rule 1 is triggered, reusing the underlying feature extraction parameters of the image classifier;

[0133] Parameter sharing and resource optimization: Parameter dependency graphs are constructed to mark the parameter dependency paths of the product recommendation model (e.g., "image classifier → attribute encoder → recommendation ranker") to ensure consistent call of shared parameters. Dynamic resource allocation is also implemented, with a closed-loop control algorithm used to adjust the sharing ratio during peak periods (e.g., "Double 11").

[0134] ,

[0135] Stabilize end-to-end response latency below 200ms;

[0136] The asynchronous task scheduling system prioritizes tasks with high alignment scores ( >0.8), allocate more GPU resources.

[0137] Substantive technical effects of this embodiment:

[0138] Improved recommendation accuracy: The cross-modal alignment mechanism increased the matching degree (F1 score) between recommended items and user intent from 78% to 89%; user click-through rate increased by 15%, especially in mixed input (text + image) scenarios;

[0139] Resource efficiency optimization: The parameter sharing strategy reduces repeated calculations by 30%, and the peak server CPU load drops by 25%. Response latency is reduced from an average of 350ms to 180ms, supporting tens of thousands of concurrent requests per second.

[0140] Enhanced scenario adaptability: Reinforcement learning strategies dynamically adjust sharing ratios during promotional seasons to cope with traffic fluctuations (such as the "Double 11" peak period), ensuring that response latency remains stable below 200ms; the rule system provides deterministic guarantees in fixed categories (such as "sports shoes") to avoid unexpected performance degradation.

[0141] Example 3:

[0142] An intelligent parameter sharing system, comprising:

[0143] Feature extraction module, which is used to extract multimodal features through convolutional neural networks and cross-modal embedding models;

[0144] Dynamic adjustment strategy module, used to generate parameter sharing decisions based on multimodal features, supporting switching between reinforcement learning strategies and rule system strategies;

[0145] Parameter pool and unified management node, used to store shared parameters and coordinate cross-module collaboration;

[0146] The task scheduling module is used to dynamically allocate computing resources based on resource status and optimize the training and inference processes through asynchronous scheduling;

[0147] 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.

[0148] As an implementation method, the parameter pool and the unified management node use distributed memory sharing technology to coordinate parallel requests of multiple model modules through an asynchronous task scheduling system;

[0149] The task scheduling module has a built-in priority queue that dynamically allocates access rights to the shared parameter pool based on task urgency and module computing load, and adjusts the parameter loading order based on resource occupancy feedback.

[0150] Example 4:

[0151] This example is basically the same as Example 1, except that this example was experimentally verified in multiple typical deep learning tasks (such as image classification and text generation), and achieved significant results. The specific experimental data is as follows:

[0152] Experimental setup:

[0153] Datasets: Various standard datasets were used for validation, including ImageNet, CIFAR-10, COCO (for image classification), and OpenWebText (for text generation).

[0154] Evaluation indicators: Accuracy, F1 score, training time, inference speed, memory usage and other indicators are used for evaluation.

[0155] Comparison Baseline: This technology solution is compared with traditional static parameter sharing models (such as traditional VGG, ResNet, LSTM, etc.) to evaluate the advantages of this technology solution in resource optimization and task adaptability.

[0156] Experimental results:

[0157] Refer to the accompanying drawings in the specification Figures 1-4It can be seen that: Parameter reduction: In image classification tasks, this technical solution reduces the number of model parameters by more than 20% under the sharing strategy optimization. This is mainly due to the dynamic adjustment of the sharing range, which avoids redundant model parameter loading; Performance improvement: In multiple tasks, the performance (such as accuracy) of this solution has increased by an average of 5%-10%. In image classification tasks in particular, by sharing convolutional layer parameters, the model significantly reduces the computational load without sacrificing accuracy; Training time reduction: During the training process, due to the automatic management of intelligent shared nodes, the training time is reduced by 15%-20%. Sharing computing resources and optimizing memory usage have significantly improved overall training efficiency; Inference speed improvement: Due to the reduction in the number of model parameters and the optimization of cross-module collaboration, the inference speed has increased by 10%-15%, ensuring the efficient execution of real-time tasks.

[0158] Compared to traditional techniques, this solution demonstrates lower computational resource consumption, greater model adaptability, and significant performance improvements across multiple tasks. In traditional approaches, parameter sharing is typically statically configured and cannot be flexibly adjusted based on task requirements. However, this solution's context-aware and dynamic adjustment strategy enables the model to select the optimal parameter sharing strategy in real time.

[0159] Facing future challenges, this technical solution offers excellent scalability, adapting to larger datasets and more complex tasks. For example, for the potential emergence of multimodal data (such as mixed inputs of text, images, and audio) and more complex generation tasks (such as generating videos and 3D scenes), this solution can further optimize sharing strategies to ensure the optimal balance between resource utilization and computational efficiency.

[0160] It should be noted that the preferred embodiments of the present invention are given in the specification and drawings 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 do not serve as additional limitations on the content of the present invention. The purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive. In addition, the above-mentioned technical features continue to be combined with each other to form various embodiments not listed above, which are all considered to be within the scope of the description of the present invention; further, it is obvious to those skilled in the art that improvements or changes can be made based on the above description, and all such improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. An intelligent parameter sharing method, characterized in that: The following steps are involved: Step 1: Input data and perform multimodal feature extraction. Multimodal features include image features, text description features, and context features. The following steps are also included: 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 contextual features, including task type, device resource status, and user needs, to generate a comprehensive feature vector to guide parameter sharing decisions 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 through a dynamic adjustment strategy, where the dynamic adjustment strategy may be implemented using a reinforcement learning strategy or a rule system strategy based on cross-modal feature alignment. The task scenario requires dynamic learning capabilities, so a reinforcement learning strategy is selected; The task rules are clear and require deterministic control, so a rule system strategy is selected; Calculate the similarity score between the image feature vector and the text description feature vector to generate a cross-modal alignment weight matrix; The reinforcement learning strategy consists of the following steps: Step 3.1.1: Construct a policy network with the feature vector as input and the shared weights or independent loading flags of each model parameter as output; Step 3.1.2: Train the policy network using a deep reinforcement learning algorithm, with a reward function based on the quality of the generated results, resource consumption, and cross-modal alignment. Step 3.1.3: During the inference phase, dynamically optimize the sharing strategy based on the input features in real time to minimize computational cost and maximize task performance. The rule system strategy includes the following steps: Step 3.2.1: Preset a shared rule base, and preset cross-modal alignment conditions in the 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 system to trigger the corresponding parameter sharing strategy; Step 4: Based on the parameter sharing decision, the shared parameter pool among multiple model modules is coordinated through a unified management node, and dynamic loading and collaborative optimization of cross-module parameters are achieved based on the parameter dependency graph; Step 5: During model training or inference, dynamically adjust the parameter sharing range and computing resource allocation based on real-time task requirements and resource status to balance model performance and computing efficiency and achieve global resource optimization.

2. 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 by a unified management node. 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 call paths of shared parameters, and ensure that shared parameters are called consistently across multiple tasks; Step 4.2: Use memory sharing technology and an asynchronous task scheduling system to coordinate parallel computing requests from multiple modules, reducing duplicate computing and improving resource utilization efficiency. Step 4.3: Adjust the parameter loading order according to module priority and task urgency through the dynamic allocation strategy of the parameter pool.

3. The intelligent parameter sharing method according to claim 2, characterized in that: In step 4.1, the parameter dependency diagram between 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 computation: ， Among them, P is the parameter loading position allocation vector.

4. 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. The dynamic adjustment specifically includes the following steps: Step 5.1: In resource-constrained scenarios, load the 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 closed-loop control algorithm adaptively adjusts the sharing ratio 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.

5. The intelligent parameter sharing method according to claim 4, characterized in that: In step 5.2, the sharing ratio is adaptively adjusted through a closed-loop control algorithm. The specific formula is: ， in, is the sharing ratio, For target delay, is the current delay, is the learning rate.

6. An intelligent parameter sharing system, used to implement the intelligent parameter sharing method according to any one of claims 1 to 5, 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, supporting 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 based on resource status and optimize the training and inference 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.

7. The intelligent parameter sharing system according to claim 6, characterized in that: The parameter pool and the unified management node adopt 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, 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