Organizable modular neural architecture search method and system

Abstract

The invention relates to the technical field of computer engineering, and discloses an organizable modular neural architecture searching method and system.The organizable modular neural architecture searching method comprises the steps of building a modular searching space, defining task types and feature parameters, determining the task types based on rules, collecting and defining network components, and verifying and sorting the network components; creating an initial search space and setting hyper-parameters of the components; the search space and the search strategy are dynamically adjusted, and the search process is optimized in real time according to task requirements; parameter adjustment and objective function estimation are carried out through real-time monitoring and optimization strategies. According to the invention, through modularized search space, flexible and efficient neural network architecture design is realized. The predefined module reduces search complexity, improves efficiency, allows a user to customize components, and enhances network flexibility. And dynamic adjustment of a search strategy and real-time monitoring ensure automatic design optimization, and model performance and stability are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and machine learning technology, and specifically to a method and system for searching modular neural architectures. Background Art

[0002] With the rapid development of artificial intelligence and deep learning technology, neural networks have achieved remarkable results in various application fields, such as image classification, natural language processing, and time series prediction. However, designing an efficient neural network architecture is a complex and time-consuming task. Traditional manual design methods rely on expert knowledge and often require repeated trials and adjustments. Neural Architecture Search (NAS) is an automated method that searches for the best network architecture through algorithms, greatly reducing the burden of manual design.

[0003] Traditional NAS methods usually require a lot of computing resources and lack flexibility, and cannot customize the search space according to specific tasks. In addition, these methods often fail to fully utilize existing excellent network components and structures during the search process. Therefore, how to flexibly combine and optimize network components while ensuring search efficiency has become a hot topic and difficulty in current research.

[0004] The modular neural architecture search method (M-NAS) proposed in this paper constructs a modular search space containing a variety of combinable network components, allowing users to select appropriate network components according to specific task requirements and dynamically adjust the search strategy. It can not only achieve the required network structure, but also customize modular components according to task types, and finally search for the optimal hyperparameters in each architecture type. This method not only improves the search efficiency, but also enhances the flexibility and applicability of network architecture design, providing new ideas and tools for neural network optimization research. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is that the traditional NAS method usually requires a large amount of computing resources, lacks flexibility, and cannot customize the search space according to specific tasks.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: a method for organizing modular neural architecture search, comprising: constructing a modular search space, defining task types and feature parameters, determining task types based on rules, collecting and defining network components, verifying and organizing network components, creating an initial search space and setting component hyperparameters; dynamically adjusting the search space and search strategy, and optimizing the search process in real time according to task requirements; and adjusting parameters and estimating objective functions through real-time monitoring and optimization strategies.

[0008] As a preferred solution of the organizeable modular neural architecture search method described in the present invention, after defining the task type and feature parameters, it also includes extracting data types, data dimensions and shapes, and output requirements from the task description and sample data provided by the user.

[0009] As a preferred solution of the organizeable modular neural architecture search method described in the present invention, the rule-based determination of the task type includes using a predefined rule set to match the task type based on the extracted features and parameters combined with the output data type.

[0010] As a preferred solution of the organizeable modular neural architecture search method described in the present invention, the collecting and defining network components includes describing the functions of each network component, including the input-output relationship, calculation method and role of the network component in the network, and listing the adjustable hyperparameters and ranges of each network component.

[0011] As a preferred solution of the method for searching for an organized modular neural architecture described in the present invention, the verification and organization of network components includes: performing functional verification and performance testing on each network component, and judging the ability of the network component to perform the expected function by verifying the performance of the network component in a specific task; modularizing the verified network components, and each modular component Described by function and parameter set Composition, expressed as ; Build a systematic component library, including the functional description, hyperparameter definition, implementation details and verification results of each component.

[0012] As a preferred solution of the method for searching an organized modular neural architecture according to the present invention, the steps of creating an initial search space and setting the hyperparameters of the components include: Represents the combination of all components, And cover various network structures; initial search space ; Design and adjust the hyperparameters of each network component according to the specific task, and incorporate the hyperparameter space H of the specific component into the overall search space ,make .

[0013] As a preferred solution of the organizeable modular neural architecture search method described in the present invention, the parameter adjustment and objective function estimation include, through real-time monitoring and optimization strategy adjustment, using Bayesian optimization method and Gaussian process model to estimate the objective function and optimize the parameters.

[0014] A modular neural architecture search system that can be organized using any of the methods described in the present invention, wherein: a search space construction module constructs a search space containing multiple combinable network components and initializes components and parameters according to task requirements; a search space adjustment module defines task characteristics and related parameters according to user task requirements and dynamically adjusts the search space and search strategy; a search process optimization module uses a Bayesian optimization method and a Gaussian process model to estimate the objective function and optimize parameters through real-time monitoring and optimization strategy adjustment.

[0015] A computer device comprises: a memory and a processor; the memory stores a computer program, comprising: the steps of implementing any one of the methods of the present invention when the processor executes the computer program.

[0016] A computer-readable storage medium stores a computer program, comprising: when the computer program is executed by a processor, the steps of implementing any one of the methods of the present invention are implemented.

[0017] Beneficial effects of the present invention: The method of the present invention realizes flexible and efficient neural network architecture design and optimization by constructing a modular search space containing a variety of combinable network components. By pre-defining the module selection in the search space, irrelevant components are filtered out, the complexity of the search space is reduced, the search efficiency is improved, and the search process is focused on the component combination that is most likely to produce high-quality results. At the same time, the user is allowed to customize the addition or modification of modular components according to specific task requirements and personal experience, which improves the flexibility and adaptability of the network architecture and can better meet the special needs of different tasks. By dynamically adjusting the search strategy, combined with real-time monitoring and optimization strategy adjustment, the efficiency and effectiveness of the search process are ensured, the neural network architecture is automatically designed and optimized, and the burden of manual design is reduced. In addition, through modular design and hyperparameter optimization, the optimal hyperparameter configuration can be searched in each type of network architecture, which improves the performance and stability of the neural network model. The constructed modular search space covers various common network components and structures, such as convolutional layers, pooling layers, fully connected layers, residual blocks, Transformer layers, LSTM units, etc., and is suitable for a variety of tasks such as image classification, natural language processing, and time series prediction, and has broad application prospects. By optimizing the search space and dynamically adjusting the search strategy, the waste of computing resources is reduced, resource utilization is improved, and it is suitable for application in environments with limited computing resources. In summary, the modular neural architecture search method (M-NAS) of the present invention not only has significant advantages in improving search efficiency and model performance, but also enhances the flexibility and automation level of network architecture design, and has important practical application value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them: Figure 1 An overall flow chart of an organized modular neural architecture search method provided by one embodiment of the present invention; Figure 2 A task type matching flow chart of a method for organizing modular neural architecture search provided by one embodiment of the present invention; Figure 3 A structural diagram of a modular search space system that can organize a modular neural architecture search method provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0020] Example 1, reference Figure 1 and Figure 2 , as an embodiment of the present invention, provides an organized modular neural architecture search method, comprising: In S1: build a modular search space, define task types and feature parameters, determine task types based on rules, collect and define network components, verify and organize network components, create an initial search space and set component hyperparameters.

[0021] In S2: Dynamically adjust the search space and search strategy to optimize the search process in real time according to task requirements.

[0022] In S3: Parameter adjustment and objective function estimation are performed through real-time monitoring and optimization strategies.

[0023] Furthermore, we construct a modular search space that contains a variety of composable network components. By creating a flexible and extensive search space, we ensure that various network structures and hyperparameter configurations can be covered, thus providing a suitable infrastructure for different tasks.

[0024] Task type feature definition: Based on common task types, define the features and parameters of each task type. The following is a detailed classification of task types and their features: Task type feature definition. Based on common task types, define the features and parameters of each task type. Image processing tasks include image classification (input is image, output is category label), object detection (input is image, output is the location and category of the object), image segmentation (input is image, output is pixel-level classification label), image generation (input can be random noise or specific conditions, output is the generated image), image restoration (input is damaged image, output is restored image) and style transfer (input is two images, output is style-transferred image).

[0025] Natural language processing tasks include text classification (the input is text and the output is the category label of the text), sentiment analysis (the input is text and the output is the sentiment label, such as positive or negative), machine translation (the input is source language text and the output is target language text), text generation (the input is partial text or keywords and the output is generated text), question answering system (the input is questions and the output is answers), named entity recognition (the input is text and the output is entities in the text and their categories), summary generation (the input is long text and the output is a short summary) and text error correction (the input is text containing errors and the output is the corrected text).

[0026] Time series tasks include time series prediction (the input is historical time series data, and the output is the future time series value), anomaly detection (the input is time series data, and the output is the detection result of the anomaly point), and time series classification (the input is time series data, and the output is the category label of the time series).

[0027] Speech processing tasks include speech recognition (the input is speech signal and the output is text transcription), speech synthesis (the input is text and the output is speech signal) and speech emotion recognition (the input is speech signal and the output is emotion label).

[0028] Recommender system tasks include recommender systems (the input is features of users and items, and the output is recommended items).

[0029] Reinforcement learning tasks involve reinforcement learning (the input is the state of the environment and the output is the action decision).

[0030] The Generative Adversarial Network task involves generating adversarial networks (the input is noise and the output is a generated sample).

[0031] Graph neural network tasks include node classification (the input is graph structure data, and the output is the category label of the node) and link prediction (the input is graph structure data, and the output is the predicted link).

[0032] Feature extraction and parameter analysis. Extract features and parameters from the task description and sample data provided by the user, such as data type (image, text, time series, speech, graph structure, etc.), data dimension and shape, and output requirements (classification label, continuous value, pixel-level classification, etc.) Rule-based task type determination: Use predefined rule sets to match task types based on extracted features and parameters.

[0033] Specifically, if the input data is an image and the output is a category label, the task type is image classification; if the input data is an image and the output is the location and category of the target, the task type is target detection; if the input data is text and the output is a sentiment label, the task type is sentiment analysis; if the input data is a time series and the output is a future value, the task type is time series prediction; if the input data is speech and the output is text transcription, the task type is speech recognition; if the input data is user and item features and the output is a recommended item, the task type is a recommendation system.

[0034] Collect and define network components. This step includes the following sub-steps B1~B5: In B1, classification and preliminary screening of network components.

[0035] According to different task types and requirements, the available network components are classified and preliminarily screened. These components include but are not limited to convolutional layers, pooling layers, fully connected layers, activation functions, regularization layers, normalization layers, residual blocks, Transformer layers, recurrent neural network units, graph convolutional layers, generative adversarial network modules, etc. Each component has specific functions and application scenarios. For example, some components are suitable for image processing tasks, while others are suitable for natural language processing or time series analysis tasks.

[0036] In B2, the functional description of each network component is defined.

[0037] Provide a detailed functional description of each collected network component, including its input-output relationship, calculation method, and role in the network. The functional description should clearly state the specific operation and expected effect of the component so that users can understand and use it correctly.

[0038] In B3, define the tunable hyperparameters of the network components.

[0039] List the adjustable hyperparameters and their ranges for each network component. Specifically, the hyperparameters of the convolution layer include the convolution kernel size and step size; the hyperparameters of the pooling layer include the pooling method and window size; the hyperparameters of the recurrent neural network unit include the hidden layer size and time step. The definition of each hyperparameter must clearly specify its value range and default value to ensure the operability of the component under different configurations.

[0040] In B4, the validity of the network components is verified.

[0041] Functional verification and performance testing are performed on each network component to ensure its effectiveness and stability in different mission scenarios. For example, by verifying the performance of the component in its specific task, it is ensured that it can correctly perform the expected function and remain stable under various input conditions.

[0042] In B5, a network component library is established.

[0043] Organize the verified network components into a systematic component library, including the functional description, hyperparameter definition, implementation details and verification results of each component. The component library should be easy to expand and maintain, so that new components can be added or existing components can be modified in the future.

[0044] Furthermore, the network component library is arranged and modularized so that it can be flexibly combined.

[0045] Each modular component Described by its function and parameter set Composition, that is .

[0046] Specifically, the convolutional layer module can be described as , the pooling layer module can be described as . Making it modular enables flexibility and scalability, and components can be adjusted and updated at any time according to needs.

[0047] Create an initial search space S containing all possible component combinations C, where The initial search space should be as broad as possible to ensure that a variety of possible network structures are covered. For example, in an image classification task, the initial search space can include different combinations of convolutional layers, pooling layers, and fully connected layers.

[0048] Define component-specific hyperparameters Define specific hyperparameters for different network structure components.

[0049] Specifically, for convolutional neural networks (CNNs), we define hyperparameters such as convolution kernel size k, step size s, and pooling type; for recurrent neural networks (RNNs), we define hidden layer size h, time step t, etc. The hyperparameter space H of these specific components is incorporated into the overall search space S, so that The definition and selection of hyperparameters are crucial for network performance optimization and need to be designed and adjusted in detail according to specific tasks.

[0050] Furthermore, the module selection in the search space is predefined according to the specific task requirements. By filtering out irrelevant components and selecting the most relevant modules, the complexity of the search space is reduced, the search efficiency is improved, and the search process is ensured to focus on the component combination that is most likely to produce high-quality results.

[0051] Filter out irrelevant network components based on task type.

[0052] Specifically, in regression tasks, we exclude classification-specific components. In natural language processing tasks, we may need to exclude components related to image processing, such as convolutional layers and pooling layers. By filtering irrelevant components, we can significantly reduce the search space and improve search efficiency.

[0053] Select components suitable for the task, either manually by the user or automatically by a recommendation system. Recommendation systems generate recommendation lists based on task features and historical data using machine learning algorithms such as collaborative filtering or content-based recommendations. For example, in an image classification task, a recommendation system may suggest a combination of multiple convolutional layers and pooling layers.

[0054] Adjust and optimize the search space to ensure that it contains the most relevant components and hyperparameter configurations. The optimization of the search space can be achieved through an optimization algorithm based on the objective function FFF, such as particle swarm optimization (PSO) or genetic algorithm (GA). For example, a genetic algorithm can be used to perform multiple rounds of iterations in the initial search space to gradually optimize the component combination and hyperparameter configuration.

[0055] After adjusting and optimizing the search space, the method further includes allowing the user to add or modify modular components according to task requirements, thereby increasing the flexibility of the search space and enabling the user to adjust the network structure according to specific task requirements and personal experience, thereby improving model performance and adaptability, including: Users add or modify components Users add or modify network components based on task requirements and experience. For example, users can add custom normalization layers or regularization layers to a convolutional neural network to improve model performance.

[0056] Update the search space to include user-defined modular components. Newly added components should include their functional descriptions and parameter sets to ensure compatibility with existing components. For example, if a user adds a new activation function layer, the component combination and hyperparameter configuration in the search space need to be updated.

[0057] Verify the compatibility of newly added components with existing components. The verification process includes functional testing and performance evaluation, using cross-validation methods to evaluate the performance of new components in different task scenarios. For example, cross-validation can be used to evaluate the performance of a newly added regularization layer on different data sets to ensure its compatibility and effectiveness.

[0058] Furthermore, the search strategy is dynamically adjusted according to the search progress. Through real-time monitoring and optimization strategy adjustment, the efficiency and effectiveness of the search process can be ensured, problems in the search process can be discovered and corrected in a timely manner, and ultimately the optimal network architecture and hyperparameter configuration can be found.

[0059] Monitor search progress Monitor search progress, including performance indicators (such as accuracy, loss function value, etc.). Based on the Bayesian optimization (BO) method, the search strategy is updated through the posterior probability of the model. Bayesian optimization estimates the objective function through the Gaussian process (GP) model and obtains the function by maximizing it. Find the optimal parameter combination. The definition of the acquisition function is: , in, is the mean estimate of the objective function, is the standard deviation, is a trade-off parameter.

[0060] Dynamically adjust the search strategy based on the monitoring results, for example, change the parameters of the search algorithm, increase or decrease the scope of the search space. Dynamic adjustment strategies can include adjusting the learning rate, modifying the number of network layers, etc. For example, if the current search progresses slowly, the scope of the search space can be increased to find more possible network structures.

[0061] Evaluate the effect of adjustment strategies in real time to ensure the efficiency of the search process. Evaluation methods include real-time monitoring and evaluation reports based on performance indicators. Optimization algorithms such as gradient descent (SGD) and random search (RS) can be applied to this process. Through real-time evaluation, problems in the search process can be discovered and corrected in a timely manner to ensure the accuracy and effectiveness of the final search results.

[0062] Through the steps of the above implementation scheme, users can select appropriate network components according to specific task requirements in a flexible and combinable search space, and finally find the optimal network architecture and hyperparameter configuration by dynamically adjusting the search strategy and optimizing hyperparameters. The present invention not only improves the search efficiency, but also enhances the flexibility and applicability of network architecture design, providing new ideas and tools for neural network optimization research.

[0063] If the above functions are implemented in the form of 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 the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0064] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0065] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0066] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0067] Embodiment 2, in an exemplary embodiment, an organizeable modular neural architecture search system is also provided, including a search space construction module, which constructs a search space containing a variety of combinable network components and initializes components and parameters according to task requirements.

[0068] The search space adjustment module defines task characteristics and related parameters according to user task requirements and dynamically adjusts the search space and search strategy.

[0069] The search process optimization module uses Bayesian optimization method and Gaussian process model to estimate the objective function and optimize parameters through real-time monitoring and optimization strategy adjustment.

[0070] The specific module structure is as follows Figure 3 As shown, the search space construction module includes a task type feature definition and extraction module, a rule matching module, a modular search space construction module, a search space design module, a component classification module and a network component management module.

[0071] After the user inputs the task requirements and data, the task type feature definition and extraction module defines the task type features and parameters according to the user input, extracts the relevant feature information, and passes the task type features and parameters to the rule matching module.

[0072] The rule matching module matches the extracted features and parameters based on predefined rules, determines the task type, and passes the determined task type to the building module of the modular search space.

[0073] The building modules of the modular search space include a search space design module and a component classification module. The component classification module classifies different types of network components according to their functions; the search space design module constructs the search space of combinable network components, designs the combination methods of different components, and passes the component combination information to the network component management module.

[0074] The network component management module accepts inputs from search space design and component classification, and processes four subtasks in parallel: component collection, functional description, component verification, and component organization. It manages network components and collects, classifies, verifies, and organizes system components. It passes component functional descriptions, verification results, and organized component information to the search space creation and adjustment module.

[0075] The search space creation and adjustment module creates the initial search space, defines the hyperparameters of the components, and dynamically adjusts the search space and search strategy. It is connected to the user customization and monitoring optimization module and receives feedback from user customization and real-time monitoring optimization.

[0076] The user customization and monitoring optimization module allows users to customize network components, monitor the search process and optimize the search strategy, and use real-time data to optimize the search process. It is connected to the "search space creation and adjustment module" to provide feedback to adjust the search strategy and search space.

[0077] Embodiment 3, in an exemplary embodiment, a computer program product is further provided, including a computer program, and when the computer program is executed by a processor, the following method steps are implemented, including: Build a modular search space, define task types and feature parameters, determine task types based on rules, collect and define network components, verify and organize network components, create an initial search space and set component hyperparameters; Dynamically adjust the search space and search strategy to optimize the search process in real time according to task requirements; Parameter adjustment and objective function estimation are performed through real-time monitoring and optimization strategies.

[0078] In one embodiment, the computer program may perform the following steps: Build a modular search space: Build a search space that contains a variety of network components that can be freely combined to meet different task requirements.

[0079] Define task type features and parameters: Extract task features and parameters from the task description and sample data provided by the user.

[0080] Extract features and parameters: Extract relevant features such as data type, dimension, and shape by analyzing the task input data.

[0081] Rule-based task type determination: Use a predefined set of rules to match the task type, for example, if the input is an image and the output is a class label, it is an image classification task.

[0082] Collect and define network components: Define the functions, hyperparameters, and application scenarios of network components in detail to ensure the operability of components under different configurations.

[0083] Verify and organize network components: Perform functional verification and performance testing on components to ensure their effectiveness and stability in different mission scenarios.

[0084] Create an initial search space: Create an initial search space that contains all possible component combinations, ensuring that a wide range of network structures are covered.

[0085] Define component hyperparameters: Set appropriate hyperparameter ranges for each component and include them in the overall search space.

[0086] Dynamically adjust the search space and search strategy: Dynamically adjust the scope and strategy of the search space based on the monitoring results, such as adjusting the learning rate or the number of network layers.

[0087] Allow users to customize components: Users can add or modify network components according to task requirements and integrate customized components into the search space.

[0088] Monitor and optimize the search process: By monitoring the search progress in real time, use Bayesian optimization methods and Gaussian process models to adjust the search strategy to ensure the efficiency of the search process.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for searching for an organized modular neural architecture, characterized in that: include: Build a modular search space, define task types and feature parameters, determine task types based on rules, collect and define network components, verify and organize network components, create an initial search space and set component hyperparameters; Dynamically adjust the search space and search strategy to optimize the search process in real time according to task requirements; Parameter adjustment and objective function estimation are performed through real-time monitoring and optimization strategies.

2. The method for searching an organizable modular neural architecture according to claim 1, wherein: After defining the task type and characteristic parameters, it also includes: Extract data types, data dimensions and shapes, and output requirements from the task description and sample data provided by the user.

3. The method for searching for an organizable modular neural architecture according to claim 2, wherein: The rule-based determination of the task type includes using a predefined rule set to match the task type according to the extracted features and parameters in combination with the output data type.

4. The method for searching for an organizable modular neural architecture according to claim 3, wherein: The collecting and defining of network components includes: Describe the function of each network component, including the input-output relationship, calculation method, and role in the network, and list the adjustable hyperparameters and ranges of each network component.

5. The method for searching for an organizable modular neural architecture according to claim 4, wherein: The verification and arrangement of the network components includes performing functional verification and performance testing on each network component, and judging the ability of the network component to perform the expected function by verifying the performance of the network component in a specific task; Modularize the proven network components. Described by function and parameter set Composition, expressed as ; Build a systematic component library, including the functional description, hyperparameter definition, implementation details and verification results of each component.

6. The method for searching for an organizable modular neural architecture according to claim 5, wherein: The steps of creating the initial search space and setting the hyperparameters of the components include: use Represents the combination of all components, And covers various network structures; Initial search space ; Design and adjust the hyperparameters of each network component according to the specific task, and incorporate the hyperparameter space H of the specific component into the overall search space ,make .

7. The method for searching for an organizable modular neural architecture according to claim 6, wherein: The parameter adjustment and objective function estimation include, Through real-time monitoring and optimization strategy adjustment, Bayesian optimization method and Gaussian process model are used to estimate the objective function and optimize parameters.

8. An organisable modular neural architecture search system using any of the methods of claims 1 to 7, characterized in that: include, Search space construction module, which constructs a search space containing a variety of composable network components and initializes components and parameters according to task requirements; The search space adjustment module defines the task characteristics and related parameters according to the user's task requirements and dynamically adjusts the search space and search strategy; The search process optimization module uses Bayesian optimization method and Gaussian process model to estimate the objective function and optimize parameters through real-time monitoring and optimization strategy adjustment.

9. A computer device comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, the steps of the organizeable modular neural architecture search method as described in any one of claims 1-7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the organizeable modular neural architecture search method as described in any one of claims 1-7 are implemented.

Images