Multi-cause regulation AI large model training method and intelligent decision-making system

Abstract

The invention relates to the technical field of multi-modal feature fusion, in particular to a training method of a multi-cause regulation AI large model and an intelligent decision making system. Through data preprocessing, feature expression optimization, class center construction of support data, efficient distance calculation, dynamic parameter adjustment and a self-adaptive hyper-parameter and field self-adaptive strategy, the multi-cause regulation AI large model is obtained; according to the technical scheme, the robustness and accuracy of model training are remarkably improved. Meanwhile, the modularized hardware implementation scheme ensures the high efficiency and stability of the overall operation of the system, and can adapt to the requirements of the data size and the field change, thereby providing an efficient, stable and intelligent neural network model training platform for an intelligent decision-making system.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal feature fusion, and specifically to a training method for a multi-factor regulated AI large model and an intelligent decision-making system. Background Art

[0002] Existing neural network training methods generally rely on large-scale data for training, and improve the consistency of data distribution through methods such as batch normalization and ordinary data preprocessing; feature extraction mainly relies on the output of pre-trained models, and subsequent classification and regression tasks are achieved through conventional network layers, and the feature expression methods are relatively simple; parameter updates usually adopt fixed optimizers and fixed update strategies, and hyperparameter settings rely on manual tuning or grid search, lacking a dynamic adaptive mechanism; for the problem of domain migration, existing methods usually directly fine-tune pre-trained models, making it difficult to cope with the challenges brought by changes in data volume and data distribution.

[0003] Most existing systems are implemented in software, with low integration of hardware platforms, and there are bottlenecks in data transmission and collaborative scheduling between modules;

[0004] Lack of a complete mechanism for real-time feedback and hyperparameter adaptive adjustment during the training process, resulting in difficulty for the model to continuously maintain the best performance.

[0005] 1. In the existing technology for data normalization and batch normalization processing, the problem of inconsistent data distribution in different batches cannot be fully solved, resulting in slow model convergence speed and easy to fall into local optimum;

[0006] 2. When traditional methods calculate category features, they usually only rely on a single feature expression, making it difficult to capture the distribution characteristics of data within the category. At the same time, in a large-scale data environment, the distance calculation efficiency is low;

[0007] 3. In the existing technology, parameter updates mostly adopt fixed strategies and cannot be dynamically adjusted according to data characteristics and fluctuations during the training process, and it is easy to cause training instability due to large fluctuations in parameter updates;

[0008] 4. Traditional methods rely on manual tuning or static search methods for hyperparameter adjustment, and only use simple fine-tuning means for the problem of domain migration, making it difficult to achieve adaptive optimization;

[0009] 5. Existing training devices have scattered modules and low data bus transmission efficiency, making it difficult to achieve efficient collaboration; at the same time, lack of real-time monitoring and feedback mechanism during the training process, making it difficult to adjust training strategies in a timely manner. Summary of the Invention

[0010] The purpose of the present invention is to provide a training method for a multi-factor regulated AI large model and an intelligent decision-making system, so as to solve the problems raised in the above-mentioned background technology. Through data preprocessing, feature expression optimization, construction of class centers for support data, efficient distance calculation, dynamic parameter adjustment, as well as adaptive hyperparameters and domain adaptation strategies, this technical solution significantly improves the robustness and accuracy of model training. At the same time, its modular hardware implementation scheme ensures the high efficiency and stability of the overall operation of the system, and can adapt to the requirements of data volume size and domain changes, thus providing an efficient, stable, and intelligent neural network model training platform for the intelligent decision-making system.

[0011] To achieve the above purpose, the present invention provides the following technical solutions:

[0012] A method for training a neural network model, comprising the following steps:

[0013] Obtain a neural network model pre-trained based on second training data, first training data, and the class information of the first training data. The first training data includes support data and query data, and the support data is all or part of the data of each category in the first training data, and the query data is all or part of the data of each category in the first training data;

[0014] Preprocess the first training data. The preprocessing steps include normalization and batch normalization to ensure consistent data distribution in different batches;

[0015] Use the neural network model to extract features from the first training data, and perform multi-perspective feature fusion and improved deep hashing processing on the extracted features to obtain a compact and robust feature representation;

[0016] Calculate the class center features of each category according to the features of the support data, where each one-dimensional value of the class center features of each category is the average value of the corresponding feature dimensions of the support data within that category;

[0017] Use the efficient approximate nearest neighbor ANN algorithm and the mini-batch method to calculate the feature distances between the query data features and the class center features of each category;

[0018] According to the feature distances between the class center features of each category and the query data features, and the average distances between the features of the first training data within each category, dynamically adjust the parameters of some layers in the neural network model to obtain an adjusted neural network model, where regularization, gradient smoothing, and an adaptive optimizer are introduced during the parameter adjustment process to ensure update stability;

[0019] When the data volume of the first training data is less than a preset value, the Bayesian optimization scheme is adopted to automatically search for the optimal hyperparameters within the predefined hyperparameter search space, and on this basis, the parameters of the partial layers are adjusted;

[0020] When the data volume of the first training data is greater than or equal to the preset value, according to the preset hyperparameters corresponding to the neural network model and in combination with the distance information between the query data and the class center features, an incremental training and domain adaptation module is used to perform smooth migration on the pre-trained model to adapt to the distribution of new domain data;

[0021] The training result is fed back, the performance of the model on the validation set is monitored in real time, and hyperparameter fine-tuning and model retraining are automatically triggered when necessary according to the monitoring result.

[0022] In a specific embodiment, the deep hashing processing includes, after the neural network model extracts features, first passing through a batch normalization layer, and then using a contrastive learning mechanism to constrain positive and negative samples to obtain a more robust and compact feature representation.

[0023] In a specific embodiment, in the dynamic parameter adjustment step, the parameter update adopts a smooth update mechanism with a weight update buffer, and this buffer uses the parameter update mean value of the past several training cycles as a reference for parameter adjustment in this cycle to reduce the volatility of single-step updates.

[0024] In a specific embodiment, the hyperparameter adaptive adjustment scheme adopts Bayesian optimization when the data volume is small, and sets a triggering mechanism when the data volume is large. When the validation set metrics show no significant improvement in consecutive several iterations, fine-tuning is triggered to further optimize the learning rate, batch size, and network structure parameters.

[0025] An intelligent decision-making system for the method according to any one of claims 1-4, the intelligent decision-making system comprising:

[0026] An acquisition module, configured to acquire a neural network model, first training data, and its category information;

[0027] A preprocessing module, configured to perform normalization and batch normalization processing on the first training data;

[0028] A feature extraction module, configured to extract features of the first training data by using the neural network model, and perform multi-view feature fusion and deep hashing processing;

[0029] A calculation module, configured to calculate the distance between the class center features composed of support data and the query data features by using the ANN and mini-batch methods;

[0030] A parameter adjustment module, which is used to dynamically adjust the parameters of some layers in the neural network model according to the feature distance information and the hyperparameters obtained by preset or Bayesian optimization, and introduce a regularization and smooth update mechanism;

[0031] A domain adaptation module, which is used to achieve smooth migration of the pre-trained model to a new data domain when the data volume is sufficient;

[0032] A feedback module, which is used to monitor the training effect in real time and trigger hyperparameter fine-tuning or retraining according to the feedback.

[0033] In a specific embodiment, the acquisition module, the preprocessing module, the feature extraction module, the calculation module, the parameter adjustment module, the domain adaptation module and the feedback module are all interconnected through a high-speed data bus and cooperate on a hardware platform.

[0034] Through multiple innovative steps and modular design, this technical solution realizes comprehensive optimization in aspects such as feature extraction, model update, and hyperparameter adaptive adjustment during the training process of the neural network model, thus bringing the following beneficial effects:

[0035] 1. Data preprocessing and feature expression optimization

[0036] Normalization and batch normalization processing ensure the consistency of the distribution of different batches of data, reducing the impact of data deviation on training stability and convergence speed;

[0037] After extracting features through the pre-trained model, combined with multi-view feature fusion and improved deep hashing processing, a compact and robust feature representation is obtained. This not only improves the expression ability of features, but also is beneficial to subsequent distance calculation and classification tasks, thereby improving the model recognition accuracy.

[0038] 2. Class center construction based on support data and efficient distance calculation

[0039] Calculate the class center features of each category using support data, and integrate the features within the category by taking the average value, effectively reflecting the central distribution of each category;

[0040] Adopt an efficient approximate nearest neighbor (ANN) algorithm and mini-batch method to calculate the feature distance between the query data and the class centers of each category, which not only reduces the computational complexity, but also accelerates the distance calculation process, facilitating real-time update in a large-scale data environment.

[0041] 3. Dynamic parameter adjustment and improvement of model stability

[0042] Dynamically adjust the parameters of some layers of the neural network based on the feature distance information between the support data and the query data. By introducing mechanisms such as regularization, gradient smoothing, and adaptive optimizers, the volatility during the parameter update process is effectively alleviated, ensuring the stability of the training process;

[0043] Adopt a smooth update mechanism with a weight update buffer. By referring to the parameter update mean of several past training cycles, the oscillation risk brought by single-step updates is reduced, and the continuity and reliability of model parameter adjustment are improved.

[0044] 4. Hyperparameter Adaptive Adjustment and Domain Adaptation

[0045] When the amount of data is small, automatically search for the optimal hyperparameters within the predefined hyperparameter search space through Bayesian optimization, reducing the manual hyperparameter tuning work and improving the adaptability of the model in the few-shot environment;

[0046] When the amount of data is sufficient, through incremental training and domain adaptation modules, achieve the smooth migration of the pre-trained model to new domain data, effectively solve the challenges brought by data distribution changes, and monitor the effect in real time on the validation set. Automatically trigger hyperparameter fine-tuning or retraining when necessary to maintain the optimal performance of the model.

[0047] 5. System Integration and Efficient Implementation

[0048] Each module of the training device (acquisition, preprocessing, feature extraction, calculation, parameter adjustment, domain adaptation, and feedback modules) is interconnected through a high-speed data bus, realizing collaborative work on the hardware platform, improving the overall operation efficiency and data processing speed;

[0049] The modular design makes the system structure clear, easy to expand and maintain, and also convenient for integration and application in the intelligent decision-making system, thus providing accurate and efficient intelligent support in the actual decision-making process. Detailed Implementation Modes

[0050] Next, clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0051] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the present invention.

[0052] Embodiment 1

[0053] This embodiment is implemented based on the following hardware and software platforms:

[0054] Hardware platform:

[0055] Server: An NVIDIA Tesla V100 GPU server is adopted

[0056] Data bus: Based on the PCIe 3.0 high-speed data bus to achieve fast data transmission between modules. Software environment:

[0057] Operating system: Linux Ubuntu 18.04

[0058] Programming language: Python 3.7

[0059] Deep learning framework: PyTorch 1.8.0

[0060] Other tools: Use the FLANN library to implement approximate nearest neighbor search

[0061] Step description and specific parameters

[0062] Step 1: Obtain the pre-trained model and the first training data

[0063] Pre-trained model:

[0064] Select the ResNet-50 model, pre-train it based on the ImageNet dataset, and load its weights.

[0065] First training data:

[0066] The dataset includes 10 categories, with 100 samples in each category. Among them, 30 samples are selected as support data and 70 samples are selected as query data;

[0067] At the same time, provide the category labels for each sample.

[0068] Step 2: Data preprocessing

[0069] Normalization:

[0070] Common normalization processing is performed on the image data: the mean is set to [0.485, 0.456, 0.406], and the standard deviation is set to [0.229, 0.224, 0.225].

[0071] Batch normalization:

[0072] Insert a batch normalization layer at the front end of the network. Use a mini-batch size of 64 and set the BN layer momentum to 0.1 to ensure consistent data distributions across different batches.

[0073] Step 3: Feature extraction, fusion, and deep hashing processing

[0074] Feature extraction:

[0075] Input the preprocessed data into the pre-trained ResNet-50 to extract the 2048-dimensional features output by the last convolutional block;

[0076] Meanwhile, extract the features of the intermediate layer (such as the conv4 block), with the dimension set to 2048.

[0077] Multi-view feature fusion:

[0078] Concatenate the above two-way features to form a 4096-dimensional fused feature vector.

[0079] Improved deep hashing processing:

[0080] For the fused features, first pass through a batch normalization layer, then map to a 128-dimensional vector through a fully connected layer, and then use the Sigmoid activation function to obtain an output between 0 and 1;

[0081] Utilize the contrastive learning mechanism to further optimize the compactness and robustness of the 128-dimensional features by setting distance constraints for positive samples (same category) and negative samples (different categories).

[0082] Step 4: Calculate the class center features

[0083] For each category, calculate the class center using the fused features of the support data (or the vector before mapping to the 128-dimensional continuous features):

[0084] Class center = (1 / N) × ∑(support data features), where N is the number of support data for this category (N = 30 in this example);

[0085] For each dimension, take the average value of the numerical values of all support samples in this dimension.

[0086] Step 5: Calculate the approximate nearest neighbor distance

[0087] Use the FLANN library to implement the ANN algorithm and set the approximate search parameter epsilon = 0.5;

[0088] Adopt the mini-batch method, process 128 query samples in each batch, and calculate the Euclidean distance between each query sample and the center of each category.

[0089] Step 6: Dynamic parameter adjustment and model update

[0090] Layer to be adjusted: Select the last two fully connected layers of the network for dynamic parameter adjustment.

[0091] Optimizer: Use the Adam optimizer, and set the initial learning rate to 1×10 -4 ;

[0092] Regularization: Apply L2 regularization to the layer to be adjusted, and set the regularization coefficient λ to 0.001;

[0093] Gradient smoothing: Use the exponential moving average method to smooth the gradient, and set the decay rate to 0.9;

[0094] Weight update buffer: Set to save the average value of parameter updates in the past 5 training cycles as the reference for this cycle's update to reduce the fluctuation of single-step updates.

[0095] Step 7: Hyperparameter adaptive adjustment and domain adaptation

[0096] Situation A: Small amount of data (total number of samples < 1000)

[0097] Use Bayesian optimization to search for the optimal hyperparameters within the predefined hyperparameter search space:

[0098] Learning rate: Search range [1×10 -5 , 1×10 -3

[0099] Batch size: Candidate values {16, 32, 64, 128}

[0100] Number of training cycles: Search range [10, 50]

[0101] Set the number of Bayesian optimization iterations to 20 times, and adjust the parameters of the corresponding layer according to the optimization results.

[0102] Situation B: Large amount of data (total number of samples ≥ 1000)

[0103] Adopt incremental training and introduce a domain adaptation module.

[0104] During the training process, monitor the metrics on the validation set (accounting for 20% of the total data) every time a cycle is completed;

[0105] When the accuracy improvement of the validation set is less than 0.1% for 5 consecutive cycles, trigger the domain adaptation module,​

[0106] Use the MMD (Maximum Mean Discrepancy) loss as the domain adaptation loss, and set its weight factor to 0.5;

[0107] If the validation set loss does not decrease by 1% for 10 consecutive epochs, automatically trigger hyperparameter fine-tuning and retraining.

[0108] Step 8: Feedback and real-time monitoring mechanism

[0109] After each epoch, use the feedback module to monitor the performance of the validation set (metrics such as accuracy, loss value, etc.);

[0110] If the monitored metrics fluctuate or continuously decline, automatically trigger hyperparameter fine-tuning (such as adjusting the learning rate, batch size) and model retraining;

[0111] Record all training process logs for subsequent debugging and model optimization.

[0112] Modular system implementation

[0113] The following modules are interconnected through a high-speed data bus on the hardware platform to achieve collaborative processing:

[0114] Acquisition module:

[0115] Implement data reading and pre-trained model loading, and use SSD high-speed storage for data storage.

[0116] Preprocessing module:

[0117] Implement image normalization and batch normalization functions, and use multi-threaded data preprocessing technology.

[0118] Feature extraction module:

[0119] Use the pre-trained ResNet-50 and intermediate layer extractor to achieve feature extraction, and perform multi-view feature fusion and deep hashing processing.

[0120] Calculation module:

[0121] Responsible for calculating the class center features composed of support data, and using the FLANN and mini-batch algorithms to calculate the distance between query data and the class center.

[0122] Parameter adjustment module:

[0123] According to the distance information and hyperparameter search results, dynamically adjust the parameters of the specified layers in the model, and achieve smooth update with a weight update buffer.

[0124] Domain adaptation module:

[0125] When the amount of data is sufficient, the adaptive loss function is used to achieve smooth migration of the pre-trained model to new domain data.

[0126] Feedback module:

[0127] Monitor validation set metrics in real time and automatically trigger hyperparameter fine-tuning or retraining to ensure that the model remains in the optimal state.

[0128] Summary of training process

[0129] The first training data (support data and query data) and category information are input into the system.

[0130] After data preprocessing, it is sent to the pre-trained model for feature extraction, and feature fusion and deep hashing are performed to obtain 128-dimensional compact features.

[0131] The class center feature of each category is calculated based on the supporting data, and the ANN algorithm is used to calculate the distance between the query sample and the class center.

[0132] According to the distance information, the parameters of some layers of the neural network are dynamically adjusted, and regularization, gradient smoothing and buffer smoothing update strategies are adopted.

[0133] Depending on the amount of data, select Bayesian optimization or incremental training and domain adaptation modules to automatically adjust hyperparameters.

[0134] During the training process, the validation set performance is monitored in real time, and fine-tuning and retraining are automatically triggered based on the feedback until the predetermined training cycle (for example, 50 cycles) is reached or the early stopping condition is met.

[0135] Through the above steps, this embodiment can achieve:

[0136] Efficient and stable feature extraction ensures consistent data distribution across batches;

[0137] Multi-view fusion and deep hashing are used to obtain more compact and robust feature representation;

[0138] Dynamic parameter adjustment and hyperparameter adaptive strategy effectively improve the convergence speed and accuracy of model training;

[0139] Modular system design, through high-speed data bus collaborative processing, achieves significant improvement in overall operating efficiency;

[0140] Adaptive training strategies and domain adaptation functions are provided for both small sample and large sample environments to meet the needs of different application scenarios.

[0141] This embodiment details the specific implementation process of the neural network model training method, including steps such as data preprocessing, feature extraction, multi-view feature fusion, deep hashing processing, class center calculation, ANN distance calculation, dynamic parameter adjustment, hyperparameter adaptation, and domain adaptation, and gives clear parameter indicators. The above embodiment enables those skilled in the art to implement an efficient, stable, and intelligent neural network model training platform according to the teaching content of this specification, thereby achieving the expected technical effects.

[0142] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to embrace all changes that fall within the meaning and scope of the equivalent elements of the claims within the present invention.

[0143] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for training a neural network model, characterized in that: The following steps are involved: Acquire a neural network model pre-trained based on second training data, first training data, and category information of the first training data, wherein the first training data includes support data and query data, and the support data is all or part of the data of each category in the first training data, and the query data is all or part of the data of each category in the first training data; Preprocessing the first training data, wherein the preprocessing step includes normalization and batch normalization processing to ensure that data distribution of different batches is consistent; Using the neural network model to extract features from the first training data, and applying multi-view feature fusion and improved deep hashing to the extracted features to obtain a compact and robust feature representation; According to the characteristics of the supporting data, the class center feature of each category is calculated, wherein each dimension of the class center feature of each category is the average value of the corresponding feature dimension of the supporting data in the category; Using an efficient approximate nearest neighbor ANN algorithm and a mini-batch method, the feature distance between the query data feature and the center feature of each category is calculated; According to the feature distance between the class center feature of each category and the query data feature, and the average distance between the features of the first training data in each category, dynamically adjust the parameters of some layers in the neural network model to obtain an adjusted neural network model, wherein regularization, gradient smoothing and adaptive optimizer are introduced in the parameter adjustment process to ensure update stability; When the amount of the first training data is less than a preset value, automatically searching for optimal hyperparameters in a predefined hyperparameter search space through a Bayesian optimization scheme, and adjusting the parameters of the partial layers on this basis; When the amount of the first training data is greater than or equal to a preset value, according to the preset hyperparameters corresponding to the neural network model and in combination with the distance information between the query data and the class center feature, the pre-trained model is smoothly migrated using incremental training and domain adaptation modules to adapt to the distribution of new domain data; Provide feedback on training results, monitor the performance of the model on the validation set in real time, and automatically trigger hyperparameter fine-tuning and model retraining when necessary based on the monitoring results.

2. The method according to claim 1, characterized in that The deep hashing process includes first passing through a batch normalization layer after extracting features from a neural network model, and then using a contrastive learning mechanism to constrain positive and negative samples to obtain a more robust and compact feature representation.

3. The method according to claim 1, characterized in that In the dynamic parameter adjustment step, the parameter update adopts a smooth update mechanism with a weight update buffer, which uses the mean of parameter updates of several past training cycles as a reference for parameter adjustment in this cycle to reduce the volatility of single-step updates.

4. The method according to claim 1, characterized in that: The hyperparameter adaptive adjustment scheme adopts Bayesian optimization when the data volume is small, and sets a trigger mechanism when the data volume is large. When the validation set indicators have no significant improvement in several consecutive iterations, fine-tuning is triggered to further optimize the learning rate, batch size and network structure parameters.

5. An intelligent decision-making system for the method according to any one of claims 1 to 4, characterized in that: The intelligent decision-making system comprises: An acquisition module, used to acquire a neural network model, first training data and category information thereof; A preprocessing module, used for normalizing and batch normalizing the first training data; A feature extraction module, used to extract features of the first training data using the neural network model, and perform multi-view feature fusion and deep hashing processing; A calculation module is used to calculate the distance between the class center features of the support data and the query data features using ANN and mini-batch method; A parameter adjustment module, which is used to dynamically adjust the parameters of some layers in the neural network model according to the feature distance information and the hyperparameters preset or obtained by Bayesian optimization, and introduce regularization and smoothing update mechanisms; The domain adaptation module is used to achieve smooth migration of pre-trained models to new data domains when the amount of data is sufficient; The feedback module is used to monitor the training effect in real time and trigger hyperparameter fine-tuning or retraining based on the feedback.

6. The intelligent decision-making system according to claim 5, characterized in that: The acquisition module, preprocessing module, feature extraction module, calculation module, parameter adjustment module, domain adaptation module and feedback module are all interconnected through a high-speed data bus and are collaboratively implemented on a hardware platform.