An autonomous learning and training method for AI digital asset content

Abstract

The present invention relates to the technical field of artificial intelligence model training, and discloses an autonomous learning and training method for AI digital asset content. The method first obtains original data and preprocesses it to obtain a data feature matrix, generates an initial feature representation through an adaptive multi-layer perceptron network, then optimizes it with deep reinforcement learning, generates synthetic data through a generative adversarial network, and adjusts the network parameters with a contrastive learning algorithm to obtain the final digital asset content for training the AI model. This method effectively solves the existing problems in the processing of AI digital asset content, can improve data quality and utilization efficiency, expand the data volume, enhance the generalization ability of the model, optimize the model training process, improve the training accuracy, meet the requirements of real-time and security, reduce labor costs, improve training autonomy, and provide strong support for the development of AI technology.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence model training, and particularly to an autonomous learning and training method for AI digital asset content. Background Art

[0002] At present, with the booming development of artificial intelligence (AI), the training of AI models highly depends on a large amount of high-quality data. As a key data source, the processing and utilization of digital asset content are crucial for the development of AI. However, there are many problems in the current processing of AI digital asset content, which seriously restrict the further breakthrough and application expansion of AI technology.

[0003] The uneven data quality is a prominent problem. In reality, the sources of digital asset content are extensive and the formats are diverse, such as pictures, texts, audios, and videos. Taking image digital assets as an example, they may be taken under different resolutions, different shooting devices, and different lighting conditions, resulting in huge differences in image data in terms of clarity, color accuracy, and noise level; text digital assets may have problems such as spelling mistakes, grammar mistakes, semantic ambiguity, and inconsistent encoding formats; audio digital assets may be affected by recording environment noise interference, inconsistent sampling rates, etc. These quality problems make it difficult for the original data to be directly used for efficient AI model training. If low-quality data is directly used, it will lead to poor model training effects, such as low accuracy and poor generalization ability.

[0004] The insufficient data volume is also a key factor restricting the performance improvement of AI models. In many specific fields, such as medical image analysis and rare species identification, high-quality digital asset data is extremely scarce. The acquisition of medical image data not only requires professional equipment and complex processes, but also involves many restrictions such as patient privacy, resulting in slow data accumulation; for image or audio data of rare species, due to the small number of the species itself and special living environments, it is extremely difficult to collect. The limited data volume makes it difficult for AI models to learn comprehensive and accurate feature patterns. When facing new data, the models lack sufficient generalization ability and cannot accurately perform tasks such as classification, prediction, or generation.

[0005] Traditional AI model training methods have dual bottlenecks in efficiency and accuracy when dealing with digital asset content. On the one hand, traditional data preprocessing methods are often based on fixed rules or simple statistical means, and it is difficult to perform adaptive processing according to the complex characteristics of the data. For example, when normalizing image data, the commonly used fixed-range normalization method may not be able to fully retain the key detail information in the image, resulting in the loss of important features in subsequent model training; on the other hand, during the model training process, traditional supervised learning and unsupervised learning methods rely on a large amount of manual annotation or preset model structures, which not only consume a large amount of human and time costs, but also tend to fall into local optimal solutions when dealing with complex and diverse digital asset content, and cannot fully explore the potential value in the data. With the continuous expansion and deepening of AI application scenarios, the requirements for the real-time performance and security of digital asset content processing are also getting higher and higher. In scenarios such as real-time video surveillance and real-time early warning of financial transaction risks, a large amount of digital asset data needs to be quickly processed and analyzed, and traditional processing methods are difficult to meet this real-time requirement; at the same time, digital asset content contains a large amount of sensitive information, such as personal identity information, business secrets, etc. How to ensure the security and privacy of data during AI training and prevent data leakage and abuse has become an important problem to be solved urgently. Summary of the Invention

[0006] The purpose of the present invention is to provide an autonomous learning training method for AI digital asset content to solve the problems proposed in the above background technology.

[0007] To achieve the above purpose, the present invention provides the following technical solution: An autonomous learning training method for AI digital asset content, the method includes:

[0008] Obtain the original data of the digital asset content, and preprocess the original data to obtain a preprocessed data feature matrix;

[0009] Based on the preprocessed data feature matrix, generate an initial feature representation based on an adaptive multi-layer perceptron network;

[0010] Use the initial feature representation to optimize the features through a deep reinforcement learning framework to obtain an optimized feature representation;

[0011] Based on the optimized feature representation, generate synthetic data of the digital asset content based on a generative adversarial network;

[0012] Use the synthetic data to compare with the original data, and adjust the parameters of the generative adversarial network through a contrastive learning algorithm;

[0013] Generate the final digital asset content according to the adjusted generative adversarial network;

[0014] Use the final digital asset content for model training to obtain a trained AI model.

[0015] Preferably, the generating an initial feature representation based on an adaptive multi-layer perceptron network according to the preprocessed data feature matrix includes:

[0016] Perform normalization processing on the preprocessed data feature matrix to obtain a normalized data feature matrix;

[0017] Use the normalized data feature matrix to perform a non-linear transformation through the hidden layer of the adaptive multi-layer perceptron network, and the calculation formula is:

[0018]

[0019] where, is the output of the hidden layer, is the activation function, is the weight matrix of the hidden layer, is the bias vector of the hidden layer, is the normalized data feature matrix;

[0020] Perform a linear transformation on the output of the hidden layer through the output layer to obtain an initial feature representation, and the calculation formula is:

[0021]

[0022] where, is the initial feature representation, is the weight matrix of the output layer, is the bias vector of the output layer.

[0023] Preferably, the optimizing the features through a deep reinforcement learning framework using the initial feature representation to obtain an optimized feature representation includes:

[0024] Define the state space of the deep reinforcement learning framework as the initial feature representation, and the action space as the set of feature adjustment operations;

[0025] Construct a reward function, which is related to the diversity and effectiveness of the features;

[0026] Use the proximal policy optimization algorithm to perform iterative optimization in the deep reinforcement learning framework, update the parameters of the policy network and the value network, and obtain an optimized policy network;

[0027] Use the optimized policy network to operate on the initial feature representation to obtain an optimized feature representation.

[0028] Preferably, the synthetic data for generating digital asset content based on the optimized feature representation includes:

[0029] Input the optimized feature representation into the generator of the generative adversarial network. The generator consists of multiple transposed convolutional layers, and the calculation formula is:

[0030]

[0031] Where, is the generator, is the optimized feature representation, is the th transposed convolutional layer;

[0032] After the generator generates synthetic data, the discriminator discriminates between the synthetic data and the real data. The discriminator consists of multiple convolutional layers, and the calculation formula is:

[0033]

[0034] Where, is the discriminator, is the input data, is the th convolutional layer;

[0035] Through adversarial training, continuously adjust the parameters of the generator and the discriminator to make the generator generate more realistic synthetic data.

[0036] Preferably, the parameters of the generative adversarial network are adjusted by the contrastive learning algorithm using the synthetic data and the original data, including:

[0037] Calculate the cosine similarity between the synthetic data and the original data. The calculation formula is:

[0038]

[0039] Where, sim is the cosine similarity, and are the values of the synthetic data and the original data in the th dimension respectively, is the data dimension; Construct a contrastive loss function based on the cosine similarity. The calculation formula is:

[0040]

[0041] Where, is the contrastive loss function, is the temperature parameter, is the number of data samples, is the cosine similarity of other samples;

[0042] Adjust the parameters of the generator and discriminator of the generative adversarial network by using a contrastive loss function.

[0043] Preferably, generating the final digital asset content according to the adjusted generative adversarial network includes:

[0044] Input the optimized feature representation into the generator of the adjusted generative adversarial network again;

[0045] The generator outputs the final digital asset content, and the calculation formula is:

[0046]

[0047] Wherein, is the final digital asset content, is the adjusted generator, is the optimized feature representation.

[0048] Preferably, using the final digital asset content for model training to obtain a trained AI model includes:

[0049] Select a convolutional neural network as the architecture of the target AI model;

[0050] Use the final digital asset content as training data and input it into the target AI model for training; adopt the stochastic gradient descent algorithm to update the parameters of the target AI model, and the calculation formula is:

[0051]

[0052] Wherein, is the updated parameter, is the current parameter, is the learning rate, is the gradient of the loss function with respect to the current parameter; after multiple rounds of training, a trained AI model is obtained.

[0053] Preferably, it further includes the step of performing quality evaluation on the final digital asset content after generating it: calculating the information entropy of the final digital asset content to evaluate the uncertainty and richness of the data; calculating the structural similarity index between the final digital asset content and the original data to evaluate the structural similarity degree between the two; judging whether the quality of the final digital asset content meets the requirements according to the information entropy and the structural similarity index.

[0054] Preferably, if the quality of the final digital asset content does not meet the requirements, return to the step of preprocessing the original data and perform subsequent operations again.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] Through the preprocessing of the original data of digital asset content, the present invention converts it into a unified and standardized data feature matrix, effectively solving the problem of uneven data quality. In the processing of image digital assets, the resolution and color mode are unified, and noise interference is removed, enabling subsequent model training to focus on key features and greatly improving the usability of the data. At the same time, an initial feature representation is generated based on an adaptive multi-layer perceptron network, which can adaptively extract key features according to the complex characteristics of the data. Compared with traditional fixed-rule feature extraction methods, more valid information is retained, further improving the data quality and enhancing the model's understanding and utilization efficiency of the data.

[0057] Feature optimization is carried out using a deep reinforcement learning framework, and synthetic data of digital asset content is generated in combination with a generative adversarial network, successfully expanding the data volume. In data-scarce fields such as medical image analysis and rare species research, the diversity of training samples is increased through synthetic data. The generator and discriminator in the generative adversarial network are trained against each other, making the generated synthetic data highly similar to the real data in terms of features and distribution. This not only enriches the materials for model training but also enables the model to learn a wider range of feature patterns, significantly enhancing the generalization ability of the model and improving the model's performance on new data. The contrastive learning algorithm used in the invention adjusts the parameters of the generative adversarial network, which can optimize the generation process according to the similarity between the synthetic data and the original data, making the generated data closer to the real data. During model training, the final digital asset content is used as training data, and algorithms such as stochastic gradient descent are combined to update the parameters of the target AI model. This approach avoids the problem that traditional training methods are prone to falling into local optimal solutions. Taking a convolutional neural network as an example, in an image classification task, the accuracy of the model trained by the method of the present invention is significantly improved compared with traditional methods, and it can more accurately identify the object categories in the image, reducing the situation of misclassification.

[0058] In terms of real-time performance, the autonomous learning and training method of the present invention has been optimized, resulting in a faster processing speed. In real-time video surveillance scenarios, it can quickly preprocess video data, extract features, and train models, promptly detect abnormal behaviors and issue warnings, meeting the requirements of real-time processing in practical applications. In terms of security, the data processing and protection mechanisms throughout the training process are more perfect. When processing digital asset content containing sensitive information, it avoids the risks brought by data leakage and abuse, ensuring the security and privacy of the data, and providing strong support for the application of AI in fields with extremely high requirements for data security such as finance and healthcare. Different from traditional training methods that rely on a large number of manual annotations and preset model structures, the autonomous learning and training method of the present invention can automatically learn features and patterns from data. In the processing of text digital assets, there is no need to manually annotate information such as text sentiment tendency and semantic categories one by one, and the model can autonomously learn these features, reducing a large amount of human input. At the same time, the entire training process has higher autonomy, can adaptively adjust the training strategy according to data characteristics and training effects, improves the training efficiency and quality, reduces the technical dependence on professionals, and makes AI model training more convenient and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 FIG. is a working principle diagram of the autonomous learning and training method described in the present invention;

[0060] Figure 2 FIG. is an implementation step diagram for optimizing features through a deep reinforcement learning framework using the initial feature representation to obtain an optimized feature representation;

[0061] Figure 3 FIG. is an implementation step diagram for generating synthetic data based on a generative adversarial network;

[0062] Figure 4 FIG. is an implementation step diagram for adjusting the parameters of a generative adversarial network through a contrastive learning algorithm. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings 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.

[0064] Please refer to Figures 1-4 , the present invention provides a technical solution: an autonomous learning and training method for AI digital asset content, the method comprising:

[0065] Obtain the original data of the digital asset content, and preprocess the original data to obtain a preprocessed data feature matrix. For example, in an image recognition project, the original data is a large number of pictures. Perform preprocessing operations such as denoising and normalizing the size of these pictures, and convert each picture into a feature matrix with a fixed dimension for subsequent processing.

[0066] Based on the preprocessed data feature matrix, generate an initial feature representation based on an adaptive multi-layer perceptron network. Use the network to perform non-linear and linear transformations on the preprocessed data to extract the initial features of the data.

[0067] Utilize the initial feature representation to optimize the features through a deep reinforcement learning framework to obtain an optimized feature representation. Define a suitable state space, action space, and reward function, and use the proximal policy optimization algorithm to optimize the features.

[0068] Based on the optimized feature representation, generate synthetic data of the digital asset content based on a generative adversarial network. The generator and discriminator are trained against each other to generate more realistic synthetic data.

[0069] Compare the synthetic data with the original data, and adjust the parameters of the generative adversarial network through a contrastive learning algorithm. Calculate the cosine similarity between the two and construct a contrastive loss function to optimize the generative adversarial network.

[0070] Based on the adjusted generative adversarial network, generate the final digital asset content. Input the optimized feature representation into the generator of the adjusted generative adversarial network again to obtain the final digital asset content.

[0071] Use the final digital asset content for model training to obtain a trained AI model. Select a suitable neural network architecture, use the final digital asset content for training, and update the model parameters using the stochastic gradient descent algorithm.

[0072] The present invention will be further described below in conjunction with Embodiments 1 to 5:

[0073] Embodiment 1:

[0074] This embodiment details how to process the preprocessed data based on an adaptive multi-layer perceptron network to generate an initial feature representation, which is the basis for subsequent feature optimization and data generation, ensuring the extraction of valuable initial feature information from the original data.

[0075] In practical applications, taking a speech recognition project as an example, assume that the original data of the digital asset content obtained is a large number of speech audio files. First, preprocess these audio files, perform operations such as sampling and filtering on the audio signals, and convert them into a preprocessed data feature matrix.

[0076] The preprocessed data feature matrix is normalized to obtain the normalized data feature matrix. Normalization is to make the data in the same scale range for the convenience of model learning. For example, the maximum-minimum normalization method is adopted to map the data to the interval [0,1]. Suppose a certain eigenvalue in the original data feature matrix is , the maximum value of this feature is max, and the minimum value is min, then the normalized eigenvalue is calculated by the formula:

[0077]

[0078] Using the normalized data feature matrix, a non-linear transformation is performed through the hidden layer of the adaptive multi-layer perceptron network. The adaptive multi-layer perceptron network has the ability to learn and adjust its own parameters to adapt to the data features. Taking an adaptive multi-layer perceptron network with one hidden layer as an example, the output of the hidden layer is calculated by the formula:

[0079]

[0080] Among them, suppose the activation function selects the ReLU function, that is

[0081]

[0082] The weight matrix of the hidden layer and the bias vector of the hidden layer are continuously adjusted and optimized during the training process. Suppose the normalized data feature matrix is a matrix of size , represents the number of samples, represents the feature dimension. The size of the weight matrix of the hidden layer is , is the number of neurons in the hidden layer, and the size of the bias vector of the hidden layer is . After calculation, the output of the hidden layer is obtained, and its size is .

[0083] The output of the hidden layer is linearly transformed through the output layer to obtain the initial feature representation. The calculation formula is:

[0084]

[0085] The size of the weight matrix of the output layer is , is the dimension of the initial feature representation of the output, and the output layer bias vector has a size of . Through this linear transformation, the output of the hidden layer is converted into the final initial feature representation , whose size is . In a speech recognition project, the obtained initial feature representation can contain multi-faceted feature information such as the frequency, duration, and intonation of speech, providing an important data basis for subsequent feature optimization and model training.

[0086] Example 2:

[0087] This embodiment details how to optimize the initial feature representation using a deep reinforcement learning framework. By defining a suitable state space, action space, and reward function, and adopting the proximal policy optimization algorithm, the optimized feature representation becomes more diverse and effective, improving the performance of the model.

[0088] In an actual scenario of image classification, feature optimization is carried out using a deep reinforcement learning framework. First, the state space of the deep reinforcement learning framework is defined as the initial feature representation. For example, the initial feature representation obtained through Example 1 is a matrix of size , where each row represents the initial features of an image sample, and each column represents different feature dimensions. This matrix is used as the state of the deep reinforcement learning framework, and the agent makes decisions in this state space.

[0089] The action space is a set of operations for adjusting the features. For image features, actions can include operations such as scaling, translation, and adding noise to certain feature dimensions. For example, define an action as multiplying the values of certain specific feature dimensions by a random number between to achieve the scaling operation of the features; or adding a random number between to the values of certain feature dimensions to achieve the translation operation.

[0090] Construct a reward function, which is related to the diversity and effectiveness of the features. In an image classification task, the diversity of the features can be measured by calculating the correlation between the features. The lower the correlation, the higher the diversity; the effectiveness of the features can be measured by the accuracy of image classification using the current features. Assume that the feature representation in the current state is , and define the reward function as:

[0091]

[0092] where represents the correlation coefficient between the features, represents using the feature The accuracy of image classification and are weight coefficients used to balance the importance of diversity and effectiveness in the reward function. For example, set .

[0093] The Proximal Policy Optimization (PPO) algorithm is used for iterative optimization in the deep reinforcement learning framework to update the parameters of the policy network and the value network, obtaining an optimized policy network. The PPO algorithm makes the training more stable and efficient by restricting the magnitude of policy updates. During training, the agent selects actions from the policy network based on the current state, and after executing the actions, new states and rewards are obtained. Based on this data, the parameters of the policy network and the value network are updated. Assuming the parameters of the policy network are , and the parameters of the value network are , through multiple iterations, continuously adjust and so that the policy network can select better actions.

[0094] The optimized policy network is used to operate on the initial feature representation to obtain an optimized feature representation. In the image classification scenario, the optimized policy network selects a series of actions based on the initial feature representation, such as scaling and translating certain feature dimensions. After these operations, an optimized feature representation is obtained, which has been improved in terms of both diversity and effectiveness and is more conducive to subsequent image classification tasks.

[0095] Example 3:

[0096] This example specifically describes how to generate synthetic data of digital asset content based on a generative adversarial network. Through the adversarial training of the generator and the discriminator, the quality of the synthetic data is continuously improved, providing rich data for subsequent contrast learning and model training.

[0097] Taking a project of generating handwritten digit images as an example, synthetic data of digital asset content is generated based on a generative adversarial network. The optimized feature representation is input into the generator of the generative adversarial network. The generator consists of multiple transposed convolutional layers, and the calculation formula is:

[0098]

[0099] Assume the generator has 3 transposed convolutional layers. The first transposed convolutional layer takes the optimized feature representation as its input, and its size is (assuming the optimized feature representation is compressed to such a dimension). After passing through the first transposed convolutional layer, the output size becomes , and the convolutional kernel size is , with a step size of 1; the second transposed convolutional layer takes the output of the first layer as its input. After passing through this layer, the output size becomes , with a convolutional kernel size of , and a step size of 2; the third transposed convolutional layer takes the output of the second layer as its input. After passing through this layer, the output size becomes , and what is obtained is the generated handwritten digit image (grayscale image).

[0100] After the generator generates synthetic data, the discriminator discriminates between the synthetic data and the real data. The discriminator consists of multiple convolutional layers, and the calculation formula is:

[0101]

[0102] Assume that the discriminator has 4 convolutional layers. The first convolutional layer takes as its input the image data of size (which can be either synthetic data or real data), with a convolutional kernel size of , and a step size of 2. The output size becomes The second convolutional layer takes the output of the first layer as its input, with a convolutional kernel size of , and a step size of 2. The output size becomes The third convolutional layer takes the output of the second layer as its input, with a convolutional kernel size of , and a step size of 2. The output size becomes ; The fourth convolutional layer takes the output of the third layer as its input. After passing through this layer, it outputs a scalar value used to determine whether the input image is real data or synthetic data.

[0103] Through adversarial training, the parameters of the generator and the discriminator are continuously adjusted to make the generator generate more realistic synthetic data. During the training process, the generator attempts to generate synthetic data that can deceive the discriminator, while the discriminator attempts to accurately distinguish between real data and synthetic data. Assume that the loss function of the generator is , and the loss function of the discriminator is , and the adversarial training algorithm adopted is the min-max game algorithm. The goal of the generator is to minimize , that is

[0104]

[0105] where is the real data, is the random noise (here replaced by the optimized feature representation), is the distribution of the real data, is the distribution of noise. The goal of the discriminator is to maximize ,Right now

[0106]

[0107] Through continuous iterative training, the handwritten digit images generated by the generator become more and more realistic and can better simulate realistic handwritten digit data.

[0108] Embodiment 4:

[0109] This embodiment aims to construct a loss function based on the comparison of synthetic data and original data through a contrastive learning algorithm, and optimize and adjust the parameters of the generative adversarial network to improve the quality of the generated data, make it more similar to the original data at the feature level, and enhance the effectiveness and authenticity of the data generated by the generative adversarial network.

[0110] Taking the facial image generation task as an example, assuming that we have generated a batch of synthetic facial image data through a generative adversarial network, we now need to use a contrastive learning algorithm to adjust the parameters of the generative adversarial network.

[0111] Calculate the cosine similarity between the synthesized data and the original data. Convert both the original facial image and the synthesized facial image into feature vectors, for example, using a pre-trained convolutional neural network (such as VGG16) to extract the features of the image. Assume that the original image feature vector is , the synthetic image feature vector is ,here is the dimension of the feature vector, assuming According to the cosine similarity formula

[0112]

[0113] For example, for a pair of original images and synthesized images, the calculated cosine similarity is is 0.6.

[0114] Then, we construct a contrast loss function based on cosine similarity. Setting temperature parameters , in a batch containing 100 image samples (including original images and synthesized images), Take 20 (i.e. select 20 sample pairs for comparison calculation). The calculation formula of the comparison loss function is:

[0115]

[0116] in is the cosine similarity for other sample pairs. In actual calculation, traverse these 20 sample pairs, calculate their cosine similarities respectively, and then substitute them into the formula to calculate the contrast loss function value. Suppose the calculated contrast loss function value is 0.8. Finally, use the contrast loss function to adjust the parameters of the generator and discriminator of the generative adversarial network. Through the backpropagation algorithm, transfer the gradient of the contrast loss function to the generator and discriminator. Suppose the parameters of the generator are , and the parameters of the discriminator are . Use the Adam optimizer to update the parameters. The learning rate is set to 0.0001. For the generator, the parameter update formula is

[0117]

[0118] where is the learning rate, is the number of training epochs; for the discriminator, the parameter update formula is

[0119]

[0120] After multiple rounds of training, continuously adjust the parameters of the generator and discriminator so that the generated facial images are closer to the original facial images in terms of details, facial feature distribution, etc.

[0121] Example 5:

[0122] This example mainly shows how to use the finally generated digital asset content to train an AI model to obtain training results that meet actual requirements. At the same time, evaluate the finally generated digital asset content through specific quality evaluation indicators, and decide whether to re - perform the data processing and model training process based on the evaluation results to ensure the reliability and practicability of the entire technical solution.

[0123] In the object detection project in the autonomous driving scenario, use the finally generated digital asset content for model training and quality evaluation.

[0124] Select the convolutional neural network (CNN) as the architecture of the target AI model. Here, choose the classic YOLO (You Only Look Once) series network, such as YOLOv5. Use the finally generated digital asset content as training data, which are synthetic images containing targets such as vehicles, pedestrians, and traffic signs generated after a series of previous processing. During training, input these synthetic images and their corresponding annotation information (category, location, etc. of the targets) into the YOLOv5 model.

[0125] Use the stochastic gradient descent algorithm to update the parameters of the target AI model. The formula is

[0126]

[0127] where is the current parameter, is the updated parameter, is the learning rate, set to , is the gradient of the loss function with respect to the current parameter. The loss function combines the classification loss (such as cross - entropy loss) and localization loss (such as mean - square error loss) in the object detection task. During the training process, the model continuously adjusts the parameters according to the feedback of the loss function. After 500 rounds of training, the model gradually converges, and the detection accuracy for various objects continues to improve.

[0128] After generating the final digital asset content, its quality is evaluated. Calculate the information entropy of the final digital asset content to evaluate the uncertainty and richness of the data. For the synthesized autonomous driving scene images, consider the pixel values of the image as a probability distribution. According to the information entropy formula

[0129]

[0130] calculate the information entropy, where is the probability that the pixel value is . Suppose the calculated information entropy of a certain synthesized image is bits. The higher the information entropy, the more uniform the distribution of pixel values in the image and the higher the richness of the data.

[0131] Calculate the structural similarity index (SSIM) between the final digital asset content and the original data to evaluate the structural similarity between the two. SSIM measures the similarity by comparing the brightness, contrast, and structural information of the images, and its value ranges from [- 1,1]. The closer the value is to 1, the more similar the structures are. Use a dedicated SSIM calculation function to calculate the synthesized image and the original image. Suppose the obtained SSIM value is .

[0132] According to the information entropy and the structural similarity index, judge whether the quality of the final digital asset content meets the requirements. Set the threshold of the information entropy to bits, and the threshold of SSIM to . If the information entropy and of a certain synthesized image, then it is considered that the quality of this synthesized image meets the requirements; if not, for example, the information entropy of an image is 6.5 bits and the SSIM value is 0.75, then return to the step of pre - processing the original data and re - perform the subsequent operations to improve the quality of the synthesized data, thereby improving the performance of the object detection model.

[0133] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus.

[0134] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An autonomous learning training method for AI digital asset content, characterized in that: The method comprises: Acquire original data of digital asset content, and preprocess the original data to obtain a preprocessed data feature matrix; the digital asset content includes pictures, text, audio and video; According to the preprocessed data feature matrix, generating an initial feature representation based on an adaptive multi-layer perception network; Using the initial feature representation, optimizing the features through a deep reinforcement learning framework to obtain an optimized feature representation; Generate synthetic data of digital asset content based on a generative adversarial network according to the optimized feature representation; Using the synthetic data to compare with the original data, and adjusting the parameters of the generative adversarial network through a contrastive learning algorithm; Generate the final digital asset content based on the adjusted generative adversarial network; Using the final digital asset content to perform model training to obtain a trained AI model; The utilizing the initial feature representation and optimizing the feature through a deep reinforcement learning framework to obtain an optimized feature representation includes: Define the state space of the deep reinforcement learning framework as the initial feature representation, and the action space as the set of adjustment operations on the features; Construct a reward function that is related to the diversity and effectiveness of features; The proximal policy optimization algorithm is used to perform iterative optimization in the deep reinforcement learning framework, update the parameters of the policy network and the value network, and obtain the optimized policy network; Use the optimized strategy network to operate the initial feature representation to obtain the optimized feature representation; The step of comparing the synthetic data with the original data and adjusting the parameters of the generative adversarial network by a contrastive learning algorithm includes: Calculate the cosine similarity between the synthetic data and the original data, the calculation formula is: Among them, sim is the cosine similarity, and The synthetic data and the original data are The values ​​in the dimensions, is the data dimension; the contrast loss function is constructed based on the cosine similarity, and the calculation formula is: in, is the contrast loss function, is the temperature parameter, is the number of data samples, is the cosine similarity of other samples; The contrastive loss function is used to adjust the parameters of the generator and discriminator of the generative adversarial network.

2. The autonomous learning training method for AI digital asset content according to claim 1, characterized in that: Generating an initial feature representation based on an adaptive multi-layer perception network according to the preprocessed data feature matrix includes: Normalizing the preprocessed data feature matrix to obtain a normalized data feature matrix; Using the normalized data feature matrix, a nonlinear transformation is performed through the hidden layer of the adaptive multi-layer perception network. The calculation formula is: in, is the hidden layer output, is the activation function, is the hidden layer weight matrix, is the hidden layer bias vector, is the normalized data feature matrix; The hidden layer output is linearly transformed through the output layer to obtain the initial feature representation. The calculation formula is: in, is the initial feature representation, is the output layer weight matrix, is the output layer bias vector.

3. The autonomous learning training method for AI digital asset content according to claim 2, characterized in that: The step of generating synthetic data of digital asset content based on a generative adversarial network according to the optimized feature representation includes: The optimized feature representation is input into the generator of the generative adversarial network. The generator consists of multiple layers of transposed convolutional layers, and the calculation formula is: in, For the generator, is the optimized feature representation, For the Layer transposed convolutional layer; After the generator generates synthetic data, the discriminator distinguishes the synthetic data from the real data. The discriminator consists of multiple convolutional layers, and the calculation formula is: in, is the discriminator, For input data, For the Layer convolutional layer; Through adversarial training, the parameters of the generator and discriminator are continuously adjusted to enable the generator to generate more realistic synthetic data.

4. The autonomous learning training method for AI digital asset content according to claim 3, characterized in that: The final digital asset content generated according to the adjusted generative adversarial network includes: The optimized feature representation is input into the generator of the adjusted generative adversarial network again; The generator outputs the final digital asset content, and the calculation formula is: in, For the final digital asset content, is the adjusted generator, is the optimized feature representation.

5. The autonomous learning training method for AI digital asset content according to claim 4, characterized in that: The method of using the final digital asset content to perform model training to obtain a trained AI model includes: Select convolutional neural network as the architecture of the target AI model; The final digital asset content is used as training data and input into the target AI model for training. The parameters of the target AI model are updated using the stochastic gradient descent algorithm. The calculation formula is: in, is the updated parameter, is the current parameter, is the learning rate, is the gradient of the loss function with respect to the current parameters; after multiple rounds of training, a trained AI model is obtained.

6. The autonomous learning training method for AI digital asset content according to claim 1, characterized in that: It also includes the steps of evaluating the quality of the final digital asset content after it is generated: calculating the information entropy of the final digital asset content and evaluating the uncertainty and richness of the data; Calculate the structural similarity index between the final digital asset content and the original data to evaluate the structural similarity between the two; Based on the information entropy and structural similarity index, determine whether the quality of the final digital asset content meets the requirements.

7. The autonomous learning training method for AI digital asset content according to claim 6, characterized in that: If the quality of the final digital asset content does not meet the requirements, return to the step of preprocessing the original data and perform subsequent operations again.

Images