Data processing method and device based on AI and storage medium

Abstract

The invention discloses a data processing method and device based on AI and a storage medium. The method comprises the following steps: acquiring a feature matrix to be subjected to dimension reduction, calculating a first similarity between data points in the feature matrix in a first space by using a Gaussian kernel function, and randomly mapping each data point in a plurality of data points corresponding to the feature matrix in the first space to a second space, the dimension of the first space is greater than that of the second space; and calculating a second similarity between data points in the feature matrix in a second space by using T distribution, and adjusting the position of each data point in the second space based on the first similarity and the second similarity so as to reduce the dimension of the feature matrix. The technical problem of low data processing efficiency in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to an AI-based data processing method, device and storage medium. Background Art

[0002] In the field of big data and artificial intelligence, as the amount of data grows rapidly, the processing of high-dimensional data has become an important challenge. High-dimensional data usually contains a large number of features. Although these features can provide rich information in theory, in practical applications, high-dimensional data will bring about the "curse of dimensionality" problem, resulting in inaccurate distance measurement between data points, thus affecting the performance of machine learning models. In addition, processing high-dimensional data is often accompanied by huge computing overhead and low processing efficiency, which brings great difficulties to real-time data processing and large-scale data analysis.

[0003] In order to alleviate the above problems, dimensionality reduction techniques have been widely studied and applied. Dimensionality reduction methods improve the efficiency of data processing by mapping high-dimensional data to low-dimensional space and reducing the number of features. However, traditional dimensionality reduction methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA), often fail to preserve the intrinsic structure of the data well when processing complex data, resulting in large information loss. In addition, these methods have limited performance when processing nonlinear data and cannot effectively capture the nonlinear relationship between data points.

[0004] In recent years, AI-based dimensionality reduction methods have gradually emerged, which solve the problem of nonlinear dimensionality reduction by introducing kernel functions and probabilistic models. Such methods can better preserve the similarity structure between data and reduce information loss during dimensionality reduction. However, in the existing technology, the computational complexity of these methods is still relatively high, especially when processing large-scale data sets, which easily leads to low processing efficiency.

[0005] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0006] The embodiments of the present invention provide an AI-based data processing method, device, and storage medium to at least solve the technical problem of low efficiency of data processing in the prior art.

[0007] According to one aspect of an embodiment of the present invention, there is provided an AI-based data processing method, comprising: obtaining a feature matrix to be reduced in dimension, calculating a first similarity between each data point in the feature matrix in a first space using a Gaussian kernel function, and randomly mapping each data point of a plurality of data points corresponding to the feature matrix in the first space to a second space, wherein the dimension of the first space is greater than the dimension of the second space; calculating a second similarity between each data point in the feature matrix in a second space using a T distribution, and adjusting the position of each data point in the second space based on the first similarity and the second similarity to reduce the dimension of the feature matrix.

[0008] According to another aspect of an embodiment of the present invention, there is also provided an AI-based data processing device, including: an acquisition module, configured to acquire a feature matrix to be reduced in dimension; a mapping module, configured to use a Gaussian kernel function to calculate a first similarity between each data point in the feature matrix in a first space, and randomly map each data point of a plurality of data points corresponding to the feature matrix in the first space to a second space, wherein the dimension of the first space is greater than the dimension of the second space; a dimensionality reduction module, configured to use a T distribution to calculate a second similarity between each data point in the feature matrix in a second space, and adjust the position of each data point in the second space based on the first similarity and the second similarity to reduce the dimension of the feature matrix.

[0009] In an embodiment of the present invention, a feature matrix to be reduced in dimension is obtained, a Gaussian kernel function is used to calculate a first similarity between each data point in the feature matrix in a first space, and each data point in a plurality of data points corresponding to the feature matrix in the first space is randomly mapped to a second space, wherein the dimension of the first space is greater than the dimension of the second space; a second similarity between each data point in the feature matrix is ​​calculated in the second space using T distribution, and based on the first similarity and the second similarity, the position of each data point in the second space is adjusted to reduce the dimension of the feature matrix. Through the above scheme, the technical problem of low efficiency of data processing in the prior art is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0011] Figure 1 is a flow chart of an AI-based data processing method according to an embodiment of the present invention;

[0012] Figure 2 is a flowchart of another AI-based data processing method according to an embodiment of the present invention;

[0013] Figure 3 is a flow chart of a method for applying a data processing method to a virtual digital human driver according to an embodiment of the present invention;

[0014] Figure 4 is a flow chart of a method for extracting data features according to an embodiment of the present invention;

[0015] Figure 5 is a flow chart of a method for feature fusion according to an embodiment of the present invention;

[0016] Figure 6 is a schematic diagram of the structure of an AI-based data processing device according to an embodiment of the present invention;

[0017] Figure 7 A schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] According to an embodiment of the present invention, a method embodiment of an AI-based data processing method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0021] Figure 1 is an AI-based data processing method according to an embodiment of the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0022] Step S102, obtain the feature matrix to be reduced in dimension, use a Gaussian kernel function to calculate the first similarity between each data point in the feature matrix in the first space, and randomly map each data point of the multiple data points corresponding to the feature matrix in the first space to the second space, wherein the dimension of the first space is greater than the dimension of the second space.

[0023] First, obtain the feature matrix to be reduced. The feature matrix is ​​a matrix composed of feature vectors of multiple data points, where each row usually represents a data point and each column represents a feature.

[0024] Next, a Gaussian kernel function is used to calculate the first similarity between each data point in the feature matrix in the first space. For example, each data point in the feature matrix is ​​mapped onto a high-dimensional sphere as the first space, and the Euclidean distance between each data point is calculated by the Gaussian kernel function, wherein the high-dimensional sphere is a sphere with a spatial dimension greater than a preset dimension; and a first similarity matrix representing the first similarity between data points is generated according to the Euclidean distance.

[0025] Afterwards, each of the plurality of data points corresponding to the feature matrix in the first space is randomly mapped to the second space. For example, a random mapping matrix is ​​generated according to the first similarity matrix; and each of the data points is mapped from the first space to the second space based on the random mapping matrix.

[0026] Step S104, using T distribution to calculate the second similarity between each data point in the feature matrix in the second space, and based on the first similarity and the second similarity, adjusting the position of each data point in the second space to reduce the dimension of the feature matrix.

[0027] First, the second similarity between each data point in the feature matrix is ​​calculated in the second space using the T distribution. For example, for each data point randomly mapped into the second space, the Euclidean distance between each data point in the second space is calculated to generate a distance matrix; based on the distance matrix, the T distribution function is used to perform probability distribution calculation on the distance between each data point in the second space to generate a second similarity matrix representing the second similarity between the data points.

[0028] Next, based on the first similarity and the second similarity, the position of each data point in the second space is adjusted. For example. Based on the first similarity matrix and the second similarity matrix, the divergence value between each data point in the second space is calculated to generate a divergence matrix; using the divergence matrix, an objective function representing the difference between the first similarity and the second similarity is constructed; by minimizing the objective function, the position of each data point in the second space is adjusted to obtain the adjusted data point position as the feature matrix after dimensionality reduction.

[0029] In some embodiments, a divergence matrix may be generated by the following method: for each pair of corresponding data points in the first similarity matrix and the second similarity matrix, the difference between the data point pairs is calculated to obtain a difference matrix; based on the difference matrix, the difference between the data point pairs is quantified using a divergence function to generate a divergence value; based on the divergence value, the divergence matrix is ​​generated, wherein the divergence matrix is ​​used to represent the position change of each data point in the first space and the second space.

[0030] Figure 2 is an AI-based data processing method according to an embodiment of the present invention, such as Figure 2 As shown, the method comprises the following steps:

[0031] Step S202, obtaining a feature matrix to be reduced in dimension.

[0032] First, extract the feature matrix X∈R from the input data set m×n , where m is the number of data points and n is the feature dimension of each data point. Each row of the feature matrix X corresponds to a feature vector xi = [x i1 ,x i2 ,...,x in ], each column represents a certain dimension of features.

[0033] The feature matrix can be generated by feature extraction techniques such as principal component analysis (PCA), factor analysis (FA), high-dimensional features extracted by convolutional neural networks (CNN), or other feature vectors obtained from multimodal data such as images, text, audio, and video.

[0034] For example, in image recognition tasks, the feature vector xi can be obtained by flattening the feature map output by the convolution layer. Although these high-dimensional feature vectors can represent the complex information of the data, they may introduce unnecessary noise and redundant information due to their high dimensions. Therefore, these features need to be reduced in dimension to improve the efficiency and accuracy of subsequent algorithms.

[0035] Step 204: Calculate the first similarity using a multi-core Gaussian kernel function.

[0036] The multi-core Gaussian kernel function is used to calculate the similarity between data points in the feature matrix. Different from the traditional single Gaussian kernel, the embodiment of the present invention uses a multi-core Gaussian kernel function to improve the accuracy of similarity calculation.

[0037] First, map each data point xi in the feature matrix to the high-dimensional spherical space S n-1 , that is, the position of each data point in the high-dimensional spherical space is:

[0038]

[0039] Among them, ‖x i ‖ is the L2 norm of the feature vector. The mapping ensures that each data point is located on a unit sphere with a radius of 1, thereby unifying the scale of the data and preventing data of different scales from causing inaccurate similarity calculations. xi is the feature vector of the i-th data point in the original feature matrix, located in the high-dimensional space, and xi′ is the feature vector of the i-th data point mapped to the high-dimensional sphere, located on the unit sphere.

[0040] Next, the multi-kernel Gaussian kernel function is used to calculate the first similarity between each pair of data points in the feature matrix. The multi-kernel Gaussian kernel function is a weighted combination of multiple Gaussian kernel functions, each with a different bandwidth parameter σ k The formula for multi-core Gaussian kernel similarity is as follows:

[0041]

[0042] Among them, α k is the weight of the kth kernel function, satisfying σ k is the corresponding bandwidth parameter, which is used to control the scale of the kernel function. k The value of can be selected by cross-validation and other methods to ensure that each kernel function can capture the similarity between data points at different scales. P is the number of Gaussian kernel functions used, that is, the number of types of kernel functions.

[0043] In addition, considering that the distribution of data points may be uneven, the present invention introduces a local density weight γi to enhance the similarity calculation between data points in a dense area. The local density weight can be calculated by the following formula:

[0044]

[0045] The weight value reflects the density of the area where each data point is located. Data points in high-density areas will be given higher weights so that they stand out more in the similarity calculation. i is the local density weight of the i-th data point, reflecting the data density of the area where the point is located, m is the total number of data points in the data set, and σ is a parameter that controls distance scaling and is used to adjust the sensitivity of the density weight. The first similarity matrix S1∈Rm×m is generated, and its element S1(i,j) represents the similarity between data point i and data point j:

[0046]

[0047] Step S206: adaptively and randomly mapping to a low-dimensional space.

[0048] The data points are mapped from a high-dimensional first space to a low-dimensional second space. The invention enhances the flexibility and robustness of the mapping by introducing an adaptive random mapping parameter β.

[0049] First, based on the calculated first similarity matrix S1, a random mapping matrix R∈R is generated n×d , where n is the dimension of the original feature matrix and d is the target dimension after dimensionality reduction. The generation process of the mapping matrix R is as follows:

[0050]

[0051] Among them, N(0,β) means that it obeys a Gaussian distribution with a mean of 0 and a variance of β. The value of parameter β is adaptively adjusted according to the distribution of the data to ensure that the local structure of the data point can be preserved during the mapping process. The mapped data point is expressed as:

[0052] X′=X·(R·β)

[0053] Among them, X represents the original high-dimensional feature matrix with a dimension of m×n. R is the random mapping matrix with a dimension of n×d, where d <n。

[0054] In order to further optimize the mapping effect, the present application also sets an objective function based on local preservation projection, that is, to try to maintain the local similarity relationship of data points during mapping. The objective function set based on local preservation projection is as follows:

[0055]

[0056] By optimizing the objective function, a low-dimensional space representation X′∈R that preserves the local structure is generated. m×d .

[0057] Step S208: Calculate the second similarity using T distribution in the low-dimensional space.

[0058] The similarity between data points in the low-dimensional space is calculated using T distribution. Compared with the similarity calculation in the first space, T distribution can better handle the data distribution in the low-dimensional space, especially for high-density data points.

[0059] First, calculate the Euclidean distance between data points in low-dimensional space and generate a distance matrix D∈R m×m :

[0060] D(i,j)=‖x i ′-x′ j ‖ 2

[0061] Next, the distance is transformed using T distribution to calculate the second similarity matrix S2∈Rm×m. The similarity calculation formula of T distribution is:

[0062]

[0063] Where P(i|j) represents the similarity probability of data point xi′ to data point xj′ in low-dimensional space. D(i,j) represents the square of the Euclidean distance between data points xi′ and xj′ in low-dimensional space. The formula in the denominator represents the normalization factor for calculating the similarity probability for all k not equal to i, ensuring that the sum of all P(i|j) is 1. This formula ensures that data points in high-density areas are given higher similarity weights, while avoiding high similarity between data points that are far away.

[0064] Step S210, constructing an objective function and minimizing the divergence for optimization.

[0065] An objective function is constructed to quantify the difference between the first similarity and the second similarity in the dimensionality reduction process, and the position of the data points in the low-dimensional space is optimized by minimizing the objective function.

[0066] First, a divergence matrix is ​​constructed to measure the similarity difference between each pair of data points in the first space and the second space.

[0067] Next, the objective function formula is constructed based on the divergence matrix as follows:

[0068]

[0069] In order to avoid overfitting, in some other embodiments, a regularization term Ω(X′) is introduced into the objective function, and the regularization term is used to constrain the distribution of data points after dimensionality reduction. The objective function with the regularization term is expressed as:

[0070]

[0071] Wherein, Ω(X′) is a regularization term, which is used to constrain the distribution of data points in the low-dimensional space to ensure that the distribution of data points is more uniform or meets specific application requirements. The regularization term can be, for example, the L2 norm, which is used to control the amplitude of the data point position.

[0072] Finally, the feature matrix X′ after dimensionality reduction is output, which contains the new position of each data point in the low-dimensional space. This feature matrix not only retains the main similarities between data points, but also improves the similarity retention effect and the generalization ability of the model by introducing local density weights, regularization terms and multi-core Gaussian kernel functions.

[0073] The embodiments of the present application can effectively reduce the dimension of a high-dimensional feature matrix to a lower dimension while retaining the similarity structure of the data, thereby improving the efficiency and accuracy of data processing.

[0074] The present application also provides a method for applying the above data processing method to driving a virtual digital human, such as Figure 3 As shown, the method comprises the following steps:

[0075] Step S302: acquiring speech data and emotion data corresponding to the speech data, extracting speech features from the speech data, and extracting emotion features from the emotion data.

[0076] like Figure 4 As shown, step S302 includes the following:

[0077] Step S3022, obtaining voice data and emotion data.

[0078] First, the original voice data is obtained from the voice signal input by the user. The voice data can be collected through a variety of devices, such as a microphone, a recording device, or an audio interface of a mobile terminal. In order to ensure the accuracy and completeness of the data, the voice acquisition module will pre-process the collected audio signal, such as denoising, normalization, etc. At the same time, the system will also obtain the emotional data corresponding to the voice data. Emotional data can be obtained through user feedback, expression recognition, heart rate monitoring, etc., or the voice data can be preliminarily analyzed based on a pre-built emotional model to obtain the corresponding emotional label.

[0079] Step S3024, extracting speech features from the speech data.

[0080] After acquiring the speech data, the speech feature extraction module converts the original speech data into multiple feature vectors representing speech characteristics through a series of signal processing algorithms (such as MFCC, spectrogram, time-frequency analysis, etc.). Specifically, the long speech signal is divided into multiple short time frames for feature extraction. The time domain signal is converted into a frequency domain signal through short-time Fourier transform, and the frequency components are extracted from it. Finally, MFCC (Mel-frequency cepstral coefficients) are extracted.

[0081] Step S3026, extracting emotion features from the emotion data.

[0082] After obtaining the emotion data corresponding to the speech data, the emotion data will be feature extracted. The emotion feature extraction module selects different processing methods based on the type of emotion data. The emotion data can be obtained through speech analysis, and then further refined analysis can be performed based on the emotion model to extract specific emotion category features such as anger, joy, sadness, etc. Emotion data can also come from non-speech methods, such as expression recognition or biometric monitoring (such as heart rate, skin galvanic response, etc.). In this way, the corresponding emotion features can be extracted based on a specific emotion inference algorithm.

[0083] Step S304: fusing the speech feature and the emotion feature based on the factor graph to obtain a fused feature.

[0084] like Figure 5 As shown, the feature fusion method includes the following steps:

[0085] Step S3042, constructing a factor graph model.

[0086] First, we construct a factor graph model. The factor graph contains two main elements: variable nodes and factor nodes. In this solution, variable nodes represent different dimensions of speech features and emotional features, such as the frequency features, pitch features of speech, and the excitement and pleasure of emotions. Factor nodes represent the probabilistic dependency between these features, which is used to characterize the potential correlation between speech and emotional features.

[0087] Step S3044, initializing the node values ​​in the factor graph.

[0088] After the factor graph is constructed, the previously extracted speech features and emotion features are used as the initial input values ​​of the variable nodes. The dimension of each speech feature and emotion feature vector corresponds to the corresponding variable node in the factor graph. The system will assign initial probability distribution values ​​to each variable node according to a predetermined initialization algorithm. These values ​​come from a specific prior model or empirical distribution obtained through training.

[0089] Step S3046, execute the information transmission algorithm.

[0090] In the factor graph model, the information transfer algorithm is used to transfer and update the probability information between the variable nodes and the factor nodes. Through iterative calculation, the posterior probability distribution of each variable node is gradually updated. After receiving the probability information of the adjacent factor nodes, each variable node updates its own state by combining its own characteristics and feedback from other nodes. The speech feature node will adjust according to the feedback information of the emotion feature node. For example, speech features such as pitch and speech speed may be affected by emotional features such as emotional excitement and tension. Conversely, the emotion feature node will also update its own state according to the feedback of the speech feature.

[0091] Step S3048, generating fusion features.

[0092] After the information transfer process converges, the states of each node in the factor graph will reach a stable state. At this point, the final fusion feature is extracted. The fusion feature is composed of the posterior probability distribution of all variable nodes in the factor graph, reflecting the interdependence of speech and emotion features in multidimensional space. The fusion feature contains not only the independent information of speech and emotion, but also the complex relationship between them, which can better support subsequent application scenarios, such as emotion recognition, speech sentiment analysis, or virtual human expression generation.

[0093] Step S3049, performing dimensionality reduction processing on the fused features.

[0094] The fused features are often high-dimensional vectors. To ensure the efficiency and accuracy of the neural network processing, the system performs dimensionality reduction processing on these features. For example, the fused features are processed by the AI-based data processing method provided in the above embodiment. The specific processing method is as described above and will not be repeated here.

[0095] Step S306, performing animation parameter mapping processing on the fusion feature to obtain controller parameters for driving the facial expression and mouth shape of the virtual object.

[0096] Firstly, a mapping model based on deep neural network (DNN) is used to realize the mapping of fusion features to control parameters. The neural network model mainly consists of input layer, hidden layer and output layer, which is responsible for converting complex fusion features into specific parameters that can be used for animation control. The input layer receives the standardized fusion feature vector, and the dimension of the feature depends on the specific size of the fusion feature generated by the previous factor graph. The hidden layer is the key part of the neural network, which consists of multiple fully connected neurons, each of which uses a nonlinear activation function (such as ReLU or Sigmoid) for feature transformation. This part can capture the complex relationship between speech and emotional features through multi-layer nonlinear mapping, and then generate control parameters that adapt to various expressions and mouth shape changes. The depth of the hidden layer and the number of neurons in each layer are carefully designed and optimized to ensure that the model can effectively process the fusion features and provide accurate mapping. The output layer generates a set of controller parameters for driving the facial expressions and mouth shapes of the virtual object. The output control parameters include various animation parameters for controlling facial movements, such as mouth opening and closing, eyebrow movements, eye movements, etc.

[0097] To ensure that the neural network mapping model can accurately generate facial expressions and lip control parameters, the model is trained. The training data contains paired samples of speech and emotion features and corresponding animation control parameters. During the training process, the system uses supervised learning to repeatedly adjust the weight parameters of the model by calculating the error between the predicted control parameters and the actual control parameters. Loss functions (such as mean square error) are used to optimize the neural network to continuously improve and increase prediction accuracy. Through training with a large amount of data, the neural network can learn the complex nonlinear relationship between fusion features and animation parameters, thereby generating high-precision control parameters that meet the needs of actual applications.

[0098] After that, facial expression control parameters are generated. After the fusion features after dimensionality reduction are input into the neural network, a set of facial expression control parameters are generated through the network. The control parameters are directly used to drive the expression changes of the virtual person, including but not limited to the raising or frowning of eyebrows, the opening and closing of eyes, and the changes of mouth. The generated facial expression parameters are automatically controlled by the neural network to ensure that the expression of the virtual object matches the voice and emotional characteristics, providing realistic and delicate visual performance. For example, when the voice contains anger or sadness, the neural network can generate corresponding control parameters for lowering eyebrows or lowering the corners of the mouth to reflect the emotional state of the virtual person.

[0099] Finally, lip synchronization parameters are generated. In addition to generating facial expression control parameters, the neural network is also responsible for generating lip control parameters that are synchronized with the speech signal. Based on the phoneme information in the speech features, the neural network can accurately predict the opening and closing of the mouth and the movement trajectory of the lips, ensuring that the lip movement of the virtual object matches the speech pronunciation. Lip synchronization processing is trained with a large amount of speech and lip data, and the model can learn the correspondence between speech pronunciation and lip movement, thereby achieving a natural lip synchronization effect in virtual human animation.

[0100] Step S308: using the controller parameters to drive the virtual object.

[0101] All generated controller parameters are output and passed to the animation engine of the virtual object. Based on these controller parameters, the animation engine drives the facial expressions and lip movements of the virtual person in real time. Through the mapping processing of the neural network, the facial expressions and lip movements of the virtual person are highly consistent with the voice and emotional characteristics, which can provide natural and smooth animation performance, significantly improve the interactive experience between users and virtual objects, and enhance the sense of immersion.

[0102] The present application also provides an AI-based data processing device, such as Figure 6 As shown, it includes: an acquisition module 62, configured to acquire a feature matrix to be reduced in dimension; a mapping module 64, configured to use a Gaussian kernel function to calculate a first similarity between each data point in the feature matrix in a first space, and randomly map each data point in a plurality of data points corresponding to the feature matrix in the first space to a second space, wherein the dimension of the first space is greater than the dimension of the second space; a dimensionality reduction module 66, configured to use a T distribution to calculate a second similarity between each data point in the feature matrix in a second space, and based on the first similarity and the second similarity, adjust the position of each data point in the second space to reduce the dimension of the feature matrix.

[0103] It should be noted that the AI-based data processing device provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the AI-based data processing device provided in the above embodiment and the AI-based data processing method embodiment belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0104] Figure 7 FIG. 1 shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiment of the present disclosure. It should be noted that: Figure 7The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0105] like Figure 7 As shown, the electronic device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage part 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for system operation are also stored. The CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0106] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed.

[0107] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A data processing method based on AI, characterized in that: include: Obtain a feature matrix to be reduced in dimension, calculate a first similarity between each data point in the feature matrix in a first space using a Gaussian kernel function, and randomly map each data point in a plurality of data points corresponding to the feature matrix in the first space to a second space, wherein the dimension of the first space is greater than the dimension of the second space; A second similarity between each data point in the feature matrix is ​​calculated in a second space using T distribution, and a position of each data point in the second space is adjusted based on the first similarity and the second similarity to reduce the dimension of the feature matrix.

2. The method according to claim 1, characterized in that Calculating a first similarity between each data point in the feature matrix in the first space using a Gaussian kernel function includes: Mapping each data point in the feature matrix onto a high-dimensional spherical surface as the first space, and calculating the Euclidean distance between each data point by using the Gaussian kernel function, wherein the high-dimensional spherical surface is a spherical surface whose spatial dimension is greater than a preset dimension; A first similarity matrix representing the first similarity between data points is generated according to the Euclidean distance.

3. The method according to claim 2, characterized in that Randomly mapping each of the plurality of data points corresponding to the feature matrix in the first space to a second space includes: Generate a random mapping matrix according to the first similarity matrix; Based on the random mapping matrix, each of the data points is mapped from the first space to the second space.

4. The method according to claim 2, characterized in that: Calculating the second similarity between each data point in the feature matrix in the second space using T distribution includes: For each data point randomly mapped into the second space, calculating the Euclidean distance between each data point in the second space to generate a distance matrix; Based on the distance matrix, a T distribution function is used to perform probability distribution calculation on the distance between each data point in the second space to generate a second similarity matrix representing the second similarities between the data points.

5. The method according to claim 4, characterized in that Adjusting the position of each data point in the second space based on the first similarity and the second similarity includes: Based on the first similarity matrix and the second similarity matrix, calculating the divergence value between each data point in the second space to generate a divergence matrix; constructing, using the divergence matrix, an objective function representing the difference between the first similarity and the second similarity; By minimizing the objective function, the position of each data point in the second space is adjusted to obtain the adjusted data point position as the feature matrix after dimensionality reduction.

6. The method according to claim 5, characterized in that Calculating the divergence value between each data point in the second space based on the first similarity matrix and the second similarity matrix to generate a divergence matrix includes: For each pair of corresponding data points in the first similarity matrix and the second similarity matrix, calculating the difference between the data point pairs to obtain a difference matrix; Based on the difference matrix, using a divergence function to quantify the differences of the data point pairs to generate a divergence value; Based on the scatter value, the scatter matrix is ​​generated, wherein the scatter matrix is ​​used to represent the position change of each data point in the first space and the second space.

7. A method for driving a virtual object, characterized in that: include: Acquire speech data and emotion data corresponding to the speech data, extract speech features from the speech data, and extract emotion features from the emotion data; Based on the factor graph, the speech feature and the emotion feature are fused to obtain a fused feature, and the fused feature is subjected to dimensionality reduction processing by the method described in any one of claims 1 to 6; Performing animation parameter mapping processing on the fusion features after dimension reduction to obtain controller parameters for driving facial expressions and lip shapes of virtual objects; The virtual object is driven based on the controller parameters.

8. An AI-based data processing device, characterized in that: include: An acquisition module is configured to acquire a feature matrix to be reduced in dimension; A mapping module is configured to calculate a first similarity between each data point in the feature matrix in a first space using a Gaussian kernel function, and randomly map each data point in a plurality of data points corresponding to the feature matrix in the first space to a second space, wherein a dimension of the first space is greater than a dimension of the second space; The dimension reduction module is configured to calculate the second similarity between each data point in the feature matrix in the second space using T distribution, and adjust the position of each data point in the second space based on the first similarity and the second similarity to reduce the dimension of the feature matrix.

9. A computer device, characterized in that: include: Memory and processor, The memory stores a computer program; The processor is used to execute the computer program stored in the memory, and when the computer program is run, the processor is enabled to execute the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images