Electroencephalogram emotion recognition method based on graph neural network and federal learning
By constructing dynamic graph sequences and fusing features based on graph neural networks and federated learning, the problems of data heterogeneity and small sample size in EEG emotion recognition are solved, achieving efficient EEG emotion recognition and model generalization, and improving recognition accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-10
AI Technical Summary
Existing brain disease diagnostic models face problems such as data heterogeneity, privacy protection and data silos, small sample size and insufficient feature extraction, especially when using rs-fMRI data for EEG emotion recognition, resulting in poor model training performance and insufficient generalization ability.
This study employs a graph neural network and federated learning approach. By acquiring resting-state functional magnetic resonance imaging time series and non-image personalized data from multiple local sites, a dynamic graph sequence is constructed and features are extracted. Shared features are extracted by combining graph isomorphic network modules and Transformer modules. Personalized data is processed using independent personalized layers, and a personalized federated learning strategy is used for model training to achieve feature fusion and EEG emotion recognition.
It improves the accuracy of EEG emotion recognition and the model's generalization ability, solves the problems of data heterogeneity and small sample size, ensures privacy protection, and enhances the model's stability and feature extraction capabilities.
Smart Images
Figure CN121817889A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, and in particular to a brainwave emotion recognition method based on graph neural networks and federated learning. Background Technology
[0002] In recent years, the use of brain imaging data for the diagnosis of brain diseases has been a rapidly developing area in neuroscience. Resting-state functional magnetic resonance imaging (rs-fMRI), as a non-invasive technique, analyzes the functional connectivity networks of the brain by detecting blood oxygenation level-dependent signal fluctuations in neurons during resting states, and has been widely used in the diagnostic research of brain diseases.
[0003] Based on rs-fMRI data, researchers typically construct functional connectivity networks and analyze them using deep learning techniques such as graph neural networks. However, existing technologies face several key challenges: Data heterogeneity: Training existing diagnostic models requires large-scale datasets. However, data collected from multiple medical institutions (sites) exhibits significant feature distribution heterogeneity due to factors such as scanner model, acquisition parameters, and batch effects. Simply training on a centralized dataset often yields poor results. Privacy protection and data silos: Sharing raw rs-fMRI training data across multiple sites is extremely difficult in reality due to patient privacy, data ownership, and related regulatory restrictions, leading to the "data silo" problem. Small sample size: The data sample size of a single medical institution is usually limited, which restricts the generalization ability and stability of the constructed classification model. Insufficient feature extraction: Traditional graph neural network methods are mostly based on static functional connectivity networks, ignoring the dynamic changes in functional magnetic resonance imaging data over time. Furthermore, relying solely on a single brain atlas for brain region segmentation may fail to fully capture the complex functional organizational characteristics of the brain. Summary of the Invention
[0004] This application provides a brainwave emotion recognition method based on graph neural networks and federated learning to address the problems of data heterogeneity, privacy protection, small sample size, and insufficient feature extraction in existing technologies. The method is applied to a brainwave emotion recognition model, which includes a shared feature layer and independent personalized layers.
[0005] Acquire resting-state functional magnetic resonance imaging time-series and non-image personalized data distributed across multiple local sites;
[0006] Preprocessing of resting-state functional magnetic resonance imaging time series;
[0007] The preprocessed resting-state functional magnetic resonance imaging time series were used to construct dynamic map sequences based on multiple preset brain atlases and using the sliding window technique;
[0008] The dynamic graph sequence is input into the shared feature layer for feature extraction to obtain a spatiotemporal shared feature vector. The shared feature layer includes a graph isomorphic network module and a Transformer module.
[0009] Non-image personalized data is input into an independent personalized layer for feature extraction to obtain a personalized feature vector. The independent personalized layer is composed of a multi-layer sensing mechanism.
[0010] Feature fusion processing is performed on spatiotemporally shared feature vectors and personalized feature vectors to obtain fused features;
[0011] The fused features are mapped to the probability of EEG emotion categories by a classifier to obtain the EEG emotion recognition result.
[0012] Optionally, the preprocessed resting-state functional magnetic resonance imaging time series is constructed based on multiple preset brain atlases and using a sliding window technique to create a dynamic map sequence, including:
[0013] The resting-state functional magnetic resonance imaging time series is divided into multiple time windows according to the preset sliding window size and step size;
[0014] Within each time window, the correlation coefficients between time series of each region of interest divided according to the preset first or second brain map are calculated, and a functional connectivity matrix is constructed based on the correlation coefficients.
[0015] The correlation coefficients of the functional connection matrix are filtered according to preset threshold conditions to obtain the adjacency matrix;
[0016] The functional connection matrix is used as the node feature matrix, and a dynamic graph sequence for each time window is constructed based on the node feature matrix and the adjacency matrix.
[0017] Optionally, the step of filtering the correlation coefficients of the functional connectivity matrix according to a preset threshold condition to obtain the adjacency matrix includes:
[0018] All correlation coefficients in the functional connectivity matrix are sorted by numerical value, and the top k% of correlation coefficients are selected as strong connections.
[0019] Set the correlation coefficients of unselected functional connections to zero, retain the correlation coefficients of strong connections, and generate an adjacency matrix containing only strong connections.
[0020] Optionally, the step of inputting the dynamic graph sequence into the shared feature layer for feature extraction to obtain a spatiotemporal shared feature vector includes:
[0021] The dynamic graph sequence of each time window constructed based on the first or second brain map is input into the graph isomorphic network module to aggregate neighbor node information and obtain a node-level feature matrix.
[0022] The mean readout layer is used to aggregate node features in the node-level feature matrix into graph-level vectors.
[0023] Input the graph-level vector sequence of all time windows into the Transformer encoder to model the dynamic temporal pattern of brain functional connectivity, and obtain the output feature vector of the first brain map branch or the second brain map branch;
[0024] The output feature vectors of the first and second brain map branches are concatenated, and the concatenated feature vector is input into a multilayer perceptron structure for nonlinear transformation and dimensionality reduction to obtain a spatiotemporal shared feature vector.
[0025] Optionally, the dynamic graph sequence of each time window constructed based on the first or second brain map is input into the graph isomorphic network module to aggregate neighbor node information, thereby obtaining a node-level feature matrix that satisfies the following formula:
[0026]
[0027] in, It is a node-level feature matrix. It is a multilayer perceptron structure. It is the node feature matrix. It is an adjacency matrix. These are learnable parameters. It is a learnable scalar.
[0028] Optionally, the mean readout layer aggregates the node features in the node-level feature matrix into a graph-level vector, satisfying the following formula:
[0029]
[0030] in, A vector at the graph level, For the first The set of nodes in a dynamic graph sequence within a time window. For the number of nodes, For the first Nodes under each time window eigenvectors.
[0031] Optionally, the training process of the EEG emotion recognition model includes:
[0032] A personalized federated learning strategy is used to train the shared feature layer and the independent personalized layer. In each round of federated learning, the model parameters of the shared feature layer are uploaded to the central server for aggregation and update, while the model parameters of the independent personalized layer are retained and updated locally on the local site.
[0033] Optionally, the loss function used in the training process of the EEG emotion recognition model satisfies the following formula:
[0034]
[0035] in, Cross-entropy classification loss, This is the orthogonality regularization term corresponding to the first brain map. For the orthogonality regularization term corresponding to the second brain map, For hyperparameters
[0036] This application provides a brainwave emotion recognition method based on graph neural networks and federated learning. The method acquires resting-state functional magnetic resonance imaging (fMRI) time series and non-image personalized data distributed across multiple local sites; preprocesses the resting-state fMRI time series; constructs a dynamic graph sequence based on multiple preset brain atlases using a sliding window technique; inputs the dynamic graph sequence into a shared feature layer for feature extraction to obtain a spatiotemporal shared feature vector, the shared feature layer including a graph isomorphic network module and a Transformer module; inputs the non-image personalized data into an independent personalized layer for feature extraction to obtain a personalized feature vector, the independent personalized layer consisting of a multilayer perceptron; performs feature fusion processing on the spatiotemporal shared feature vector and the personalized feature vector to obtain fused features; and maps the fused features to brainwave emotion category probabilities using a classifier to obtain the brainwave emotion recognition result, thus improving the accuracy of brainwave emotion recognition and the model's generalization ability. Attached Figure Description
[0037] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0038] Figure 1 A flowchart illustrating the EEG emotion recognition method based on graph neural networks and federated learning provided in this application embodiment;
[0039] Figure 2 This is a schematic diagram of the processing flow of the shared feature layer provided in the embodiments of this application;
[0040] Figure 3 This is a flowchart illustrating the personalized federated learning mechanism provided in the embodiments of this application.
[0041] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0042] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0043] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.
[0044] In this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.
[0045] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0046] Figure 1 This is a flowchart illustrating the EEG emotion recognition method based on graph neural networks and federated learning provided in this application embodiment. This EEG emotion recognition method based on graph neural networks and federated learning is applied to an EEG emotion recognition model, which includes a shared feature layer and independent personalized layers. Figure 1 As shown, the EEG emotion recognition method based on graph neural networks and federated learning provided in this embodiment includes:
[0047] S1: Acquire resting-state functional magnetic resonance imaging time series and non-image personalized data distributed across multiple local sites.
[0048] The multi-site resting-state functional magnetic resonance imaging (fMRI) data were derived from the publicly available REST-meta-MDD dataset, which contains rs-fMRI data from 1300 patients with major depressive disorder (MDD) and 1128 healthy controls from multiple hospitals in China. Preferably, data from sites 15, 20, 21, and 25 were used for validation. Simultaneously, relevant demographic information (such as age, sex, and years of education) was obtained as non-image-personalized data.
[0049] S2: Preprocessing of resting-state functional magnetic resonance imaging time series.
[0050] Specifically, the DPARSF standard workflow was used to preprocess the rs-fMRI data, including: discarding the first 10 time points, temporal slice correction, head motion correction, spatial smoothing, removal of covariates (such as head motion parameters, white matter signals, etc.), spatial normalization (registration to MNI space), and temporal bandpass filtering (0.01-0.1Hz). To ensure consistency, all time series were truncated to 230 time points.
[0051] S3: The preprocessed resting-state functional magnetic resonance imaging time series are used to construct dynamic graph sequences based on multiple preset brain atlases and the sliding window technique.
[0052] Specifically, it includes the following steps:
[0053] S31: Divide the resting-state functional magnetic resonance imaging time series into multiple time windows according to the preset sliding window size and step size.
[0054] For example, setting the window size At each time point, the step size Time points, time series Classified as Overlapping time windows.
[0055] S32: Within each time window, calculate the correlation coefficient between the time series of each region of interest divided according to the preset first or second brain map, and construct the functional connectivity matrix based on the correlation coefficient.
[0056] The pre-defined first and second brain atlases are two different types of brain atlases. These atlases could be, for example, an Automated Anatomical Labelling (AAL) atlas or a Harvard-Oxford Atlas. The brain is divided according to the AAL atlas. The brain is divided into regions of interest according to the Harvard-Oxford Atlas. Areas of interest.
[0057] In each time window Inside, calculation The Pearson correlation coefficients between the time series of each region of interest yield the functional connectivity matrix of the window. .
[0058] S33: Filter the correlation coefficients of the functional connection matrix according to the preset threshold conditions to obtain the adjacency matrix.
[0059] Specifically, to reduce noise and highlight strong connections, the functional connection matrix is... All correlation coefficients are sorted by numerical value, and the top k% (preferred) are selected. The correlation coefficients of the functional connections are used as strong connections; the correlation coefficients of unselected connections in the functional connection matrix are set to zero, while the correlation coefficients corresponding to strong connections are retained, generating an adjacency matrix containing only strong connections. .
[0060] S34: Use the functional connection matrix as the node feature matrix, and construct a dynamic graph sequence for each time window based on the node feature matrix and the adjacency matrix.
[0061] Specifically, each subject is represented as a dynamic graph sequence. Each of the graphs , This is the node feature matrix.
[0062] S4: Input the dynamic graph sequence into the shared feature layer for feature extraction to obtain the spatiotemporal shared feature vector.
[0063] The shared feature layer includes a Graph Isomorphic Network (GIN) module and a Transformer module.
[0064] Figure 2 This is a schematic diagram of the processing flow of the shared feature layer provided in an embodiment of this application. The dynamic graph sequence is input into the shared feature layer for feature extraction to obtain a spatiotemporal shared feature vector, including:
[0065] S41: Input the dynamic graph sequence of each time window constructed based on the first or second brain map into the graph isomorphic network module to aggregate neighbor node information and obtain a node-level feature matrix.
[0066] Specifically, the graph for each time window Input the graph isomorphic network module. Graph isomorphic networks update the representation of the central node by aggregating neighbor node information, and their update rules are as follows:
[0067]
[0068] in, It is a node-level feature matrix. It is a multilayer perceptron structure. It is the node feature matrix. It is an adjacency matrix. These are learnable parameters. It is a learnable scalar. Graph isomorphic networks can effectively distinguish different graph structures and capture the complex topological characteristics of functional connections in the brain.
[0069] S42: The mean readout layer is used to aggregate the node features in the node-level feature matrix into a graph-level vector.
[0070] Specifically, the graph isomorphic network outputs a node-level feature matrix. A mean-based readout layer is used to aggregate node features into graph-level vectors. It satisfies the following formula:
[0071]
[0072] in, A vector at the graph level, For the first The set of nodes in a dynamic graph sequence within a time window. For the number of nodes, For the first Nodes under each time window eigenvectors.
[0073] S43: Input the graph-level vector sequence of all time windows into the dynamic temporal pattern of the Transformer encoder to model the brain's functional connectivity, and obtain the output feature vector of the first or second brain map branch.
[0074] Specifically, the graph-level vector sequences of all time windows The input is fed into the Transformer encoder for analysis and processing, yielding the output features, as shown below:
[0075]
[0076] in, This represents the output feature vector. It represents the Transformer encoder.
[0077] Understandably, Transformer's self-attention mechanism can effectively capture long-range dependencies between any two time points in a sequence, thereby modeling the dynamic temporal patterns of brain functional connectivity.
[0078] S44: Concatenate the output feature vectors of the first brain map branch and the second brain map branch, and input the concatenated feature vector into a multilayer perceptron structure for nonlinear transformation and dimensionality reduction to obtain a spatiotemporal shared feature vector.
[0079] For example, the automatic anatomical labeling atlas branch and the Harvard-Oxford atlas branch each output feature vectors. and The two vectors are concatenated and input into a multilayer perceptron for nonlinear transformation and dimensionality reduction to obtain the final spatiotemporal shared feature vector. :
[0080]
[0081] in, Represents a multilayer perceptron. This represents vector multiplication.
[0082] S5: Input non-image personalized data into an independent personalized layer for feature extraction to obtain personalized feature vectors.
[0083] The independent personalization layer consists of multiple sensing mechanisms.
[0084] Understandably, a separate personalization layer is used to process non-image personalization data, such as age, gender, and years of education. This data can vary significantly between different sites.
[0085] Specifically, standardized demographic information The input is fed into a multilayer perceptron to extract personalized feature vectors. :
[0086]
[0087] in, The mean, The standard deviation is denoted as .
[0088] S6: Perform feature fusion processing on the spatiotemporal shared feature vector and the personalized feature vector to obtain the fused feature.
[0089] For example, spatiotemporal shared feature vectors and personalized feature vectors can be concatenated or summed to obtain fused features.
[0090] S7: The fused features are mapped to the probability of EEG emotion categories through a classifier to obtain the EEG emotion recognition result.
[0091] Specifically, the fused features are input into the final multilayer perceptron classifier, and the probability of the EEG emotion category is obtained by processing it through the Softmax function. :
[0092] .
[0093] In an optional embodiment, the training process of the EEG emotion recognition model includes: training a shared feature layer and an independent personalized layer using a personalized federated learning strategy, wherein the personalized federated learning strategy is a personalized federated averaging strategy. In each round of federated learning, the model parameters of the shared feature layer are uploaded to a central server for aggregation and updating, while the model parameters of the independent personalized layer are kept on the local site for local updating and do not participate in aggregation. Figure 3 This is a flowchart illustrating the personalized federated learning mechanism provided in the embodiments of this application.
[0094] For example, the training process includes:
[0095] (1) Local training: Each local site Its model parameters, including shared layer parameters, are trained independently using its local data (rs-fMRI and demographic information). and personalized layer parameters The optimal number of training rounds is 20 or 25.
[0096] (2) Parameter upload: After local training is completed, each site Only update the shared layer parameters Uploaded to the central server.
[0097] (3) Global aggregation: The central server performs a weighted average of the shared layer parameters uploaded by all sites to obtain the global shared parameters. .
[0098] (4) Parameter distribution: The server will distribute the aggregated parameters. Distribute back to all local sites.
[0099] (5) Model update: local site use Replace its local shared layer parameters, but retain its local personalized layer parameters. .
[0100] Repeat steps (1)-(5) above until the model converges (e.g., the number of federated learning rounds is preferably 30 or 35 rounds).
[0101] Furthermore, to improve model stability and feature quality, this invention employs a combined loss function during training:
[0102]
[0103] in, For the sample size, For real labels, The predicted probability output by the model. For hyperparameters, This is the orthogonality regularization term corresponding to the first brain map. This is the orthogonality regularization term corresponding to the second brain map. Understandably, the first term is the cross-entropy classification loss, and the second term is the orthogonality regularization term used to influence the node features output by the graph isomorphic network. They are independent of each other, reducing feature redundancy.
[0104] The EEG emotion recognition method based on graph neural networks and federated learning provided in this application acquires resting-state functional magnetic resonance imaging (fMRI) time series and non-image personalized data distributed across multiple local sites; preprocesses the resting-state fMRI time series; constructs a dynamic graph sequence based on multiple preset brain atlases and using a sliding window technique; inputs the dynamic graph sequence into a shared feature layer for feature extraction to obtain a spatiotemporal shared feature vector, the shared feature layer including a graph isomorphic network module and a Transformer module; inputs the non-image personalized data into an independent personalized layer for feature extraction to obtain a personalized feature vector, the independent personalized layer consisting of a multilayer perceptron; performs feature fusion processing on the spatiotemporal shared feature vector and the personalized feature vector to obtain fused features; and maps the fused features to EEG emotion category probabilities using a classifier to obtain the EEG emotion recognition result, thereby improving the accuracy of EEG emotion recognition and the generalization ability of the model.
[0105] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.
[0106] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A brainwave emotion recognition method based on graph neural networks and federated learning, characterized in that, An application is made to an EEG emotion recognition model, the EEG emotion recognition model comprising: a shared feature layer and an independent personalized layer, the method comprising: Acquire resting-state functional magnetic resonance imaging time-series and non-image personalized data distributed across multiple local sites; Preprocessing of resting-state functional magnetic resonance imaging time series; The preprocessed resting-state functional magnetic resonance imaging time series were used to construct dynamic map sequences based on multiple preset brain atlases and using the sliding window technique; The dynamic graph sequence is input into the shared feature layer for feature extraction to obtain a spatiotemporal shared feature vector. The shared feature layer includes a graph isomorphic network module and a Transformer module. Non-image personalized data is input into an independent personalized layer for feature extraction to obtain a personalized feature vector. The independent personalized layer is composed of a multi-layer sensing mechanism. Feature fusion processing is performed on spatiotemporally shared feature vectors and personalized feature vectors to obtain fused features; The fused features are mapped to the probability of EEG emotion categories by a classifier to obtain the EEG emotion recognition result.
2. The method according to claim 1, characterized in that, The preprocessed resting-state functional magnetic resonance imaging time series is constructed based on multiple preset brain atlases and using a sliding window technique to create a dynamic map sequence, including: The resting-state functional magnetic resonance imaging time series is divided into multiple time windows according to the preset sliding window size and step size; Within each time window, the correlation coefficients between time series of each region of interest divided according to the preset first or second brain map are calculated, and a functional connectivity matrix is constructed based on the correlation coefficients. The correlation coefficients of the functional connection matrix are filtered according to preset threshold conditions to obtain the adjacency matrix; The functional connection matrix is used as the node feature matrix, and a dynamic graph sequence for each time window is constructed based on the node feature matrix and the adjacency matrix.
3. The method according to claim 2, characterized in that, The step of filtering the correlation coefficients of the functional connectivity matrix according to a preset threshold condition to obtain the adjacency matrix includes: All correlation coefficients in the functional connectivity matrix are sorted by numerical value, and the top k% of correlation coefficients are selected as strong connections. Set the correlation coefficients of unselected functional connections to zero, retain the correlation coefficients of strong connections, and generate an adjacency matrix containing only strong connections.
4. The method according to claim 1, characterized in that, The step of inputting the dynamic graph sequence into the shared feature layer for feature extraction to obtain a spatiotemporal shared feature vector includes: The dynamic graph sequence of each time window constructed based on the first or second brain map is input into the graph isomorphic network module to aggregate neighbor node information and obtain a node-level feature matrix. The mean readout layer is used to aggregate node features in the node-level feature matrix into graph-level vectors. Input the graph-level vector sequence of all time windows into the Transformer encoder to model the dynamic temporal pattern of brain functional connectivity, and obtain the output feature vector of the first brain map branch or the second brain map branch; The output feature vectors of the first and second brain map branches are concatenated, and the concatenated feature vector is input into a multilayer perceptron structure for nonlinear transformation and dimensionality reduction to obtain a spatiotemporal shared feature vector.
5. The method according to claim 1, characterized in that, The dynamic graph sequence of each time window constructed based on the first or second brain map is input into the graph isomorphic network module to aggregate neighbor node information, thereby obtaining a node-level feature matrix that satisfies the following formula: ； in, It is a node-level feature matrix. It is a multilayer perceptron structure. It is the node feature matrix. It is an adjacency matrix. These are learnable parameters. It is a learnable scalar.
6. The method according to claim 1, characterized in that, The mean readout layer aggregates the node features in the node-level feature matrix into a graph-level vector, satisfying the following formula: ； in, A vector at the graph level, For the first The set of nodes in a dynamic graph sequence within a time window. For the number of nodes, For the first Nodes under each time window eigenvectors.
7. The method according to claim 1, characterized in that, The training process of the EEG emotion recognition model includes: A personalized federated learning strategy is used to train the shared feature layer and the independent personalized layer. In each round of federated learning, the model parameters of the shared feature layer are uploaded to the central server for aggregation and update, while the model parameters of the independent personalized layer are retained and updated locally on the local site.
8. The method according to claim 7, characterized in that, The loss function used in the training process of the EEG emotion recognition model satisfies the following formula: ； in, Cross-entropy classification loss, This is the orthogonality regularization term corresponding to the first brain map. For the orthogonality regularization term corresponding to the second brain map, This is a hyperparameter.