Neural network online learning method and system based on blockchain
By using blockchain technology to selectively update neural network parameters based on the degree of attention, the problems of slow training and overfitting/underfitting in existing methods are solved, and fast and accurate neural network training is achieved.
Patent Information
- Application Number
- CN202210363411.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-08
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-04-08
AI Technical Summary
Existing neural network training methods fail to fully utilize learned features when new datasets are generated, resulting in slow convergence during retraining. Furthermore, difficulties in data sharing between different hospitals lead to low network accuracy and a tendency to overfit or underfit.
By leveraging blockchain technology, the attention level can be calculated using the feature matrix and accuracy of a trained DNN network. This allows for selective updates to network parameters, and training with new datasets to quickly learn more accurate features, reducing the risk of overfitting and underfitting.
It achieves rapid convergence of neural networks, improves training accuracy, reduces retraining time, and enhances the network's ability to adapt to new data.
Smart Images

Figure CN114677563B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of neural network online learning, and in particular to a neural network online learning method and system based on blockchain. Background Art
[0002] With the continuous advancement of medical technology, medical image data has become an important analytical data for disease diagnosis and treatment. At the same time, with the development of artificial intelligence technology, the use of neural networks to analyze and process image data has become a development trend in the field of data analysis.
[0003] However, since the labeling of medical image data requires professionals, data sets are difficult to obtain and are small in number. Also, since it is difficult for different hospitals to share data sets on a large scale, the accuracy of the neural network trained by each hospital using its own data set is low, and it is prone to overfitting or underfitting problems. The existing federated learning method or other methods provide a method to train an accurate neural network using the features of the data sets of all hospitals without sharing data sets. However, as new image data are labeled, new data sets are constantly generated, and the size of the data sets continues to grow. At this time, new data sets are needed to retrain the parameters of the old neural network.
[0004] The existing technology has the following defects: the existing method is to use the new data set and the previous old data set as a large data set to retrain the parameters of the old neural network, but this method requires all old data sets to participate in the training, and does not fully utilize the features that the neural network has learned, resulting in slow convergence when the network is retrained. Summary of the invention
[0005] In order to solve the above technical problems, the purpose of the present invention is to provide a method and system for online learning of a neural network based on blockchain. The technical solutions adopted are as follows:
[0006] In a first aspect, an embodiment of the present invention provides a method and system for online learning of a neural network based on blockchain, the method comprising the following steps:
[0007] According to the trained initial DNN network stored on the blockchain, the feature matrix corresponding to each medical image in the initial training set and the accuracy rate corresponding to the feature matrix are obtained;
[0008] Calculate the expected feature matrix based on the feature matrix corresponding to each medical image in the initial training set and the accuracy rate corresponding to each feature matrix;
[0009] Input each medical image in the training data set into the encoder of the initial DNN network to obtain a feature matrix corresponding to each medical image in the training data set and an accuracy rate corresponding to the feature matrix;
[0010] Calculate the degree of attention corresponding to each medical image in the initial training set according to the expected feature matrix and the feature matrix corresponding to each medical image in the training data set;
[0011] Determine whether the attention level corresponding to each medical image in the initial training set is greater than the set attention level threshold. If so, use the corresponding medical image and the training data set together to update the initial DNN network, and store the updated DNN network on the blockchain.
[0012] In a second aspect, another embodiment of the present invention provides a blockchain-based neural network online learning system, comprising a memory and a processor, wherein the processor executes a computer program stored in the memory to implement the above-mentioned blockchain-based neural network online learning method.
[0013] Preferably, the method for calculating the degree of attention corresponding to each medical image in the initial training set includes:
[0014] Obtaining the abnormality degree corresponding to each feature matrix according to the expected feature matrix, the feature matrix corresponding to each medical image in the training data set, and the corresponding accuracy rate;
[0015] Calculate the initial training set difference based on the feature matrix corresponding to each medical image in the initial training set and the expected feature matrix;
[0016] Calculate the training set difference based on the feature matrix corresponding to each medical image in the initial training set and the feature matrix corresponding to each medical image in the to-be-trained data set;
[0017] Calculating the sub-attention degree corresponding to each medical image in the initial training set according to the abnormality degree, the initial training set difference, and the training set difference;
[0018] The multiple sub-attention levels corresponding to each medical image in the initial training set are summed to obtain the corresponding attention level of each medical image in the initial training set.
[0019] Preferably, the calculation formula of the sub-attention degree is:
[0020]
[0021] Among them, F(f,d) is the sub-attention level corresponding to the medical image f in the initial training set calculated based on the medical image d in the training set, and g d is the abnormality degree corresponding to the medical image d in the training data set, Ff is the feature matrix corresponding to the medical image f in the initial training set, F d is the feature matrix corresponding to the medical image d in the training data set, and F0 is the expected feature matrix.
[0022] Preferably, the abnormality degree is calculated using the following calculation formula:
[0023] g d =exp(-α d )||F d -F0||2
[0024] Among them, α d is the accuracy of the medical image d in the training data set on the initial DNN network, and || ||2 is the L2 norm.
[0025] Preferably, the calculation formula of the expected feature matrix is:
[0026]
[0027] α f is the accuracy of the medical image f in the initial training set on the initial DNN network, and S is the set of feature matrices corresponding to the medical images in the initial training set.
[0028] Preferably, the trained initial DNN network is trained using a federated learning method.
[0029] The embodiments of the present invention have the following beneficial effects:
[0030] The present invention obtains the feature matrix of medical images in the initial training set, the accuracy corresponding to the feature matrix, and the feature matrix in the data set to be trained based on the trained DNN network, and then calculates the degree of attention of each medical image in the initial training set. The degree of attention can reflect the degree of learning of the features of each medical image in the initial training set by the DNN network. The greater the degree of attention, the worse the feature learning of the medical images in the initial training set, and the smaller the degree of attention, the worse the feature learning of the medical images in the initial training set. According to whether the degree of attention is greater than the set degree of attention threshold, it is determined which medical images in the initial training set are used together with the data set to be trained to update the initial DNN network; after the hospital continuously generates new annotated data, the DNN network can quickly learn more and more accurate features, thereby reducing the risk of underfitting and overfitting, and can also make the DNN network converge quickly. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0032] Figure 1 A flowchart of a blockchain-based neural network online learning method provided by one embodiment of the present invention. DETAILED DESCRIPTION
[0033] In order to further explain the technical means and functional effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a blockchain-based neural network online learning method and system proposed by the present invention in combination with the accompanying drawings and preferred embodiments.
[0034] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0035] Since the annotation of medical image data requires professionals, data sets are difficult to obtain and are small in number. It is also difficult to share data sets on a large scale between different hospitals, resulting in low accuracy of neural networks trained by each hospital using its own data set, which is prone to overfitting or underfitting problems. The existing federated learning method provides a method to train an accurate neural network using the features of all hospital data sets without sharing data sets. However, when a hospital generates a new data set, the existing method uses the new data set and the old data set to retrain the network. This method does not fully utilize the features that the neural network has already learned, resulting in slow convergence when the network is retrained, and is not conducive to learning data features in the new data set.
[0036] The online learning method of the present invention is mainly that when a new data set is generated, each hospital can obtain the initial DNN network, and specifically select part of the images in the initial training set as samples to update the DNN network together with the newly generated data set. The main idea of the present invention is: according to the feature matrix corresponding to each medical image in the initial training set, the accuracy rate corresponding to the feature matrix and the medical image corresponding to the data set to be trained, the attention level of each image in the initial training set is calculated. The size of this attention level determines whether the image in the old training set participates in the update of the DNN network, so that the DNN network can converge quickly and learn more and more accurate features.
[0037] The following is a detailed description of a specific scheme of a blockchain-based neural network online learning method and system provided by the present invention in conjunction with the accompanying drawings.
[0038] Implementation example of a neural network online learning method based on blockchain
[0039] like Figure 1 As shown, the blockchain-based neural network online learning method of this embodiment includes the following steps:
[0040] Step S1, based on the trained initial DNN network stored on the blockchain, obtain the feature matrix corresponding to each medical image in the initial training set and the accuracy corresponding to the feature matrix.
[0041] The initial DNN network of this embodiment is trained by the federated learning method using the respective data sets of all hospitals, where the data set of each hospital is divided into a training set and a test set. This trained DNN network is called the trained initial DNN network, the training set of each hospital participating in the training of the DNN network is called the initial training set, and the test set used to test the accuracy of the DNN network is called the initial test set.
[0042] Each hospital has labeled the new medical images, that is, each hospital has a new data set. 80% of this data set is used as a new training set, and the remaining 20% is used as a new test set. This new training set is called the to-be-trained data set, which has not yet participated in the training of the initial DNN network.
[0043] In this embodiment, the initial DNN network has been stored on the blockchain. Each hospital reads the initial DNN network from the blockchain, inputs all image data in the hospital's initial training set into the DNN network, and obtains all feature maps output by the encoder of the DNN network. We call the feature map a feature matrix.
[0044] Specifically, the structure of the DNN network is an encoder-decoder structure, which is an important structure in neural network design. Among them, the role of the encoder is to extract and downsample the high-dimensional features of the image. The image is input into the encoder, which contains multiple convolutional layers, each of which outputs a multi-channel feature map. The encoder finally outputs a feature map, and the feature map F is the feature matrix described in the present invention. The size of the feature map F is much lower than that of the input image, but the feature map has a higher number of channels, and this feature map contains the high-dimensional features of the image. The task of the DNN network in this embodiment is to perform semantic segmentation on the image to obtain the lesion area on the medical image, so the feature map F should represent the characteristics of the lesion; and before the feature map F is output in the encoder, the feature map f output by other convolutional layers close to the input also includes the lesion characteristics, but also includes other features; in general, the feature map f contains many features and has a large information entropy, so the features contained in different input images will also be different; but for the feature map F, it contains high-dimensional, more abstract, and smaller information entropy features, which represent the lesion characteristics. Although the input medical images are different, the lesion features corresponding to these medical images will not be very different, so the feature maps obtained by inputting different images are similar.
[0045] In addition, the feature map F is input into the decoder, and the decoder upsamples the feature map F and outputs a semantic area. That is, the function of the decoder is to decode the feature map F containing the feature matrix into a visualized lesion area.
[0046] Each hospital will share and store all acquired feature matrices and the accuracy of the initial DNN network on each hospital's initial test set on the blockchain.
[0047] Specifically, for a hospital a, after hospital a inputs all medical images in the initial training set into the initial DNN network, it obtains a set of feature matrices and shares the set; at the same time, hospital a also shares the accuracy of the initial DNN network on the initial test set of hospital a.
[0048] In this embodiment, the hospital with the largest number of medical images in the training data set is selected, and the hospital is asked to create a new block, and then the feature matrix and accuracy shared by all hospitals are stored in the newly created block, and then the block is connected to the blockchain. After that, the hospital is rewarded. As another implementation method, a hospital can also be designated to create blocks, store data, and connect blocks to the blockchain.
[0049] Step S2, calculating the expected feature matrix according to the feature matrix corresponding to each medical image in the initial training set and the accuracy rate corresponding to each feature matrix.
[0050] For a hospital a, hospital a reads all feature matrices stored on the blockchain, as well as the accuracy rate corresponding to each feature matrix, and then calculates a centroid position F0 based on all feature matrices and the accuracy rate corresponding to each feature matrix, which is called the expected feature matrix.
[0051] The specific calculation formula of the expected feature matrix is:
[0052]
[0053] Among them, S is the set of all feature matrices, F f is any feature matrix in S. This feature matrix is obtained by inputting the medical image f in the initial training set into the encoder of the initial DNN network, and its corresponding accuracy is α f .
[0054] Among them, the feature matrix F f It is a three-dimensional matrix, which can be regarded as a feature point in a high-dimensional space. The distribution of all these feature points represents the distribution of the disease in the initial data set. Ideally, when the number of initial training sets is large enough, there is no abnormal data and the image data is widely distributed, these feature points should be concentrated together. The reason is that the data set annotated by the present invention is a medical image of different patients suffering from a certain disease. The content of these medical images may not be the same, but the feature matrix embodied is the same. However, due to the insufficient data set for training the initial DNN network, the image data distribution in the initial training set is unbalanced. For example, some diseases have collected a lot of medical images, while some diseases have very few medical images. In addition, the data annotation is abnormal or abnormal sample data is introduced, which may cause these feature points to be discretely distributed, or the distribution is not concentrated enough. This may be caused by overfitting or underfitting of the initial DNN network, which is specifically related to the data distribution in the initial data set. However, whether it is overfitting or underfitting, the distribution of these feature points can still roughly determine how the real expected feature matrix is distributed, that is, it can reflect what the expected value F0 of the disease feature is. The expected value is used to characterize the ideal value around which these feature points should be concentrated.
[0055] Step S3, input each medical image in the training data set into the encoder of the initial DNN network to obtain the feature matrix corresponding to each medical image in the training data set and the accuracy corresponding to the feature matrix.
[0056] For a hospital a, first, hospital a reads the initial DNN network from the blockchain, then obtains a medical image d in the training data set of hospital a, inputs d into the encoder of the DNN network, and outputs the corresponding feature matrix. This corresponding feature matrix is called F d, and at the same time, we get the feature matrix F d The corresponding accuracy.
[0057] Step S4, calculating the attention level corresponding to each medical image in the initial training set according to the expected feature matrix and the feature matrix corresponding to each medical image in the training data set.
[0058] The degree of attention corresponding to each medical image in the initial training set calculated in this embodiment is calculated based on the abnormality degree of the feature matrix corresponding to the medical image to be trained. Specifically, this embodiment obtains the abnormality degree corresponding to each feature matrix based on the expected feature matrix, the feature matrix corresponding to the medical image to be trained, and the corresponding accuracy. The formula for calculating the abnormality degree of the feature matrix in this embodiment is:
[0059] g d =exp(-α d )||F d -F0||2
[0060] Among them, g d It represents the feature matrix F d The degree of abnormality, α d It represents the accuracy of the initial DNN network on the dataset when an image data d in the training dataset is used as the test set; ||F d -F0||2 represents the degree of deviation between the feature matrix corresponding to the image data d and the expected feature matrix. d The smaller ||F d -F0||2 is larger, the g d There are three reasons for the difference. First, the value of F0 is biased, that is, the expected value of all feature matrix distributions calculated is not F0 but other values. This is caused by the initial training set being not good enough. Second, F d The value is biased, that is, the feature matrix obtained by inputting a medical image d in the training data set into the DNN network is wrong, indicating that the initial DNN network has not learned the features of the medical image d; third, F d There are deviations between F0 and F0.
[0061] This embodiment is more inclined to the third reason, because: on the one hand, F0 is obtained based on a large number of initial training sets, so even if there is a deviation, the deviation will not be too large, so no matter what the actual value of F0 is, the calculated F0 can still roughly reflect how the feature matrix should be distributed. d is obtained from medical image d, but the initial DNN network did not learn the features of d, and the features of d are different from the medical images in the initial training set, which leads to F d There is a deviation.
[0062] This embodiment first assumes that F0 is correct, uses the medical image d to update the parameters of the initial DNN network, and then re-acquires F0 after the initial DNN network parameters are updated; through d, different medical images are taken to iteratively update the DNN network parameters and F0, and finally the DNN network can learn correct and reasonable features that conform to the actual situation.
[0063] The image data d represents a medical image in the set to be trained. In the actual process, all the data in the set to be trained need to be used to assign a degree of attention to each medical image in the initial training set.
[0064] Since little new data is generated in a short period of time, it is necessary to combine it with the initial training set. Secondly, some features of the data in the initial training set are not correctly learned by the DNN network, so it is necessary to introduce some medical images in the initial training set based on the degree of attention to keep the features that the DNN network can extract consistent before and after training.
[0065] The degree of attention is used to characterize whether the medical images in the initial training set need to participate in the training of the initial DNN network. In this embodiment, the degree of attention is calculated based on the feature matrix corresponding to each medical image in the training data set, wherein the sub-degree of attention corresponding to the medical image in the initial training set is first calculated based on the feature matrix corresponding to a single medical image in the training data set, and then the degree of attention corresponding to the medical image in the initial training set is calculated by summing the sub-degrees of attention.
[0066] The calculation formula of the degree of attention of the neutron in this embodiment is:
[0067]
[0068] Where f is any medical image in the initial training set, F f is the feature matrix corresponding to f, and F(f,d) is the degree of sub-attention assigned to f given the training data set.
[0069] Denotes the feature matrix F f The difference between the expected feature matrix F0, that is, the initial training set difference, the larger the value, the greater the feature matrix F f The further away from the expected value; Denotes the feature matrix F f and the feature matrix F d The difference is the training set difference. The larger the value, the greater the difference.
[0070] Among them, g dThe larger it is, the lower the accuracy of the features extracted by the DNN network on medical image d and the further it deviates from the expected value. Therefore, it is more necessary to combine medical image d with other medical images (such as medical image f) to guide the training of the DNN network to learn the features of medical image d.
[0071] In addition, if The larger the value, the less the initial DNN network has learned the feature matrix F. f and the feature matrix F d The characteristics of the corresponding samples, that is, the more the initial DNN network needs to update the parameters to make the feature matrix F f and the feature matrix F d Closer to the expected feature matrix F0, avoiding underfitting of the DNN network to the samples corresponding to these two features; The smaller it is, the more likely the initial DNN network has learned the feature matrix F. f and the feature matrix F d The corresponding sample characteristics, that is, the initial DNN network does not need to update parameters to make the feature matrix F f and the feature matrix F d Closer to the expected feature matrix F0, avoiding overfitting.
[0072] After hospital a obtains the corresponding attention levels for all feature matrices stored on the blockchain, it shares these attention levels. Similarly, other hospitals will also share the attention features they have obtained. Judging from the attention levels corresponding to the feature matrices shared by all hospitals, each feature matrix corresponds to multiple attention levels. The hospital with the largest sum of attention levels is selected. Since the sum of attention levels is the largest, it means that the data features of the training data set of this hospital are of great significance to the parameter update of the initial DNN network; the hospital sums the multiple attention levels corresponding to each feature matrix, and uses the result as the final attention level of each feature matrix. The hospital stores the final attention level in a newly generated block, then connects this block to the blockchain and rewards the hospital.
[0073] The process of deciding which hospital can generate a new block and connecting the new block to the blockchain as described in step S1 and step S4 is the consensus mechanism. This approach is conducive to rewarding hospitals to label some data sets that are conducive to initial DNN network feature learning, which is of great significance to improving the accuracy and representation capabilities of the DNN network.
[0074] Step S5, determine whether the attention level corresponding to each medical image in the initial training set is greater than the set attention level threshold. If so, the corresponding medical image and the training data set are used together to update the initial DNN network, and the updated DNN network is stored on the blockchain.
[0075] Specifically, assuming hospital a, read out the degree of attention corresponding to each image sample stored on the blockchain, and then make a judgment based on the degree of attention. If the degree of attention corresponding to any medical image is greater than the set threshold and the medical image comes from hospital a, then the medical image is selected and merged with the set to be trained to form a new training set to update the initial DNN network.
[0076] All hospitals use the new training set to update the parameters of the initial DNN network. The specific method is as follows: a hospital reads the latest parameters of the DNN network from the blockchain, uses its new training set to update the parameters of a round of DNN network training, and then stores the updated parameters on the block, and then connects the block to the blockchain. After that, other hospitals are selected in turn to continue to read the latest and most recently updated parameters of the DNN network, and then use the new training set they have obtained to update the parameters of a round of DNN network training, and also store the updated and latest parameters on the block. Repeat this process until each hospital has updated and trained a round. The updated DNN network is regarded as the initial DNN network for subsequent calculations, and steps S2 and S3 are repeated until the DNN network converges; in this process, the expected feature matrix F0 will also be constantly updated and changed. When the DNN network converges, all feature matrices will approach the expected value F0.
[0077] This embodiment obtains the feature matrix of the medical images in the initial training set, the accuracy corresponding to the feature matrix, and the feature matrix in the data set to be trained based on the trained DNN network, and then calculates the attention level of each medical image in the initial training set, and determines which medical images in the initial training set and the data set to be trained are used together to update the initial DNN network based on whether the attention level is greater than a set attention level threshold; as the hospital continuously generates new labeled data, the DNN network can quickly learn more and more accurate features, thereby reducing the risks of underfitting and overfitting, and can also enable the DNN network to converge quickly.
[0078] Implementation example of a neural network online learning system based on blockchain
[0079] The blockchain-based neural network online learning system of this embodiment includes a memory and a processor, and the processor executes a computer program stored in the memory to implement the above-mentioned blockchain-based neural network online learning method.
[0080] Since the blockchain-based neural network online learning method has been described in the embodiment of the blockchain-based neural network online learning method, this embodiment will not repeat the blockchain-based neural network online learning method.
[0081] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A blockchain-based neural network online learning method, characterized in that: The method comprises the following steps: According to the trained initial DNN network stored on the blockchain, the feature matrix corresponding to each medical image in the initial training set and the accuracy rate corresponding to the feature matrix are obtained; According to the feature matrix corresponding to each medical image in the initial training set and the accuracy rate corresponding to each feature matrix, the expected feature matrix is calculated. For a hospital a, hospital a reads all feature matrices stored on the blockchain and the accuracy rate corresponding to each feature matrix, and then calculates a centroid position based on all feature matrices and the accuracy rate corresponding to each feature matrix, which is called the expected feature matrix. Input each medical image in the training data set into the encoder of the initial DNN network to obtain a feature matrix corresponding to each medical image in the training data set and an accuracy rate corresponding to the feature matrix; According to the expected feature matrix and the feature matrix corresponding to each medical image in the training data set, the attention degree corresponding to each medical image in the initial training set is calculated. The attention degree is used to characterize whether the medical image in the initial training set needs to participate in the training of the initial DNN network. The attention degree is calculated based on the feature matrix corresponding to each medical image in the training data set, wherein the sub-attention degree corresponding to the medical image in the initial training set is first calculated based on the feature matrix corresponding to a single medical image in the training data set, and then the attention degree corresponding to the medical image in the initial training set is calculated by summing the sub-attention degrees; Determine whether the attention level corresponding to each medical image in the initial training set is greater than a set attention level threshold. If so, use the corresponding medical image and the training data set together to update the initial DNN network, and store the updated DNN network on the blockchain; The method for calculating the degree of attention corresponding to each medical image in the initial training set includes: Obtaining the abnormality degree corresponding to each feature matrix according to the expected feature matrix, the feature matrix corresponding to each medical image in the training data set, and the corresponding accuracy rate; Calculate the initial training set difference based on the feature matrix corresponding to each medical image in the initial training set and the expected feature matrix; Calculate the training set difference based on the feature matrix corresponding to each medical image in the initial training set and the feature matrix corresponding to each medical image in the to-be-trained data set; Calculating the sub-attention degree corresponding to each medical image in the initial training set according to the abnormality degree, the initial training set difference, and the training set difference; The multiple sub-attention levels corresponding to each medical image in the initial training set are summed to obtain the corresponding attention level of each medical image in the initial training set; The calculation formula of the sub-attention degree is: in, is the medical image in the initial training set calculated based on the medical image d in the training set The corresponding sub-attention level, is the abnormality degree corresponding to the medical image d in the training data set, Medical images in the initial training set The corresponding feature matrix is, The feature matrix corresponding to the medical image d in the training data set, is the expected feature matrix; The calculation formula of the abnormality degree is: in, Medical images in the training dataset The accuracy on the initial DNN network, It is the L2 paradigm; The calculation formula of the expected feature matrix is: , is the accuracy of the medical image f in the initial training set on the initial DNN network, and S is the set of feature matrices corresponding to the medical images in the initial training set.
2. The blockchain-based neural network online learning method according to claim 1 is characterized in that: The trained initial DNN network is trained using a federated learning method.
3. A blockchain-based neural network online learning system, comprising a memory and a processor, characterized in that: The processor executes the computer program stored in the memory to implement the blockchain-based neural network online learning method as described in any one of claims 1-2.
Citation Information
Patent Citations
A method of training set selection based on depth learning
CN109508741A
A video classification method and device
CN109902202A
Method for maintaining fixed capacity of high-speed railway network diagram under abnormal events
CN110843870A
Construction site personnel exception aggregation detection system based on block chain and BIM
CN111523434A
Malicious behavior recognition method based on federated learning and feature extraction
CN113990454A
