An efficient training method for intelligent acceleration platform model based on human-machine hybrid intelligence

Abstract

The present invention relates to an efficient training method for an intelligent acceleration platform model based on human-machine hybrid intelligence, comprising: 1. inputting original data into a training data set; 2. performing redundant processing on labeled data by loading an information reduction algorithm to form an optimized data set; 3. judging whether iterative training is finished, if finished, exiting the training, if not finished, entering step 4; 4. continuing neural network model training, evaluating information value by using an information measurement algorithm, sorting information value according to the evaluation result, extracting data ranked before the budget scale according to the budget scale, and then manually marking the data, adding the manually marked data to the training data set for subsequent iterative training, and discarding the data ranked after the budget scale; by continuously circulating the loading information reduction algorithm and the information measurement algorithm until the set number of training iterations is reached, the purpose of efficient training is achieved through value sorting and redundancy deletion.

Description

Technical Field

[0001] The present invention belongs to the technical field of efficient training methods for intelligent acceleration platform models, and specifically relates to an efficient training method for intelligent acceleration platform models based on human-machine hybrid intelligence. Background Art

[0002] In recent years, with the rapid development of computer hardware and artificial intelligence technology, artificial intelligence recognition algorithms have enabled multiple fields such as industry and agriculture. The current algorithm mainly relies on deep learning as the main technical implementation method. However, current deep learning relies heavily on a large amount of labeled training data, which seriously consumes a lot of financial and material resources. On the one hand, in general fields, images and other data can be easily obtained through the Internet, and the labeling is relatively simple, but the data is huge and the redundancy is high, which will waste labeling time and increase the pressure on training resources and electricity consumption. On the other hand, in special fields such as medical care, the labeling of image data requires professionals, which is extremely costly. If the data is directly labeled without cleaning, it will waste the valuable time of professionals.

[0003] NVIDIA has become the core promoter of the AI ​​revolution by virtue of its efficient application of high-performance computing chips in artificial intelligence computing. Although the research and development of artificial intelligence computing chips in China started late, a number of representative companies have emerged, including Huawei Ascend, Tianshu Zhixin, Cambrian, etc. These domestic and foreign companies have launched their dedicated intelligent acceleration platforms designed for accelerating deep learning algorithms. Their ability to efficiently handle a large number of parallel calculations has greatly improved the performance of AI applications. The present invention can be deployed on an intelligent acceleration platform and cooperate with computing chips to improve the efficiency of model training.

[0004] Therefore, we develop an efficient training method for the intelligent acceleration platform model, carry out data information measurement through a human-machine hybrid approach, only annotate valuable data, perform efficient data completion, and then reduce information reduction to reduce duplicate data. This can reduce the model's demand for data, speed up training, improve training efficiency, and reduce the demand for computing resources, thereby achieving efficient model training on the intelligent acceleration platform.

[0005] In terms of data analysis and efficient model training, in 2017, Shen et al. from Cardiff University used Fisher discriminant analysis to extract the boundaries of sparse and dense areas in clusters based on clustering algorithms, and discarded dense areas to evaluate and screen data; in 2018, Sener et al. from Intel Research Institute proposed an active learning algorithm based on feature space coverage to select different data as much as possible; in 2019, Yoo et al. from Korea University conducted research on model loss functions and proposed a data screening method for learning data loss; in 2020, Jordan et al. from Princeton University analyzed the diversity and uncertainty of gradient feature vectors and proposed a data evaluation method for gradient space; in 2021, Razvan et al. from Imperial College London extracted features based on graph convolutional networks, with nodes representing data features and edges representing data similarity, and queried high-information nodes through model uncertainty evaluation; in 2022, Amin et al. from the University of Adelaide proposed a feature mixed data evaluation method to select high-information data. In China, in 2017, relevant institutions constructed an image evaluation model based on mixed features and saliency detection. In 2018, Liu et al. from Xi'an University of Science and Technology used the overlap statistics method based on image matching to count the image overlap information at different intervals, thereby reducing the redundancy of aerial image data; in 2019, Yang et al. from Tianjin University proposed a reference-free screen image data evaluation method based on stacked autoencoders; in 2020, Zhang et al. from the University of Chinese Academy of Sciences introduced generative adversarial networks into data screening and evaluation to achieve low-redundancy training set screening; in 2021, Yuan et al. from the University of Chinese Academy of Sciences introduced support vectors into data screening and evaluation, and conducted research on multi-target data screening; in 2022, Yang et al. from Tianjin University proposed a dataset optimization evaluation method for intra-class differences and inter-class confusion.

[0006] Existing domestic and foreign research has achieved certain results in efficient model training, but has not analyzed the data from the perspective of information value, information redundancy, and in-depth analysis of the overall distribution of data sets and data complexity. In the current situation, studying efficient training strategies and cleaning data from multiple levels are of great significance to the development of artificial intelligence. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention proposes an efficient training method for an intelligent acceleration platform model based on human-machine hybrid intelligence.

[0008] The above-mentioned object of the present invention is achieved by the following technical solutions:

[0009] An efficient training method for an intelligent acceleration platform model based on human-machine hybrid intelligence, characterized in that: the method is based on a constructed intelligent acceleration hardware platform, the constructed intelligent acceleration hardware platform is equipped with an intelligent acceleration card, an algorithm software package is installed on the host operating system of the platform, the algorithm software package includes an information metric algorithm, a data information reduction algorithm and a data efficient completion algorithm based on the information metric algorithm, and the method includes the following steps:

[0010] Step 1: Start iterative training and input the original data into the training data set;

[0011] Step 2: Use the information reduction algorithm to perform redundant processing on the labeled data in the training data set, eliminate the redundant data, and form an optimized data set;

[0012] Step 3: Determine whether the iterative training is finished. If so, exit the training. If not, proceed to step 4.

[0013] Step 4: Continue training the neural network model and use an efficient data completion algorithm to process the unlabeled data, including: evaluating the information value through an information measurement algorithm, sorting the information value according to the evaluation results, extracting the data ranked before the budget scale according to the budget scale, and then manually labeling the manually labeled data, adding the manually labeled data to the training data set, and then processing redundant data through the loading information reduction algorithm, and performing subsequent iterative training. At the same time, discarding the data ranked after the budget scale; through continuous cycles of the loading information reduction algorithm and the information measurement algorithm, until the set number of training iterations is reached, the efficient training of the intelligent acceleration platform model based on human-machine hybrid intelligence is completed.

[0014] Moreover, in step 2, the data in the training data set includes labeled data sets and unlabeled data sets, and the labeled data sets are denoted as The unlabeled dataset is denoted as The steps include:

[0015] 2.1 The data in is mapped to a high-dimensional space and trained with θ: {θ e ,θ c} is a model with parameters F(X, θ) = F e (X,θ e )+F c (Z, θ c ), where F e (X,θ e ): is the backbone feature extraction network, F c (Z, θ c ): is a classifier; the input data is mapped to a high-dimensional feature space by the backbone feature extraction network to obtain a spatial representation:

[0016] Z=F e (X;θ e )

[0017] The spatial representation Z is mapped to the final output by the classifier, and the output is converted into a probability distribution p(Y|X; θ c )=softmax(Fc(Z,θ c )); Update the parameters to optimize the model by minimizing the cross entropy loss H(q, p) = -∑X(q(Y)logp(Y)); When the model optimization is completed, use F e extract and Spatial representation of data in:

[0018]

[0019] 2.2, through the feature clusterer, The unlabeled data in the dataset form k unlabeled data clusters:

[0020]

[0021] At the same time, The labeled data in form k labeled data clusters:

[0022]

[0023] 2.3. For each unlabeled data cluster and labeled data cluster, the redundancy evaluator is used to evaluate the data redundancy degree; for the unlabeled data cluster, the category prototype entropy evaluation method is first used to obtain the information evaluation score S of each data in the unlabeled data cluster, and then the scores are sorted from large to small to obtain the candidate set:

[0024] {x1: S1, x2: S2, ..., x n :S n}

[0025] Where n is the total number of data in the cluster;

[0026] 2.4. Calculate the core distance, including:

[0027] First, set x1 to the core data c1, and make a radius with the core data as the center. Neighborhood; calculate the core data c1 to the rest of x i The core distance is smaller than the radius. Then x iFor highly redundant information data, it is manually eliminated from the candidate set; then new core data is selected from the updated candidate set and calculated with the rest x i Repeat the above operation for the core distance. The core distance calculation method is as follows:

[0028]

[0029] Among them, F e (c i ) represents the core features of the core data in high-dimensional space, F e (x j ) is the high-dimensional feature of the remaining candidate set data; when the number of core data reaches the required budget scale After that, all the core data are manually labeled and added to For the labeled data cluster, randomly select the core data, use the methods in steps 2.3 and 2.4 to calculate the core distance, and eliminate the high information redundant data.

[0030] Furthermore, step 4 includes:

[0031] 4.1. Use the existing labeled data in the database to train the feature extractor. f Extract the data x to be evaluated u The eigenvector e u and the data of category i in the database The eigenvector of And according to the category Find the average value and get the category cluster center p of the i-th category data in the database i ,in:

[0032] e u =E f (x u )

[0033]

[0034] 4.2 Calculate the category prototype entropy, including:

[0035] 4.2.1 Calculate the data x to be evaluated u The eigenvector e u The cluster centers p of each category i The Euclidean distance between them is obtained to obtain the distance vector l = [l1, l2, ..., l C ]:

[0036]

[0037] 4.2.2 Select the two smallest components in the distance vector l, which correspond to the data x to be evaluated u The cluster centers of the two nearest categories in the feature space; select the cluster centers of the two categories that are closest to the data to be evaluated x u The k most recent data in the feature space Using its feature vector Calculate the edge data category cluster center of these k data

[0038]

[0039] Where k is the total number of data in this category multiplied by a custom parameter α, which is set between 0 and 1;

[0040] 4.2.3 Calculate the data x to be evaluated u and the edge data category cluster center of the two closest classes The Euclidean distance d between i for:

[0041]

[0042] 4.2.4 Using the softmax function to calculate the distance vector d i Perform normalization to obtain the category probability vector p = [p1, p2] of the data:

[0043]

[0044] 4.2.5 Calculate the entropy of the data category probability vector P and obtain the data to be evaluated x j The category prototype entropy E(x u ):

[0045]

[0046] 4.2.6 Using the data to be evaluated x u The category prototype entropy E(x u ) measures the information evaluation score S of the data; E(x u ) is larger, the closer the data is to the decision boundary of the two most similar classes, and it is difficult to determine which class the data belongs to, and the larger the score S is; conversely, E(x u ) is smaller, indicating that the data is closer to the category center of a certain category, the higher the similarity with the data in the database, and the smaller the score S.

[0047] The advantages and positive effects of the present invention are:

[0048] The present invention adopts an efficient completion algorithm. First, the information value of the data is calculated using an information measurement algorithm, and each data is scored according to the information value. The information value score S is used to manually select data with a large score S, and then manually labeled according to the required budget scale B. These unlabeled high-information data are converted into labeled data and added to the training set, thereby realizing high-value data screening and labeling; and in order to reduce the redundancy of the data set, the present invention proposes an information reduction algorithm, which measures the degree of information similarity between data. Relying on the data space representation, an unsupervised close-range elimination method is proposed to realize the manual elimination of information redundant data in high-dimensional space. In summary, the present invention achieves the purpose of efficient training through value sorting and redundancy deletion. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the efficient training method of the intelligent acceleration platform model based on human-machine hybrid intelligence of the present invention;

[0050] Figure 2 It is a flow chart of the efficient data completion algorithm implemented by the present invention. DETAILED DESCRIPTION

[0051] The structure of the present invention is further described below with reference to the accompanying drawings and by way of examples. It should be noted that the present examples are descriptive rather than restrictive.

[0052] An efficient training method for intelligent acceleration platform models based on human-machine hybrid intelligence, see Figure 1-Figure 2 The method is based on an intelligent acceleration hardware platform built by installing an intelligent acceleration card. The relevant drivers are installed on the host operating system of the platform, the environment is built, the software packages that the algorithm depends on are installed, and then the algorithm source code is transplanted to the platform. The algorithm is implemented in conjunction with specific artificial intelligence vision tasks. The algorithm includes an information metric algorithm, a data information reduction algorithm, and a data efficient completion algorithm based on the information metric algorithm. The data efficient completion algorithm can realize the evaluation and storage of high-value data. The data information reduction algorithm can remove redundant information in existing data. Finally, the algorithm is deployed on an intelligent platform to achieve efficient training.

[0053] The complete process of the efficient training method of the present invention is shown in the attached Figure 1 As shown, the following steps are included:

[0054] Step 1: Start iterative training and input the original data into the training data set;

[0055] Step 2: Use the information reduction algorithm to perform redundant processing on the labeled data in the training data set, eliminate the redundant data, and form an optimized data set;

[0056] Step 3: Determine whether the iterative training is finished. If so, exit the training. If not, proceed to step 4.

[0057] Step 4: Continue training the neural network model and use an efficient data completion algorithm to process the unlabeled data, including: evaluating the information value through an information measurement algorithm, sorting the information value according to the evaluation results, extracting the data ranked before the budget scale according to the budget scale, and then manually labeling the manually labeled data, adding the manually labeled data to the training data set, and then processing redundant data through the loading information reduction algorithm, and performing subsequent iterative training. At the same time, discarding the data ranked after the budget scale; through continuous cycles of the loading information reduction algorithm and the information measurement algorithm, until the set number of training iterations is reached, the efficient training of the intelligent acceleration platform model based on human-machine hybrid intelligence is completed.

[0058] Efficient data completion algorithm: as attached Figure 2 As shown, the efficient completion algorithm first uses the information measurement algorithm to calculate the information value of the data, scores each data according to the information value, and uses the information value score S to manually select data with a large score S. According to the required budget scale B, manual labeling is performed to convert these unlabeled high-information data into labeled data and add them to the training set.

[0059] Information measurement algorithm: The process of completing the data set requires manual addition of more high-value data. Therefore, the key to completing the data lies in the evaluation of the value of data information. To address this problem, an information measurement algorithm is proposed to study the similarity between feature space data and each category to measure the difficulty of data discrimination. Relying on the data space representation, a method is proposed to find data on the decision boundary of the category cluster, thereby evaluating the value of data information and ultimately achieving the goal of efficient completion of the data set under artificial hybrid intelligence.

[0060] In order to map unlabeled data to feature space, a feature extractor is needed. The feature extractor is trained using the labeled data in the database. f Extract the data x to be evaluated u The eigenvector e u and the data of category i in the database The eigenvector of And according to the category Find the average value and get the category cluster center p of the i-th category data in the database i ,in:

[0061] e u =E f (x u )

[0062]

[0063] Then calculate the category prototype entropy as follows: calculate the data x to be evaluated u The eigenvector e u The cluster centers p of each category i The Euclidean distance between them is obtained to obtain the distance vector l = [l1, l2, ..., l c ]:

[0064]

[0065] Select the two smallest components in the distance vector l, which correspond to the data x to be evaluated. u The cluster centers of the two closest categories in the feature space correspond to x u The two most likely categories. Select the data to be evaluated x from these two categories. u The k most recent data in the feature space Using its feature vector Calculate the edge data category cluster center of these k data

[0066]

[0067] Where k is the total number of data in this category multiplied by a custom parameter α, which is set between 0 and 1. Calculate the data to be evaluated x u and the edge data category cluster center of the two closest classes The Euclidean distance d between i :

[0068]

[0069] Use the softmax function to calculate the distance vector d i Perform normalization to obtain the category probability vector P = [p1, p2] of the data:

[0070]

[0071] Calculate the entropy of the data category probability vector P and obtain the data to be evaluated x j The category prototype entropy E(x u ):

[0072]

[0073] Using the data to be evaluated x u The category prototype entropy E(x u ) measures the information evaluation score S of the data. E(x u) is larger, the closer the data is to the decision boundary of the two most similar classes, and it is difficult to determine which class the data belongs to, and the larger the score S is; conversely, E(x u ), the smaller the data is, the closer it is to the category center of a certain category, the higher the similarity with the data in the database, and the smaller its score S is.

[0074] Information reduction algorithm: Efficiently completed data sets still have data redundancy problems. In order to reduce the redundancy of data sets, the present invention proposes an information reduction algorithm. Similar data have similar information distribution, which will produce information redundancy. This algorithm measures the degree of information similarity between data. Relying on the data space representation, an unsupervised close-range elimination method is proposed to achieve the artificial elimination of information redundant data in high-dimensional space, thereby achieving the simplification of data sets under human-machine hybrid intelligence.

[0075] The existing labeled dataset is denoted as The unlabeled dataset is denoted as First, you need to To map the data in to a high-dimensional space, we need to train with θ: e ,θ c} is a model with parameters F(X, θ) = F e (X,θ e )+F c (Z, θ c ), where F e (X,θ e ): is the backbone feature extraction network, F c (Z, θ c ): is a classifier. The input data is mapped to a high-dimensional feature space by the backbone feature extraction network to obtain a spatial representation:

[0076] Z=F e (X;θ e )

[0077] The spatial representation Z is mapped to the final output by the classifier, and the output is converted into a probability distribution p(Y|X; θ c )=softmax(F c (Z, θ c )). The model can be optimized by updating the parameters by minimizing the cross entropy loss H(q, p) = -∑X(q(Y)logp(Y)). When the model optimization is completed, use F e extract and Spatial representation of data in:

[0078]

[0079] The unsupervised close-range elimination method consists of two modules, a feature clusterer and a redundancy evaluator. The features of similar data are usually clustered in one area. The purpose of the feature clusterer is to differentiate the unlabeled data and form clusters in this area. Specifically, first randomly select Take a data, and make a neighborhood with radius R with this data as the center. Set β as the minimum number of points in the domain, sample the data in the neighborhood, and record the number k. If k is less than β, the selected data is called invalid data, and the next data is randomly selected to perform the above operation until a point n is greater than β is found. If n is greater than β, the selected data is called node data n. i . Define the data within the node data neighborhood as reachable data In the neighborhood of the node data, randomly select a new reachable data as the center, and do the same as above until the next node data is found. After multiple node data searches, there is no next node data. Multiple node data and reachable data in their neighborhood form a data cluster:

[0080]

[0081] After obtaining a data cluster, continue to select data outside the cluster and repeat the above operation to obtain multiple node data, thereby obtaining a new data cluster. Until all data are summarized in the cluster, the work of the feature clusterer is completed. Obtain multiple unlabeled data clusters:

[0082]

[0083] at the same time, The labeled data in can naturally form k labeled data clusters:

[0084]

[0085] For each unlabeled data cluster and labeled data cluster, we use the redundancy evaluator to evaluate the degree of data redundancy. For the unlabeled data cluster, we first use the category prototype entropy evaluation method to obtain the information evaluation score S of each data in the unlabeled data cluster, and sort them from large to small to obtain the candidate set:

[0086] {x1: S1, x2: S2, ..., x n :S n}

[0087] Here n is the total number of data in the cluster.

[0088] First, set x1 to the core data c1, and make a radius with the core data as the center. Calculate the core data c1 to the remaining xi The core distance is smaller than the radius. Then x i For highly redundant information data, it is manually eliminated from the candidate set. Next, new core data is selected from the updated candidate set and calculated with the rest of x i Repeat the above operation to find the core distance. The core distance calculation method is as follows:

[0089]

[0090] Among them, F e (c i ) represents the core features of the core data in high-dimensional space, F e (x j ) is the high-dimensional feature of the remaining candidate set data. When the number of core data reaches the required budget scale After that, all the core data are manually labeled and added to For the labeled data cluster, core data are randomly selected, similar to the above method, the core distance is calculated, and the high information redundant data is eliminated.

[0091] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. An efficient training method for an intelligent acceleration platform model based on human-machine hybrid intelligence, characterized by: The method is based on a constructed intelligent acceleration hardware platform, wherein the constructed intelligent acceleration hardware platform is installed with an intelligent acceleration card, and an algorithm software package is installed on the host operating system of the platform, and the algorithm software package includes an information metric algorithm, a data information reduction algorithm, and a data efficient completion algorithm based on the information metric algorithm. The method comprises the following steps: Step 1: Start iterative training and input the original data into the training data set; Step 2: Use the information reduction algorithm to perform redundant processing on the labeled data in the training data set, eliminate the redundant data, and form an optimized data set; Step 3: Determine whether the iterative training is finished. If so, exit the training. If not, proceed to step 4. Step 4: Continue training the neural network model and use an efficient data completion algorithm to process the unlabeled data, including: evaluating the information value through an information measurement algorithm, sorting the information value according to the evaluation results, extracting the data ranked before the budget scale according to the budget scale, and then manually labeling the manually labeled data, adding the manually labeled data to the training data set, and then processing redundant data through the loading information reduction algorithm, and performing subsequent iterative training. At the same time, discarding the data ranked after the budget scale; through continuous cycles of the loading information reduction algorithm and the information measurement algorithm, until the set number of training iterations is reached, the efficient training of the intelligent acceleration platform model based on human-machine hybrid intelligence is completed.

2. The efficient training method of intelligent acceleration platform model based on human-machine hybrid intelligence according to claim 1 is characterized in that: In step 2, the data in the training dataset includes labeled datasets and unlabeled datasets. The labeled dataset is denoted as The unlabeled dataset is denoted as The steps include: 2.1 The data in is mapped to a high-dimensional space and trained with θ: {θ e ,θ c } is a model with parameters F(X, θ) = F e (X,θ e )+F c (Z, θ c ),in is the backbone feature extraction network, is a classifier; the input data is mapped to a high-dimensional feature space by the backbone feature extraction network to obtain a spatial representation: Z=F e (X;θ e ) The spatial representation Z is mapped to the final output by the classifier, and the output is converted into a probability distribution p(Y|X; θ c )=softmax(F e (Z, θ c )); Update the parameters to optimize the model by minimizing the cross entropy loss H(q, p) = -∑X(q(Y)logp(Y)); When the model optimization is completed, use F e extract and Spatial representation of data in: 2.2, through the feature clusterer, The unlabeled data in the dataset form k unlabeled data clusters: At the same time, The labeled data in form k labeled data clusters: 2.

3. For each unlabeled data cluster and labeled data cluster, the redundancy evaluator is used to evaluate the data redundancy degree. For the unlabeled data cluster, the category prototype entropy evaluation method is first used to obtain the information evaluation score S of each data in the unlabeled data cluster, and the scores are sorted from large to small to obtain the candidate set: {x1:S1,x2:S2,...,x n :S n } Where n is the total number of data in the cluster; 2.

4. Calculate the core distance, including: First, set x1 to the core data c1, and make a radius with the core data as the center. Neighborhood; calculate the core data c1 to the rest of x i The core distance is smaller than the radius. Then x i For highly redundant information data, it is manually eliminated from the candidate set; then new core data is selected from the updated candidate set and calculated with the rest x i Repeat the above operation for the core distance. The core distance calculation method is as follows: Among them, F e (c i ) represents the core features of the core data in high-dimensional space, F e (x j ) is the high-dimensional feature of the remaining candidate set data; when the number of core data reaches the required budget scale After that, all the core data are manually labeled and added to For the labeled data cluster, randomly select the core data, use the methods in steps 2.3 and 2.4 to calculate the core distance, and eliminate the high information redundant data.

3. The efficient training method of intelligent acceleration platform model based on human-machine hybrid intelligence according to claim 1 is characterized in that: Step 4 includes: 4.

1. Use the existing labeled data in the database to train the feature extractor. f Extract the data x to be evaluated u The eigenvector e u and the data of category i in the database The eigenvector of And according to the category Find the average value and get the category cluster center p of the i-th category data in the database i ,in: And u =And f (x u ) 4.2 Calculate the category prototype entropy, including: 4.2.1 Calculate the data x to be evaluated u The eigenvector e u The cluster centers p of each category i The Euclidean distance between them is obtained to obtain the distance vector l = [l1, l2, ..., l c ]: 4.2.2 Select the two smallest components in the distance vector l, which correspond to the data x to be evaluated u The cluster centers of the two nearest categories in the feature space; select the cluster centers of the two categories that are closest to the data to be evaluated x u The k most recent data in the feature space Using its feature vector Calculate the edge data category cluster center of these k data Where k is the total number of data in this category multiplied by a custom parameter α, which is set between 0 and 1; 4.2.3 Calculate the evaluation data x u and the edge data category cluster center of the two closest classes The Euclidean distance d between i for: 4.2.4 Using the softmax function to calculate the distance vector d i Perform normalization to obtain the category probability vector P = [p1, p2] of the data: 4.2.5 Calculate the entropy of the data category probability vector P and obtain the data to be evaluated x j The category prototype entropy E(x u ): 4.2.6 Using the data to be evaluated x u The category prototype entropy E(x u ) measures the information evaluation score S of the data; E(x u ) is larger, the closer the data is to the decision boundary of the two most similar classes, and it is difficult to determine which class the data belongs to, and the larger the score S is; conversely, E(x u ) is smaller, indicating that the data is closer to the category center of a certain category, the higher the similarity with the data in the database, and the smaller the score S.

Images