Model optimization method, device and computer-readable storage medium based on machine learning

By adjusting the loss function value and updating the noise probability value according to the label noise probability value in model training, the model overfitting problem caused by label noise samples is solved, and the generalization ability of the model is improved.

CN114548257BActive Publication Date: 2024-11-08CHINA UNIONPAY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210150801.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-14
Publication Date
2024-11-08
Estimated Expiration
2042-02-14

AI Technical Summary

Technical Problem

During model training, the presence of label noise samples leads to overfitting of the model, and it is difficult for the prior art to effectively identify and reduce its impact.

Method used

Through multiple model iteration operations, the loss function value is adjusted according to the label noise probability value of the sample, the weight of the label noise sample is reduced, and the label noise probability value is updated using unsupervised learning to prevent the model from being overfitted.

Benefits of technology

Effectively identify and reduce the impact of label noise samples on model training, prevent overfitting, and improve the generalization ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114548257B_ABST
    Figure CN114548257B_ABST
Patent Text Reader

Abstract

The present disclosure provides a model optimization method, device and computer-readable storage medium based on machine learning. The method includes: multiple model iteration operations, at least one model iteration operation in the multiple model iteration operations includes: training the model according to multiple samples, sample labels, and the probability value of the sample being a label noise sample to obtain a first model, wherein the first loss function value of the single sample is determined according to the label prediction value corresponding to the single sample, the label of the single sample, and the probability value of the single sample being a label noise sample, and the training of the model is guided according to the first loss function values ​​of the multiple samples; the multiple samples are input into the first model to obtain the label prediction value of each sample, and the second loss function value of each sample is determined according to the label prediction value of each sample and the label of each sample; the probability value of each sample being a label noise sample is updated according to the second loss function value of the multiple samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure belongs to the field of artificial intelligence, and specifically relates to a model optimization method, device and computer-readable storage medium based on machine learning. Background Art

[0002] This section is intended to provide a background or context for embodiments of the present disclosure. No description herein is admitted to be prior art by virtue of its inclusion in this section.

[0003] The accuracy of sample labels is crucial for model training. However, it is impossible to completely avoid label errors in some samples. In this disclosure, samples with incorrect labels are called label noise samples, and samples with correct labels are called label normal samples. During model training, it is necessary to identify label noise samples and reduce the impact of label noise samples on model training to prevent the model from overfitting these label noise samples. Summary of the invention

[0004] The present invention provides a model optimization method, device and computer-readable storage medium based on machine learning.

[0005] The following scheme is provided in an embodiment of the present disclosure: a model optimization method based on machine learning, the model optimization method comprising multiple model iteration operations, at least one model iteration operation of the multiple model iteration operations comprising:

[0006] The model is trained according to multiple samples, labels of each sample, and probability values ​​of each sample being a label noise sample to obtain a first model, wherein a first loss function value of the single sample is determined according to a label prediction value corresponding to the single sample, the label of the single sample, and the probability value of the single sample being a label noise sample, and the training of the model is guided according to the first loss function values ​​of the multiple samples;

[0007] Input the multiple samples into the first model to obtain a label prediction value of each sample, and determine a second loss function value of each sample according to the label prediction value of each sample and the label of each sample;

[0008] The probability value of each sample being a label noise sample is updated according to the second loss function value of the multiple samples.

[0009] The following scheme is provided in an embodiment of the present disclosure: A model optimization device based on machine learning includes: a model iteration module, which is used to perform multiple model iteration operations, at least one of the multiple model iteration operations includes:

[0010] The model is trained according to multiple samples, labels of each sample, and probability values ​​of each sample being a label noise sample to obtain a first model, wherein a first loss function value of the single sample is determined according to a label prediction value corresponding to the single sample, the label of the single sample, and the probability value of the single sample being a label noise sample, and the training of the model is guided according to the first loss function values ​​of the multiple samples;

[0011] Input the multiple samples into the first model to obtain a label prediction value of each sample, and determine a second loss function value of each sample according to the label prediction value of each sample and the label of each sample;

[0012] The probability value of each sample being a label noise sample is updated according to the second loss function value of the multiple samples.

[0013] The following scheme is provided in an embodiment of the present disclosure: a model optimization device based on machine learning, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the aforementioned model optimization method based on machine learning.

[0014] The following solution is provided in an embodiment of the present disclosure: a computer-readable storage medium storing a program, which, when executed by a processor, enables the processor to execute the aforementioned model optimization method based on machine learning.

[0015] The probability of incorrect labeling of each sample is analyzed through unsupervised learning. During the model training process, the loss function value is corrected by this probability value, and the weights of samples with a higher probability of incorrect labeling are intentionally reduced, thereby avoiding overfitting of the model to samples with a higher probability of incorrect labeling.

[0016] It should be understood that the above description is only an overview of the technical solution of the present disclosure, so that the technical means of the present disclosure can be more clearly understood and implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present disclosure more obvious and easy to understand, the specific implementation methods of the present disclosure are specifically described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The present invention is a flowchart of a model iteration operation in a model optimization method based on machine learning according to an embodiment of the present invention.

[0018] Figure 2 A schematic diagram of the distribution of the second loss function values ​​of samples obtained by the model optimization method according to an embodiment of the present disclosure.

[0019] Figure 3 This is a variation trend graph of the second loss function value of the sample obtained by the model optimization method according to an embodiment of the present disclosure.

[0020] Figure 4 It is a complete flowchart of the model optimization method based on machine learning according to an embodiment of the present disclosure.

[0021] Figure 5 It is a structural schematic diagram of a model optimization device based on machine learning according to an embodiment of the present disclosure.

[0022] Figure 6 It is a structural schematic diagram of a model optimization device based on machine learning according to another embodiment of the present disclosure. DETAILED DESCRIPTION

[0023] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0024] The embodiments of the present disclosure provide a model optimization method based on machine learning. The model is, for example, a neural network model or other machine learning model. The application scenarios of the model can be known machine learning application scenarios such as speech recognition, image recognition, and decision-making. From the perspective of the device, the executor of the model optimization method can be one or more electronic devices, more specifically, the processing modules therein; from the perspective of the program, the executor of the model optimization method can be the programs installed on these electronic devices.

[0025] The model optimization method includes multiple model iteration operations, at least one of the multiple model iteration operations includes Figure 1 Steps 101 to 103 in the flowchart shown.

[0026] In step 101, model training is performed based on multiple samples, labels of each sample, and probability values ​​of each sample being a label noise sample to obtain a first model, wherein a first loss function value of the single sample is determined based on the label prediction value corresponding to the single sample, the label of the single sample, and the probability value of the single sample being a label noise sample, and the training of the model is guided by the first loss function values ​​of the multiple samples.

[0027] Each sample can be a multi-dimensional vector. The label of the sample can be 1 or 0. The probability value of each sample being a label noise sample is also between 0 and 1.

[0028] Model training refers to multiple rounds of iterations of the parameters or hyperparameters of the model so that the sum of the first loss function values ​​of the multiple samples converges to a sufficiently small value. The present disclosure does not limit how to update the parameters or hyperparameters of the model according to the sum of the first loss function values. For example, conventional gradient descent or stochastic gradient descent can be used.

[0029] Since the first loss function value contains the information about the probability value of the sample being a label noise sample, for samples with relatively large probability values, the weights of these samples can be appropriately reduced during the model training process to prevent overfitting of these samples during the model training process.

[0030] In step 102, the multiple samples are input into the first model to obtain a label prediction value of each sample, and a second loss function value of each sample is determined according to the label prediction value of each sample and the label of each sample.

[0031] The second loss function value represents the actual difference between the sample's label prediction value and its true value. In this step, it is not important whether the label is correctly labeled. Since the weights of samples with a higher probability of label errors are intentionally reduced in step 101, the second loss function values ​​of these samples tend to increase as the number of model iterations increases.

[0032] In step 103, the probability value of each sample being a label noise sample is updated according to the second loss function value of the multiple samples.

[0033] This step analyzes the probability value of each sample's label annotation error through unsupervised learning, and the probability value of each sample being a label noise sample in step 101 in the next round of model iteration operation is updated.

[0034] It should be noted that the number of samples in the next round of model iteration operation can be the same as the number of samples in the current round of model iteration operation. If the probability value of a certain sample being a label noise sample is large enough, the sample can also be directly eliminated in the next round of model fall operation.

[0035] In this disclosure, for the sake of distinction, the iterative optimization process of the parameters or hyperparameters of the model in step 101 is referred to as "model training", and a round of operations including steps 101 to 103 is referred to as "model iteration operation". "Model iteration operation" includes "model optimization operation" and the operation of updating the probability value of a sample being a label noise sample.

[0036] The following introduces several methods for calculating the value of the first loss function.

[0037] In some implementations, the first loss function value of the sample numbered i among the multiple samples is calculated according to the following formula:

[0038]

[0039] in, is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample;

[0040] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0041] In some implementations, the first loss of the sample numbered i among the multiple samples is calculated according to the following formula:

[0042]

[0043] in, is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample;

[0044] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0045] In some implementations, the first loss of the sample numbered i among the multiple samples is calculated according to the following formula:

[0046]

[0047] in, is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i is the probability value of the sample numbered i being a label noise sample, h i is the probability value that the sample numbered i is a positive sample;

[0048] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0049] In some implementations, the second loss function value of the sample numbered i among the multiple samples is calculated according to the following formula:

[0050]

[0051] in, is the second loss function value of sample number i, y i is the label of the sample numbered i, h i is the probability value that the sample numbered i is a positive sample;

[0052] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0053] The second loss function value reflects the actual difference between the predicted value of the sample and the label. The first loss function value corrects the difference between the predicted value of the sample and the label based on the probability that the sample is a label noise sample.

[0054] Figure 3 These are experimental data obtained by the inventors of the present disclosure to verify the above method.

[0055] As the number of “model iteration operations” increases, the second loss function values ​​of some samples tend to increase, and these samples are most likely label noise samples. Figure 3 The curve in the upper middle part shows the changing trend of the average value of the second loss function of this type of samples.

[0056] As the number of “model iteration operations” increases, the second loss function values ​​of some samples tend to decrease, and these samples are likely to be samples with correct labels. Figure 3 The curve in the lower middle part shows the changing trend of the average value of the second loss function of this type of samples.

[0057] In some implementations, updating the probability value of each sample being a label noise sample according to the second loss function value of the multiple samples includes:

[0058] The probability value of a sample being a label noise sample, including:

[0059] Using the EM algorithm to estimate the parameters and mixing coefficients of the first distribution and the second distribution obeyed by the multiple samples, wherein the first distribution is the distribution of the probability values ​​of the labeled normal samples, and the second distribution is the distribution of the probability values ​​of the labeled noise samples;

[0060] The probability value of each sample being a label noise sample is updated according to the second loss function value of each sample and the parameter and mixing coefficient of the second distribution.

[0061] The EM algorithm is an iterative optimization strategy. Since each iteration in its calculation method is divided into two steps, one is the expectation step (E step) and the other is the maximum step (M step), the algorithm is called the EM algorithm (Expectation-Maximization Algorithm).

[0062] The first distribution and the second distribution are the same type of distribution. For example, both are Gaussian distributions or both are Beta distributions. The EM algorithm corresponding to the former can be called an EM algorithm based on Gaussian mixture distribution, and the EM algorithm corresponding to the latter can be called an EM algorithm based on Beta mixture distribution.

[0063] Figure 2 The bar graph in the figure shows the true distribution of the loss function value, the solid line shows the probability density diagram of the Gaussian mixture distribution obtained by fitting the EM algorithm, and the dotted line shows the probability density diagram of the Beta mixture distribution obtained by fitting the EM algorithm. The bars above the horizontal dividing line in the bar graph represent the distribution of the labeled noise samples, and the bars below the horizontal dividing line in the bar graph represent the distribution of the labeled normal samples. For ease of display, the loss function value is normalized to the range of [0,1] according to a certain algorithm.

[0064] The expression of Beta distribution is as follows:

[0065]

[0066] Where α and β are the parameters of the Beta distribution, Γ(·) is the Beta function, is the second loss function value, The second loss function value is The sample belongs to the probability density function of the Beta distribution with parameters α and β.

[0067] The expression for the Beta mixture distribution is as follows:

[0068]

[0069] Wherein, k is the number of the beta distribution, and K is the number of beta distributions. In the embodiment of the present disclosure, K=2. k The second loss function value is The probability that a sample of belongs to the kth Beta distribution (also called the mixing coefficient), is the second loss function value The probability density function of .

[0070] After calculating the second loss function value of each sample, the parameters of the two Beta distributions and the mixing coefficient are estimated by the EM algorithm.

[0071] In step E, based on the current mixing coefficient λ kand the coefficient α of each current beta distribution k and β k Calculate the second loss function value as The probability value of a sample belonging to each Beta distribution The calculation formula is as follows:

[0072]

[0073] Among them, k=0 corresponds to the labeled normal sample, and k=1 corresponds to the labeled noise sample.

[0074] In the M step, the probability value of the sample belonging to each Beta distribution according to the second loss function value is Re-estimate the mixing coefficients and parameters of each beta distribution. The calculation formula is as follows:

[0075]

[0076]

[0077] in,

[0078] Then, based on the fitted Beta mixture distribution, the probability of each sample belonging to the kth Beta distribution is updated:

[0079]

[0080] In some embodiments, in combination Figure 4 The method further includes: stopping the model iteration operation when the number of model iteration operations reaches a set threshold; or stopping the model iteration operation when the second loss function values ​​of all samples converge.

[0081] That is, it provides an exit mechanism for model iteration operations.

[0082] Based on the same technical concept, the embodiment of the present disclosure also provides a model optimization device based on machine learning, which is used to execute the model optimization method based on machine learning provided in any of the above embodiments. Specifically, the model optimization device based on machine learning can be implemented in hardware forms such as application-specific integrated circuits (ASICs) and graphics processing units (GPUs).

[0083] refer to Figure 5 The machine learning-based model optimization device includes: a model iteration module 1, which is used to perform multiple model iteration operations, at least one of the multiple model iteration operations includes:

[0084] The model is trained according to multiple samples, labels of each sample, and probability values ​​of each sample being a label noise sample to obtain a first model, wherein a first loss function value of the single sample is determined according to a label prediction value corresponding to the single sample, the label of the single sample, and the probability value of the single sample being a label noise sample, and the training of the model is guided according to the first loss function values ​​of the multiple samples;

[0085] Input the multiple samples into the first model to obtain a label prediction value of each sample, and determine a second loss function value of each sample according to the label prediction value of each sample and the label of each sample;

[0086] The probability value of each sample being a label noise sample is updated according to the second loss function value of the multiple samples.

[0087] In some implementations, the first loss function value of the sample numbered i among the multiple samples is calculated according to the following formula:

[0088]

[0089] in, is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample;

[0090] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0091] In some implementations, the first loss of the sample numbered i among the multiple samples is calculated according to the following formula:

[0092]

[0093] in, is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample;

[0094] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0095] In some implementations, the first loss of the sample numbered i among the multiple samples is calculated according to the following formula:

[0096]

[0097] in, is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i is the probability value of the sample numbered i being a label noise sample, h i is the probability value that the sample numbered i is a positive sample;

[0098] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0099] In some implementations, the second loss function value of the sample numbered i among the multiple samples is calculated according to the following formula:

[0100]

[0101] in, is the second loss function value of sample number i, y i is the label of the sample numbered i, h i is the probability value that the sample numbered i is a positive sample;

[0102] Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

[0103] In some implementations, updating the probability value of each sample being a label noise sample according to the second loss function value of the multiple samples includes:

[0104] Using the EM algorithm to estimate the parameters and mixing coefficients of the first distribution and the second distribution obeyed by the multiple samples, wherein the first distribution is the distribution of the probability values ​​of the labeled normal samples, and the second distribution is the distribution of the probability values ​​of the labeled noise samples;

[0105] The probability value of each sample being a label noise sample is updated according to the second loss function value of each sample and the parameter and mixing coefficient of the second distribution.

[0106] In some embodiments, the device also includes a stopping module 2, which is used to stop the model iteration operation when the number of model iteration operations reaches a set threshold; or, to stop the model iteration operation when the second loss function values ​​of all samples converge.

[0107] refer to Figure 6An embodiment of the present disclosure also provides a model optimization device based on machine learning, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the aforementioned model optimization method based on machine learning.

[0108] An embodiment of the present disclosure also provides a computer-readable storage medium, which stores a program. When the program is executed by a processor, the processor executes the aforementioned model optimization method based on machine learning.

[0109] Each embodiment in the present disclosure is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and computer-readable storage medium embodiments, since they are basically similar to the method embodiments, their descriptions are simplified, and the relevant parts can be referred to the partial description of the method embodiments.

[0110] The device and computer-readable storage medium provided in the embodiments of the present disclosure correspond one-to-one to the method. Therefore, the device and computer-readable storage medium also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the device and computer-readable storage medium will not be repeated here.

[0111] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, apparatuses, and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0113] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0114] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0115] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0116] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0117] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. In addition, although the operations of the disclosed method are described in a particular order in the accompanying drawings, this does not require or imply that these operations must be performed in this particular order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step, and / or one step can be decomposed into multiple steps.

[0118] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A model optimization method based on machine learning, characterized in that: The model optimization method includes multiple model iteration operations, and at least one model iteration operation among the multiple model iteration operations includes: The model is trained according to multiple samples for speech recognition or image recognition, labels of each sample, and probability values ​​of each sample being a label noise sample to obtain a first model, wherein a first loss function value of a single sample is determined according to a label prediction value corresponding to the single sample, the label of the single sample, and the probability value of the single sample being a label noise sample, and the training of the model is guided according to the first loss function values ​​of the multiple samples; Input the multiple samples into the first model to obtain a label prediction value of each sample, and determine a second loss function value of each sample according to the label prediction value of each sample and the label of each sample; The probability value of each sample being a label noise sample is updated according to the second loss function value of the multiple samples.

2. The method according to claim 1, characterized in that The first loss function value of the sample numbered i among the multiple samples is calculated according to the following formula: fl i =-[(1-p i )y i +p i z i ]log(h i )-[(1-p i )(1-y i )+p i (1-z i )]log(1-h i ), Among them, fl i is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

3. The method according to claim 1, characterized in that: The first loss of the sample numbered i among the multiple samples is calculated according to the following formula: fl i =|(1-p i )y i +p i z i -h i | Among them, fl i is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

4. The method according to claim 1, characterized in that: The first loss of the sample numbered i among the multiple samples is calculated according to the following formula: fl i =[(1-p i )y i +p i z i -h i ] 2 , Among them, fl i is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i is the probability value of the sample numbered i being a label noise sample, h i is the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

5. The method according to claim 1, characterized in that The second loss function value of the sample numbered i among the multiple samples is calculated according to the following formula: sl i =-[y i log(h i )+(1-y i )log(1-h i )], Among them, sl i is the second loss function value of sample number i, y i is the label of the sample numbered i, h i is the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

6. The method according to claim 1, characterized in that Updating the probability value of each sample being a label noise sample according to the second loss function value of the multiple samples includes: Using the EM algorithm to estimate the parameters and mixing coefficients of the first distribution and the second distribution obeyed by the multiple samples, wherein the first distribution is the distribution of the probability values ​​of the labeled normal samples, and the second distribution is the distribution of the probability values ​​of the labeled noise samples; The probability value of each sample being a label noise sample is updated according to the second loss function value of each sample and the parameter and mixing coefficient of the second distribution.

7. The method according to claim 1, characterized in that The method further comprises: When the number of model iteration operations reaches a set threshold, the model iteration operation is stopped; or, The model iteration operation is stopped when the second loss function values ​​of all samples converge.

8. A model optimization device based on machine learning, characterized in that: include: The model iteration module is used to perform multiple model iteration operations, at least one of the multiple model iteration operations includes: The model is trained according to multiple samples for speech recognition or image recognition, labels of each sample, and probability values ​​of each sample being a label noise sample to obtain a first model, wherein a first loss function value of a single sample is determined according to a label prediction value corresponding to the single sample, the label of the single sample, and the probability value of the single sample being a label noise sample, and the training of the model is guided according to the first loss function values ​​of the multiple samples; Input the multiple samples into the first model to obtain a label prediction value of each sample, and determine a second loss function value of each sample according to the label prediction value of each sample and the label of each sample; The probability value of each sample being a label noise sample is updated according to the second loss function value of the multiple samples.

9. The device according to claim 8, characterized in that The first loss function value of the sample numbered i among the multiple samples is calculated according to the following formula: fl i =-[(1-p i )y i +p i z i ]log(h i )-[(1-p i )(1-y i )+p i (1-z i )]log(1-h i ), Among them, fl i is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

10. The device according to claim 8, characterized in that The first loss of the sample numbered i among the multiple samples is calculated according to the following formula: fl i =|(1-p i )y i +p i z i -h i |, Among them, fl i is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i Indicates the probability value of the sample numbered i being a label noise sample, h i Indicates the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

11. The device according to claim 8, characterized in that The first loss of the sample numbered i among the multiple samples is calculated according to the following formula: fl i =[(1-p i )y i +p i z i -h i ] 2 , Among them, fl i is the first loss function value of sample number i, y i is the label of the sample numbered i, z i is the label prediction value of the sample numbered i, p i is the probability value of the sample numbered i being a label noise sample, h i is the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

12. The device according to claim 8, characterized in that: The second loss function value of the sample numbered i among the multiple samples is calculated according to the following formula: sl i =-[y i log(h i )+(1-y i )log(1-h i )], Among them, sl i is the second loss function value of sample number i, y i is the label of the sample numbered i, h i is the probability value that the sample numbered i is a positive sample; Wherein, 1≤i≤N, and N is the number of samples of the plurality of samples.

13. The device according to claim 8, characterized in that Updating the probability value of each sample being a label noise sample according to the second loss function value of the multiple samples includes: Using the EM algorithm to estimate the parameters and mixing coefficients of the first distribution and the second distribution obeyed by the multiple samples, wherein the first distribution is the distribution of the probability values ​​of the labeled normal samples, and the second distribution is the distribution of the probability values ​​of the labeled noise samples; The probability value of each sample being a label noise sample is updated according to the second loss function value of each sample and the parameter and mixing coefficient of the second distribution.

14. The device according to claim 8, characterized in that The device also includes a stopping module, which is used to stop the model iteration operation when the number of model iteration operations reaches a set threshold; or, to stop the model iteration operation when the second loss function values ​​of all samples converge.

15. A model optimization device based on machine learning, characterized in that: include: at least one processor; And, a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, and the instructions are executed by at least one processor so that the at least one processor can execute the machine learning-based model optimization method according to any one of claims 1 to 7.

16. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, and when the program is executed by a processor, the processor executes the model optimization method based on machine learning according to any one of claims 1 to 7.