Device and Method for Correcting Noisy Lable Data
The method corrects noise labels in neural network training data by distinguishing and refining datasets using multiple neural networks and ensemble learning, enhancing the utilization and accuracy of training data.
Patent Information
- Application Number
- KR1020230140456
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2023-10-19
- Publication Date
- 2026-11-06
- Estimated Expiration
- 2043-10-19
AI Technical Summary
Neural networks trained with datasets containing incorrect labels, such as noise labels, operate with low accuracy and fail to utilize all available data effectively.
A method involving a series of neural networks and reliability detectors to distinguish between refined and unrefined data sets, applying augmentation techniques and ensemble learning to correct noise labels, utilizing all data for training.
Enables accurate label correction, allowing all data to be utilized for neural network training, thereby improving learning efficiency and network performance.
Smart Images

Figure R1020230140456_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a data correction device and method, and more specifically, to a device and method for correcting noise in label data used in a neural network. Background Technology
[0003] Neural networks are completed through training, and training data with accurate labels is required for training neural networks. However, not all datasets provided for neural network training have correct labels, and neural networks trained with such datasets do not operate with high accuracy.
[0004] To improve the learning efficiency of neural networks, a method was proposed in which the Overfitting to Underfitting network (O2U-Net) is trained by separating a small number of data points presumed to have the correct labels.
[0005] O2U-Net is a simple and effective method for detecting noise data, but it had a problem in that it could not utilize the remaining data because it only incorporated data with a high probability of having the correct labels into the training.
[0006] A method utilizing the remaining data through semi-supervised learning was proposed, but this method also had the problem of losing label information. The problem to be solved
[0008] The disclosed embodiments provide a noise label data correction apparatus and method that corrects inaccurate labels, such as noise labels, to enable learning by utilizing all data included in a training data set. means of solving the problem
[0010] According to one aspect of the disclosed embodiment, the method comprises the step (a) of distinguishing between a refined data set X and an unrefined data set u from a raw data set; and the refined data set (X) augmented by different augmentation methods. ik A separate neural network P associated with each augmentation method is used with (where i is an index representing the augmentation method and k is an index representing the number of iterations). i k Step of training (b); trained neural network P i k Input the unrefined dataset u into each, and for each neural network P i k Ensemble the output values to obtain the correction label data u' i k Step (c) of obtaining the calibration label data u' using the second label reliability detector i Refined dataset X from i k+1 A step (d) of obtaining ; repeating steps (b) to (d) a preset number of repetitions (K-1) to refine data set X i K A step of obtaining (e); and the refined data set X i K A noise label data correction method is provided, comprising the step (f) of obtaining a final refined data set X' through post-processing of the
[0011] The above step (f) is the refined data set X i K and a separate evaluation neural network P 0 neural network P using i K A step for training (f1); and the trained neural network P i K The above unrefined dataset u is input into each of the above, and the trained neural network P i K It includes the step (f2) of obtaining the final refined data set X' through an ensemble of the output values.
[0012] Step (a) distinguishes between refined data set X and unrefined data set u using a first label reliability detector, and the first label reliability detector and the second label reliability detector distinguish between the refined data set and the unrefined data set through neural network operations.
[0013] Each neural network P of step (c) above i k The ensemble of the output values of each neural network P i k Perform a weighted average operation on the output values.
[0014] In the above weighted average calculation, the weights are for each neural network P i k It is based on the correct answer rate.
[0015] The above refined data set X i k+1 is neural network P in the next iteration i k+1 It is used for learning.
[0016] When k=1, the refined dataset X is augmented with N different augmentation methods to X0 1 ~ X N-1 1 Creates, and the created X0 1 ~ X N-1 1 P set as the initial parameter i 1 It is used for learning.
[0017] According to another aspect of the disclosed embodiment, the processor comprises: a processor; and a memory connected to the processor, wherein the processor distinguishes between a refined data set X and an unrefined data set u from a raw data set (a); and a refined data set (X) augmented by different augmentation methods. i k A separate neural network P associated with each augmentation method is used with (where i is an index representing the augmentation method and k is an index representing the number of iterations). i kStep of training (b); trained neural network P i k Input the unrefined dataset u into each, and for each neural network P i k Ensemble the output values to obtain the correction label data u' i k Step (c) of obtaining the calibration label data u' using the second label reliability detector i Refined dataset X from i k+1 A step (d) of obtaining ; repeating steps (b) to (d) a preset number of repetitions (K-1) to refine data set X i K A step of obtaining (e); and the refined data set X i K A noise label data correction device is provided that performs the step (f) of obtaining a final refined data set X' through post-processing. Effects of the invention
[0019] According to the disclosed embodiments, there is an advantage that training is possible by correcting inaccurate labels, such as noise labels, and utilizing all data included in the training data set. Brief explanation of the drawing
[0021] FIG. 1 is a diagram showing a structure for distinguishing between reliable label data and unreliable label data from training data according to an embodiment. FIG. 2 is a drawing illustrating the overall structure of a noise label data correction device according to an embodiment. FIG. 3 is a drawing illustrating the detailed structure of an NLC module according to an embodiment. FIG. 4 is a diagram showing the operation structure of a post-processing module in a noise label data correction device according to an embodiment. FIG. 5 is a block diagram illustrating the overall flow of a noise label data correction method according to an embodiment. Specific details for implementing the invention
[0022] Throughout the specification, when a part is described as “comprising” a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as “...part,” “...unit,” “module,” and “block” as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware, software, or a combination of hardware and software.
[0023] FIG. 1 is a diagram showing a structure for distinguishing between reliable label data and unreliable label data from training data according to an embodiment.
[0024] Referring to FIG. 1, D (100) represents the entire data set (Set), Os represents the first label reliability detector (110), X (120) is the data set with reliable labels detected through the first label reliability detector (110) (refined data set), and u (130) is the data set with unreliable labels detected through the first label (110) (unrefined data set).
[0025] For the entire data set (100) initially provided, the first label reliability detector (110) detects reliability, and data with reliability above a specific threshold is included in the X data set (refined data set), and data with reliability below a specific threshold is included in the u data set (unrefined data set).
[0026] Various label reliability detectors may be used as the first label reliability detector (110), and for example, an O2U-Net neural network may be used as the first label reliability detector (110).
[0027] Each data included in the entire data set (100) (e.g., an image and its label) is classified as X or u according to the reliability detected by the first label reliability detector (110).
[0028] Conventionally, it was common practice to perform training using only X, a dataset deemed reliable, and not to use u, a dataset with no reliability, for training. The present disclosure proposes a method to correct label noise of u by correcting u, a dataset with no reliability.
[0029] FIG. 2 is a diagram illustrating the overall structure of a noise label data correction device according to an embodiment.
[0030] Referring to FIG. 2, a noise label data correction device according to an embodiment includes a plurality of NLC (Noise Label Correction) modules (210, 220, 230). The number of NLC modules (K-1) is pre-set, may be arbitrarily set by a designer, or may be determined based on the proportion of reliable data in the entire data set. For example, when the proportion of reliable data (X) in the entire data set is defined as p and a pre-defined scale constant is defined as d, K, which is related to the number of NLC modules, can be defined as Equation 1 below.
[0031]
[0032] The number of NLC modules effectively refers to the number of iterations of NLC operations; when implemented in software, it means performing NLC operations repeatedly.
[0033] Each NLC module (210, 220, 230) has the same structure, and the noise data correction device of the present disclosure illustrated in FIG. 2 repeatedly performs noise label correction work in the NLC modules (210, 220, 230) to correct the noise labels of data set u with low label reliability.
[0034] A refined dataset (X, which is a dataset with high label confidence) is used for noise label correction, and each NLC module (210, 220, 230) first trains a neural network for noise label correction using the refined dataset, and then uses the trained neural network to remove noise from a dataset u with low label confidence and outputs a refined dataset.
[0035] Referring to FIG. 2, the first NLC module (210) is a refined data set X0 2 , X1 2 , X2 2 Prints.
[0036] In the above dataset, the superscript is an index associated with the number of iterations, and the subscript is an index associated with the augmentation method, which will be explained later.
[0037] The second NLC module (220) is X0, a refined data set output from the first NLC module (210). 2 , X1 2 , X2 2 It takes as input a refined dataset X0 that is obtained by repeatedly removing noise from a low-confidence dataset u. 3 , X1 3 , X2 3 Prints.
[0038] Repetitive noise removal like this is X0 K , X1 K , X2 K It continues until output is produced.
[0039] Meanwhile, refined data sets augmented by different augmentation methods are input to each NLC module (210, 220, 230). FIG. 2 illustrates a case where refined data sets augmented by three different augmentation methods are input to each NLC module (210, 220, 230).
[0040] The first NLC module (210) receives X, a refined data set obtained from U, an initial data set, and X0, obtained by augmenting X, the refined data set, with three different augmentation methods. 1 , X1 1 , X2 1 Input X0 into the first NLC module (210). 2 , X1 2 , X2 2 Acquires X0 1 is a dataset augmented by the first augmentation method, and X1 1 is a dataset augmented by the second augmentation method, and X2 1 is a dataset augmented by the third augmentation method.
[0041] Various augmentation methods for data augmentation are already known, and it will be possible to obtain data sets augmented by different augmentation methods using such known augmentation methods. In addition, in this embodiment, the case is described as an example where three independent refined data sets augmented by three augmentation methods are input into each NLC module (210, 220, 230), but it will be obvious to those skilled in the art that the number of augmentation methods can be set differently depending on the designer.
[0042] Based on a predetermined number of iterations (K-1) through the NLC module, the final data X0 K , X1 K , X2 K If acquired, X0 K , X1 K , X2 KIt is input to the post-processing module (250). The post-processing module (250) finally corrects the label of u, which is a data set with low reliability (unrefined data set), and the specific operation of the post-processing module (250) will be explained after examining the specific operation of the NLC module with reference to FIG. 3.
[0043] Figure 3 is a diagram illustrating the detailed structure of an NLC module according to an embodiment.
[0044] Referring to FIG. 3, the NLC module according to an embodiment of the present disclosure can be broadly divided into two modules, the first module is a learning module (300), and the second module is a refined data set generation module (400).
[0045] The NLC module is equipped with independent neural networks corresponding to the number of augmentation methods, and since this embodiment explains using three augmentation methods as an example, FIG. 3 shows three independent neural networks (P0 k , P1 k , P2 k ) is located in the city.
[0046] In the NLC module, the learning module (300) is a neural network (P0) for each augmentation method k , P1 k , P2 k ) is trained. Referring to FIG. 3, X0 augmented by the first augmentation method k P0 using k Train , and X1 augmented by the second augmentation method k Using P1 k Training , and X2 augmented by the third augmentation method k Using P2 k It trains . Meanwhile, the neural network (P0 k , P1 k , P2 k When learning, P0 k , P1 k , P2 k The parameter of is P 0 It can be preset as a parameter of.
[0047] X0 k , X1 k , X2 k Since it is refined data, X0 K , X1 K , X2 K Using P0 k , P1 k , P2 k It is to train it. For example, in the case of data for object recognition, X is the image and the label information of the objects contained in the image, and X0 k , X1 k , X2 k Using the label information of P0 k , P1 k , P2 k It is to teach it.
[0048] Meanwhile, when the first training is performed in the NLC module, P0 is used with the initial refined dataset (dataset with high label reliability) X. 1 , P1 1 , P2 1 Train . At this time, augment X using the first method to X0 1 Generate and augment X using the second method to X1 1 Generates, and augments X using the third augmentation method to X2 1 Creates. Created X0 1 , X1 1 , X2 1 Using P0 1 , P1 1 , P2 1 It is to teach it.
[0049] In this way, through the learning module (300), the neural network P0 corresponding to each augmentation method k , P1 k , P2 k When the training is complete, the trained neural network P0 k , P1 k , P2 kThe refined data set generation module (400) using [the module] generates a refined data set from u with a noise label.
[0050] The refined data set generation module (400) receives an unrefined data set u (a data set determined to have low label reliability in FIG. 1) as input. Referring to FIG. 3, the unrefined data set u is a neural network (P0) trained by the learning module (300). k , P1 k , P2 k It is entered as ).
[0051] Each neural network (P0 k , P1 k , P2 k ) outputs label information for input data u through neural network operations based on learned neural network parameters. According to one embodiment, the label information may be output as probability information of classes, and it will be obvious to those skilled in the art that it may be information in a different form.
[0052] Each neural network (P0 k , P1 k , P2 k Data output from ) is ensemble according to a preset method to generate a correction label dataset corresponding to the number of augmentation methods.
[0053] Referring to Fig. 3, P1 k Label data output from and P2 k The label data output from is ensemble to obtain the first correction label data, u'0 k Generates P1 k Label data output from and P2 k P1 when ensembling label data output from k and P2 k A weighted average using the accuracy rate as weights may be performed by considering the accuracy rate, or an ensemble may be formed by the average without considering the weights. Here, the accuracy rate is each Pk P as a result of training the neural network k Refined dataset X used to train the neural network k a separate evaluation neural network P 0 It refers to the accuracy rate achieved when input into. Here, the evaluation neural network P 0 may be a neural network used to separate the initial dataset D into a refined dataset X and an unrefined dataset u. For example, P 0 It may be a neural network trained by semi-supervised learning.
[0054] In addition, the second correction label data, u'1 k is P0 k Label data output from and P2 k It is generated by ensembling the label data output from. P0 also during the ensemble for generating the second correction label data. k and P2 k An ensemble may be formed by the weighted average using the correct answer rate as a weight, or by the average without considering the weight.
[0055] In addition, the third correction label data, u'2 k is P0 k Label data output from and P1 k Generates by ensembling the label data output from.
[0056] The ensemble method for generating the first calibration label data, second calibration label data, and third calibration label data exemplified in FIG. 3 is exemplary, and it will be obvious to those skilled in the art that the selection of output data of the neural network for ensemble can be varied by the designer.
[0057] When the first to third calibration label data are output, for each calibration label data, the label data with high reliability among each calibration label data is selected through the second label reliability detector (410, Op).
[0058] For the second label reliability detector (410), a known reliability detector can be used, for example, the O2U-Net neural network can be used as the second label reliability detector (410). When the O2U-Net neural network is used as the second label reliability detector (410), the parameters of the O2U-Net neural network may be updated as learning takes place. Labels assigned to an unrefined dataset may be used during the learning of the second label reliability detector.
[0059] The second label reliability detector (410) selects and outputs only the data with high reliability among each calibration label data. The second label reliability detector (410) is the first calibration label data u'0 k Data with high reliability is selected, and the selected data is augmented using the first augmentation method.
[0060] Ultimately, the k-th NLC module is X0 k cast Input u'0 k X0, which is the data selected from augmented by the first augmentation method k+1 It is to output the (refined dataset). u'1 in the same way. k X1, which is data selected from and augmented by the second augmentation method k+1 This is printed, and u'2 k X2, which is data selected from and augmented by the third augmentation method k+1 This is printed.
[0061] X0 output as shown above k+1 , X1 k+1 and X2 k+1 It is input back into the NLC module, and the NLC module is X0 k+2 , X1 k+2 and X2 k+2 It prints . This repetition continues until a preset number of repetitions is reached, and when the repetition reaches the preset number of repetitions (K-1), finally X0 K, X1 K and X2 K is printed.
[0062] FIG. 4 is a diagram showing the operation structure of a post-processing module in a noise label data correction device according to an embodiment.
[0063] Referring to FIG. 4, the post-processing module (250) primarily X0 K , X1 K and X2 K and a separate evaluation neural network P 0 P0 using K , P1 K , P2 K Train P0 K , P1 K , P2 K The parameter of is P 0 Preset with the parameter of and X0 K , X1 K and X2 K It is learned using. Here, P 0 is P used when calculating the correct answer rate 0 P identical to 0 am.
[0064] P0 K , P1 K , P2 K Once training is complete, the unrefined dataset u is used with the trained P0 K , P1 K , P2 K Input into the neural network. Unlike the NLC module, the post-processing module (250) inputs each neural network (P0 K , P1 K , P2 K The output of ) generates X', a finally refined dataset, through a weighted average or average.
[0065] As explained earlier, P0 at weighted average K , P1 K , P2 K Weights may be determined based on the accuracy rate.
[0066] Unlike the NLC module, the post-processing module (250) must output labels for all data included in the unrefined data set u, so it does not perform the refining work through a separate sorter, but instead performs a weighted average or average of the outputs of three (number of augmentation methods) neural networks to assign refined labels to all data included in the unrefined data set u.
[0067] FIG. 5 is a block diagram illustrating the overall flow of a noise label data correction method according to an embodiment.
[0068] Referring to FIG. 5, first, the refined data set X and the unrefined data set u are distinguished from the raw data set (D) through the first label reliability detector (110) (step 500).
[0069] X0, a refined dataset augmented with different augmentation methods k , X1 k , X2 k P0, the neural network associated with each immediate method, using k , P1 k , P2 k Train (Step 510). At this time, P0 k , P1 k , P2 k The parameter of is P 0 It is preset with the parameters of and training is performed using the labels of.
[0070] If k=1, X0 1 , X1 1 , X2 2 It is obtained by augmenting the initial refined data set X with different augmentation methods.
[0071] Also, when k=1, P0 1 , P1 1 , P2 1 The parameters of can be preset to arbitrary initial values.
[0072] Neural network P0 k , P1 k, P2 k When the training is complete, the trained neural network P0 k , P1 k , P2 k The unrefined dataset u is input into each to obtain output values (step 520), and neural network P0 k , P1 k , P2 k The output values of are ensembled in a preset manner to obtain the correction label dataset u'0 k , u'1 k , u'2 k Obtain (Step 530).
[0073] As explained above, the ensemble of the output values of the neural network is formed in a manner predetermined by the designer, and the example illustrated in FIG. 3 is merely one embodiment.
[0074] In addition, as previously explained, when obtaining a correction label dataset by an ensemble of output values, the ensemble of output values can be formed through a weighted average or an average, and when weighted averaging, the weights can be set based on the accuracy rate of each neural network.
[0075] For each calibration label data, refinement data is obtained using a second label reliability detector (410), and the obtained refinement data is obtained using a related augmentation method X k+1 Obtain (Step 540). If three augmentation methods are used as in the described example, X0 k+1 , X1 k+1 , X2 k+1 This is obtained.
[0076] Steps 510 through 540 are repeated up to a preset number of iterations, K-1 times, and X0 is a refined data set obtained by repeating up to K-1 times. K , X1 K , X2 K and a separate evaluation neural network P 0 P0 using K , P1 K , P2K Train (Step 550).
[0077] Learned P0 K , P1 K , P2 K Input the unrefined dataset u into each and each neural network P0 K , P1 K , P2 K The output values from are ensembled to obtain the final refined data set X' (step 560).
[0078] The final refined dataset X' is a dataset in which the noise labels of the unrefined dataset u have been corrected, and the target neural network can be trained using X'. The embodiment according to the present disclosure has the advantage of allowing more data to be used for training by efficiently removing the noise from the noise label dataset u, which is not previously used for training, so that it can be used for training.
[0079] Meanwhile, the noise label data correction device according to the embodiment can be implemented by a processor and a memory connected thereto.
[0080] Although the foregoing has been described with reference to preferred embodiments, those skilled in the art will understand that various modifications and changes can be made to the present disclosure without departing from the spirit and scope of the present disclosure as set forth in the following claims.
Claims
Claim 1 A noise label data correction method performed by a noise label data correction device including a processor and memory, comprising the step (a) of distinguishing a refined data set X and an unrefined data set u from a raw data set; and a refined data set (X) augmented by different augmentation methods. i k A separate neural network P associated with each augmentation method is used with (where i is an index representing the augmentation method and k is an index representing the number of iterations). i k Step (b) for training the neural network P i k Input the unrefined dataset u into each, and for each neural network P i k Ensemble the output values to obtain the correction label data u' i k Step (c) of obtaining the calibration label data u' using the second label reliability detector i k Refined dataset X from i k+1 A step (d) of obtaining ; repeating steps (b) to (d) a preset number of repetitions (K-1) to refine data set X i K A step of obtaining (e); and the refined data set X i K A noise label data correction method comprising the step (f) of obtaining a final refined data set X' through post-processing. Claim 2 In claim 1, the above step (f) is a separate evaluation neural network P 0 neural network P using i K Initialize the parameters of, and the above refined data set X i K Using the above neural network P i K A step for training (f1); and the trained neural network P i K The above unrefined dataset u is input into each of the above, and the trained neural network P i K A noise label data correction method comprising the step (f2) of obtaining the final refined data set X' through an ensemble of output values. Claim 3 In claim 1, step (a) distinguishes between a refined data set X and an unrefined data set u from the raw data set using a first label reliability detector, wherein the first label reliability detector performs the distinction for the raw data set through neural network operation, and the second label reliability detector performs the correction label data u' through neural network operation. i k Refined dataset X from i k+1 Noise label data correction method for obtaining Claim 4 In claim 1, each neural network P of step (c) above i k The ensemble of the output values of each neural network P i k A noise label data correction method that performs a weighted average operation on the output values. Claim 5 In paragraph 4, in the above weighted average operation, the weights are each neural network P i k Noise label data correction method based on the accuracy rate. Claim 6 In claim 1, the refined data set X i k+1 is neural network P in the next iteration i k+1 Noise label data correction method used for learning. Claim 7 In paragraph 1, when k=1, the refined data set X is augmented by N different augmentation methods to X0 1 ~ X N-1 1 Creates, and the created X0 1 ~ X N-1 1 P set as the initial parameter i 1 Noise label data correction method used for learning. Claim 8 A processor; and a memory connected to the processor, wherein the processor includes the step (a) of distinguishing between a refined data set X and an unrefined data set u from a raw data set; and a refined data set (X) augmented by different augmentation methods. i k A separate neural network P associated with each augmentation method is used with (where i is an index representing the augmentation method and k is an index representing the number of iterations). i k Step (b) for training the neural network P i k Input the unrefined dataset u into each, and for each neural network P i k Ensemble the output values to obtain the correction label data u' i k Step (c) of obtaining the calibration label data u' using the second label reliability detector i k Refined dataset X from i k+1 A step (d) of obtaining ; repeating steps (b) to (d) a preset number of repetitions (K-1) to refine data set X i K A step of obtaining (e); and the refined data set X i K A noise label data correction device that performs the step (f) of obtaining a final refined data set X' through post-processing. Claim 9 In paragraph 8, the above step (f) is a separate evaluation neural network P 0 neural network P using i K Initialize the parameters of, and the above refined data set X i K Using the above neural network P i K A step for training (f1); and the trained neural network P i K The above unrefined dataset u is input into each of the above, and the trained neural network P i K A noise label data correction device comprising the step (f2) of obtaining the final refined data set X' through an ensemble of output values. Claim 10 In claim 8, step (a) distinguishes between a refined data set X and an unrefined data set u from the raw data set using a first label reliability detector, wherein the first label reliability detector performs the distinction for the raw data set through neural network operation, and the second label reliability detector performs the correction label data u' through neural network operation. i k Refined dataset X from i k+1 Noise label data correction device for acquiring Claim 11 In claim 8, each neural network P of step (c) above i k The ensemble of the output values of each neural network P i k A noise label data correction device that performs a weighted average operation on the output values. Claim 12 In Clause 11, in the above weighted average operation, the weights are each neural network P i k Noise label data correction device based on the accuracy rate of Claim 13 In claim 8, the above refined data set X i k+1 is neural network P in the next iteration i k+1 Noise label data correction device used for learning. Claim 14 In paragraph 8, when k=1, the refined data set X is augmented by N different augmentation methods to X0 1 ~ X N-1 1 Creates, and the created X0 1 ~ X N-1 1 P set as the initial parameter i 1 Noise label data correction device used for learning.
Citation Information
Patent Citations
Superloss: a generic loss for robust curriculum learning
KR1020220047534A
Label with noise processing method and apparatus of thereof
KR1020230110143A
Machine learning device and method
WO2019176806A1
