Label noise robust learning method based on double smooth driving

Abstract

The invention provides a label noise robust learning method based on double smooth driving. The label noise robust learning method comprises the following steps: acquiring a data set with noise; initializing two networks and various parameters; building a model and a loss function; inputting the data set into the two networks for initial training; smoothing processing is carried out on the label before each iteration training; dividing the original data set into a pseudo clean data set and a pseudo noise data set through a sample selection method; respectively inputting the pseudo clean data set into opposite network training; and performing loss smoothing on the pseudo-noise data set, and inputting the pseudo-noise data set into an opposite network for training. According to the method, aiming at the problem that noise exists in a data set, a dual-network model is adopted, and the sensitivity of the model to a noise label is reduced through dual smoothing and design of a new loss function, so that the robustness and generalization ability of the model when facing noise data are improved.

Description

Technical Field

[0001] The present invention belongs to the field of training deep learning models and relates to a double smoothing driven label noise robust learning method. Background Art

[0002] Deep learning has achieved excellent performance in tasks such as image recognition, natural language processing, and speech recognition, and has become the mainstream method of machine learning. Training deep learning models usually requires a large amount of labeled data, and collecting this data is often costly and time-consuming. In practical applications, data labeling companies often improve labeling efficiency through automatic labeling, crowdsourcing labeling, etc., but these methods still face problems with data noise and inaccurate labels. Existing regularization techniques, such as data augmentation, weight decay, and dropout, can alleviate the problem of overfitting, but it is difficult to completely solve the decline in generalization ability caused by label noise, especially when the label noise is more serious. In recent years, researchers have proposed methods such as reinforcement learning, self-supervised learning, and semi-supervised learning. These cutting-edge studies have shown that solving the label noise problem is not only about correcting the data, but also about improving model performance and robustness. Summary of the invention

[0003] In view of this, the present invention aims to propose a double smoothing driven label noise robust learning method to improve the classification accuracy of the model in the presence of label noise, in order to solve the technical problems existing in the above background technology.

[0004] To achieve the above object, the technical solution of the present invention is achieved as follows: A double smoothing driven label noise robust learning method comprises the following steps: Step 1. Get a dataset with noisy labels ; each of which are all independent and identically distributed from unknown joint distributions obtained in; n represents the total number of samples in D; Step 2. Initialize various parameters, including: Number of training times t=0, number of pre-training rounds , noise screening parameter η, pseudo label threshold δ and temperature parameter τ, label smoothing parameter , loss smoothing parameter ; Step 3. Initialize two deep neural networks and ; Build collaborative learning models and loss functions, including: cross entropy loss function , smoothed CE loss function ; Step 4. Transform the noisy dataset Images in and the corresponding labels Input into two deep neural networks respectively and Warm up training, training wheel; Step 5. During the training phase, perform label smoothing on the training data set to obtain the smoothed labels: ,in, express Smoothed labels; Step 6. Smooth the labeled dataset Input into the dual network and train with cross entropy loss; Step 7. Construct a small loss selection algorithm Select samples and divide the selected samples into pseudo-clean data sets and pseudo-noise dataset ; in, Used to control how many small loss samples should be in each training stage; Step 8. Pseudo clean dataset Input to the other party network through CE loss Continue training; add the pseudo noise dataset Input the other network through smoothing CE loss Continue training; in, ; Smoothed CE loss ; Step 9. Determine whether the current number of iterations t reaches the maximum number of iterations T; if the current number of iterations t does not reach the maximum number of iterations T, return to step 5 to continue training, t=t+1; otherwise, go to step 10; Step 10. After the model training is completed, a model that can distinguish between noisy labels and clean labels on the data set is obtained, and the model is used to perform classification prediction tasks.

[0005] The present invention has the following advantages: The present invention discloses a double smoothing driven label noise robust learning method, which proposes a more accurate clean sample selection strategy by improving the traditional "small loss sample selection" method.

[0006] This method effectively solves the overfitting problem caused by relying solely on small loss samples, allowing the model to more accurately screen out clean samples, thereby improving the quality of training data and the learning effect of the model.

[0007] The present invention introduces label smoothing, which reduces the model's sensitivity to label noise and avoids overfitting by treating labels as probability distributions rather than hard labels. This enhances the generalization ability of the model, enabling the model to better handle noise and uncertainty in practical applications, thereby improving performance in test sets or practical applications.

[0008] The present invention proposes a new loss function, which utilizes the large loss samples in collaborative learning and enhances the robustness of the model by reducing the sensitivity of the model to individual samples. The loss function optimizes the training process of the model, so that it exhibits stronger adaptability and better generalization performance when facing unseen data sets. The method of the present invention improves the accuracy of the deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flowchart of the label noise robust learning method driven by double smoothing in an embodiment of the present invention; Figure 2 This is a model structure diagram of the double smoothing driven label noise robust learning method in an embodiment of the present invention.

[0010] Figure 3 It is a schematic diagram of the process of dividing a data set in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0012] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0013] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0014] Glossary: Label smoothing is a regularization technique commonly used when training classification models. The core idea is to modify the original hard label (such as "0" or "1") by introducing a small smoothing factor to generate a soft label, so that the target label is no longer strictly 0 or 1, but a probability distribution between 0 and 1. Specifically, label smoothing reduces the overfitting of the model to a single category and improves its generalization ability by allocating part of the probability of the original label to other categories. Label smoothing is often used in multi-classification problems, especially in the training process of neural networks, which can help improve the stability and performance of the model.

[0015] Sample selection refers to the method of intelligently selecting or adjusting training samples during the training process in order to reduce the negative impact of label noise on model performance. The purpose of sample selection is to reduce the adverse effects of these wrongly labeled or unreliable samples on model training by identifying and filtering them out, or by assigning these samples a smaller weight. Specifically, sample selection can be performed by setting thresholds, model-based confidence assessment, or using advanced algorithms (such as self-training, ensemble methods, etc.), with the goal of ensuring that the model relies more on high-quality labeled data during training to improve its generalization ability and robustness.

[0016] In an environment with label noise, the performance of deep neural network models is often greatly affected. Co-teaching is an effective strategy that can reduce the negative impact of noise on model training in a label noise environment by training two networks to teach each other. However, existing co-learning methods still face many challenges, such as premature convergence and inaccurate sample selection. To this end, this paper proposes an improved co-learning method, called S-Co-teaching, to improve the robustness and generalization ability of the model. Methods First, a new clean sample selection method is proposed. By analyzing the loss of the sample, the overfitting problem caused by over-reliance on small loss samples in the traditional method is avoided, and clean samples can be screened more accurately. Secondly, in a noisy label environment, by introducing label smoothing, the impact of label noise can be effectively reduced, the overfitting of the model to the noisy labels of the training set can be reduced, and the robustness of the model can be improved. Finally, by designing a new loss function, the sensitivity of the model to the noisy labels is reduced, thereby improving the robustness and generalization ability of the model in the face of noisy data.

[0017] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, the double smoothing driven label noise robust learning method includes the following steps: Step 1. Get a dataset with noisy labels ; each of which are all independent and identically distributed from unknown joint distributions obtained in ; n represents the total number of samples in D.

[0018] By adding 20%, 40% and 60% noise to the labels of dataset D, label noise datasets with different noise ratios are obtained.

[0019] Training a dataset with noisy labels for classification tasks is challenging.

[0020] Step 2. Initialize various parameters, including: Number of training times t=0, number of pre-training rounds =30, noise screening parameter η, pseudo label threshold δ and temperature parameter τ, label smoothing parameter , loss smoothing parameter ; Among them, η is a manually set parameter, which indicates the proportion of pseudo noise samples to the total number of samples in each round of iteration. For example, when η is 0.2, it means that pseudo noise samples account for 20% of the total number of samples in each round of iteration.

[0021] Step 3. Initialize two deep neural networks and ; Build collaborative learning models and loss functions, including: cross entropy loss function , smoothed CE loss function ; like Figure 2 As shown, the model as a whole consists of three modules, including label smoothing, sample selection, and smooth CE loss; Randomly initialize the parameters of the two neural networks and build a co-teaching training model.

[0022] Step 4. Transform the noisy dataset Images in and the corresponding labels Input into two deep neural networks respectively and Warm up training, training wheel; Step 5. During the training phase, perform label smoothing on the training data set to obtain the smoothed labels: ,in, express Smoothed labels; Label smoothing is a common regularization technique that is often used to reduce the overfitting of models to noisy or abnormal labels in training data and improve the generalization ability of the model; In traditional classification tasks, when using the cross entropy loss function, the model tends to make the predicted value as close to 0 or 1 as possible, which may cause the model to be overly sensitive to noise or abnormal labels in the training data.

[0023] Label smoothing assigns some confidence in the true label to other categories to smooth the label distribution, reducing the model's tendency to overfit to low-probability events in the training set.

[0024] Step 6. Smooth the labeled dataset Input into the dual network and train with cross entropy loss; Note that the output value after the activation function of the penultimate layer is m: ; The label corresponding to this sample The probability value of each class is , the loss function of this network (cross entropy loss) is ; The smoothing parameter for label smoothing training is , then the smoothed label is: .

[0025] Step 7. Construct a small loss selection algorithm Select samples and divide the selected samples into pseudo-clean data sets and pseudo-noise dataset ; in, Used to control how many small loss samples should be in each training stage; Small loss samples are likely to be correctly labeled samples, so we first choose to use small loss samples in each mini-batch to train the classifier so that it is not disturbed by noisy labels; Because the two networks are based on different initial parameters, they have different learning capabilities, so the selected small loss samples will be more reliable than those selected by a single model; After obtaining the small loss samples, the average loss is calculated and further back-propagated to update the parameters of the two networks: .

[0026] Cross entropy is used as the classification loss to minimize the distance between the predictions of the two networks for clean samples and the labels. At the same time, in order to make use of the large loss clean samples directly discarded by the traditional dual network and prevent overfitting, a smooth CE loss is proposed to optimize the cross entropy loss function.

[0027] Step 8. Pseudo clean dataset Input to the other party network through CE loss Continue training; add the pseudo noise dataset Input the other network through smoothing CE loss Continue training; in, ; Smoothed CE loss ; In the Co-teaching method, CE is used as a selection function for training loss and screening clean samples. This method selects a certain proportion of samples with smaller CE loss values ​​from the sample set as clean samples for learning. In the early stage of training, due to the memory effect of the deep neural network model, the model quickly acquires a large amount of knowledge and has a certain ability to select clean samples. Therefore, the purity rate curve rises rapidly, and the model quickly reaches a high test accuracy. However, as the iteration process continues, the purity rate of the samples selected by the model gradually decreases, which indicates that the model begins to fit some noise samples, resulting in a decrease in their loss value. At this time, since noise samples may also have lower loss values, the model cannot accurately screen out clean samples using the small loss selection strategy.

[0028] Step 9. Determine whether the current number of iterations t reaches the maximum number of iterations T; if the current number of iterations t does not reach the maximum number of iterations T, return to step 5 to continue training, t=t+1; otherwise, go to step 10; The total training loss is .

[0029] Step 10. After the model training is completed, a model that can distinguish between noisy labels and clean labels on the data set is obtained, and the model is used to perform classification prediction tasks.

[0030] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. Double smoothing driven label noise robust learning method, characterized by: The following steps are involved: Step 1. Get a dataset with noisy labels ; each of which are all independent and identically distributed from unknown joint distributions obtained in; n represents the total number of samples in D; Step 2. Initialize various parameters, including: Number of training times t=0, number of pre-training rounds , noise screening parameter η, pseudo label threshold δ and temperature parameter τ, label smoothing parameter , loss smoothing parameter ; Step 3. Initialize two deep neural networks and ; Build collaborative learning models and loss functions, including: cross entropy loss function , smoothed CE loss function ; Step 4. Transform the noisy dataset Images in and the corresponding labels Input into two deep neural networks respectively and Warm up training, training wheel; Step 5. During the training phase, perform label smoothing on the training data set to obtain the smoothed labels: ； in, express Smoothed labels; Step 6. Smooth the labeled dataset Input into the dual network and train with cross entropy loss; Step 7. Construct a small loss selection algorithm Select samples and divide the selected samples into pseudo-clean data sets and pseudo-noise dataset ; in, Used to control how many small loss samples should be in each training stage; Step 8. Pseudo clean dataset Input to the other party network through CE loss Continue training; add the pseudo noise dataset Input the other network through smoothing CE loss Continue training; in, ; Smoothed CE loss ; Step 9. Determine whether the current number of iterations t reaches the maximum number of iterations T; if the current number of iterations t does not reach the maximum number of iterations T, return to step 5 to continue training, t=t+1; otherwise, go to step 10; Step 10. After the model training is completed, a model that can distinguish between noisy labels and clean labels on the data set is obtained, and the model is used to perform classification prediction tasks.

2. The double smoothing driven label noise robust learning method according to claim 1, characterized in that: The two initialized deep neural networks are two network models with the same structure.

3. The label noise robust learning method driven by dual smoothing according to claim 1, characterized in that: Set the number of warm-up rounds In this stage, the two networks each use the cross entropy loss function to train all samples in the data set independently, using the stochastic gradient descent algorithm. The two network models after preheating will have different random parameters.

4. The double smoothing driven label noise robust learning method according to claim 1, characterized in that: The step 5 is specifically as follows: Step 5.

1. Initialize label smoothing parameters ; Step 5.

2. Note that the output value of the activation function of the penultimate layer is m, and the sample belongs to The probability value of each class is ,but , in, Indicates that the last layer corresponds to The weight of the class; Step 5.

3. See it as The class template, The value of can be regarded as a simplified representation of the Euclidean distance, that is is negative, corresponding to and weighted No. The larger the class Euclidean distance, the smaller the probability of classification; Is positive, corresponding to a small Euclidean distance, and a large probability of classification: ； Step 5.

4. The label corresponding to this sample The probability value of each class is , the loss function of this network (cross entropy loss) is ; Step 5.

5. The smoothing parameter for label smoothing training is , then the smoothed label is .

5. The double smoothing driven label noise robust learning method according to claim 1, characterized in that: Small loss samples are correctly labeled samples, and a small loss selection algorithm is constructed.

6. The label noise robust learning method driven by dual smoothing according to claim 5, characterized in that: The step 7 is specifically as follows: Step 7.

1. Construct a small loss selection algorithm: ,in, Used to control how many small loss samples should be in each training stage; Step 7.

2. At the beginning of training, keep more small loss samples in each batch, that is, use a larger , because in this case the network will first fit the clean sample, and gradually decrease with the increase of epoch , until reaching , keep fewer samples in each batch; Step 7.

3. After obtaining the small loss samples, calculate their average loss and further backpropagate to update the parameters of the two networks: ; Step 7.

4. The large loss samples screened out in step 7.3 are not discarded directly, but their losses are smoothed and then propagated together with the small loss samples for training.

7. The double smoothing driven label noise robust learning method according to claim 1, characterized in that: By smoothing the loss values ​​of large loss samples, the model's ability to handle these samples is enhanced, avoiding overfitting due to excessive attention to large loss samples, and learning the information in clean samples.

8. The label noise robust learning method with dual smoothing drive according to claim 1, characterized in that: The step 8 is specifically as follows: Step 8.

1. The original cross entropy loss is ; Step 8.

2. Define the cross entropy loss under uniform distribution ; Step 8.

3. Get the smoothed CE loss: ， in, .

9. The label noise robust learning method driven by dual smoothing according to claim 1, characterized in that: The selected samples are divided into pseudo-clean data sets through the sample selection algorithm and pseudo-noise dataset .

10. The double smoothing driven label noise robust learning method according to claim 1, characterized in that: The total loss is ,in .

Images