A confidence-driven pseudo label generation method for noisy labels

Abstract

The application discloses a confidence-driven pseudo-label generation method for noise labels, and belongs to the fields of deep learning and image classification; specifically, first, original labels are collected as training data, a part of which is labeled to obtain labels q containing noise labels; another part of training samples x is input into a feature extractor for feature extraction, and then is respectively input into a classifier and a linear mapping module of a deep neural network model, the classifier outputs a distribution p, and the linear mapping module outputs a confidence value conf between 0 and 1; then, pseudo labels are constructed and loss training is performed, and a loss L CDPL is obtained, gradients are returned to the feature extractor and the classifier; finally, the confidence conf is trained as a pseudo label allocation ratio, and gradients are returned to a mapping function h conf of the linear mapping module, so that a more reasonable and balanced pseudo label information allocation ratio can be constructed at each stage. The application solves the problem of unbalanced information allocation in the existing pseudo label algorithm, and greatly reduces the influence of noise.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of deep learning, image classification and convolutional neural network, and particularly relates to a confidence-driven pseudo-label generation method for noisy labels. BACKGROUND

[0002] In recent years, deep neural networks have achieved very superior results in various image fields and natural language processing fields. One of the reasons why deep neural networks can achieve significant results is the support of large-scale labeled data sets. However, high-quality large-scale labeled data sets are very rare, and most large-scale data sets are obtained through low-cost means such as search engines, crowdsourcing or remote supervision. These data sets all have noisy labels, which seriously affect the generalization ability of the model.

[0003] Researchers have conducted a large number of studies on the processing of noisy labels, and the best existing method is the pseudo-label algorithm based on the pseudo-label algorithm. This kind of algorithm uses the fact that deep neural networks will first fit the characteristics of most samples with a simple template during the learning process, and compares the output of the deep neural network in the early stage of training with the label to determine whether the sample label is correct or not, and directly corrects the noisy label.

[0004] In the existing pseudo-label algorithm using model output for correction, a hyperparameter is used to allocate the information of the model output and the label information according to a fixed proportion, and the specific method is as shown in formula (1), this method does not consider that the amount of information contained in the model will gradually increase with the training, thus causing the problem of uneven distribution of information in the pseudo-label. Figure 1 SUMMARY

[0005] In order to solve the problem of uneven distribution of information in the existing pseudo-label algorithm, the present application proposes a confidence-driven pseudo-label generation method for noisy labels, which dynamically allocates the information of the model output and the label information by introducing a confidence to judge the amount of information contained in the model output.

[0006] The confidence-driven pseudo-label generation method comprises the following steps:

[0007] Step 1: Divide the collected original labels into two parts, one part of the training samples is input into the feature extractor, and the other part is labeled to obtain labels containing noisy labels ;

[0008] Step 2: After the feature extractor extracts the features of the training samples , input them into the classifier and the linear mapping module of the deep neural network model respectively, and the classifier outputs the distribution ​, the linear mapping module outputs a confidence score between 0 and 1 value.

[0009] The mapping function of the linear mapping module is a mapping function from the feature space to the interval [0, 1]; that is ;

[0010] The mapping function After outputting a real number value, it is mapped to a confidence score between 0 and 1 using the function;

[0011] The function is defined as follows:

[0012]

[0013] Finally, the sample confidence score of the current relation extraction model is denoted by .

[0014] Step three, use the distribution output by the classifier to construct a pseudo label combined with the confidence score and perform loss training; the loss gradient is backpropagated to the feature extractor and the classifier;

[0015] The pseudo label calculation formula is as follows:

[0016]

[0017] wherein denotes the original label containing noise, and denotes the classifier output distribution.

[0018] When the confidence score tends to 1, the pseudo label is used to replace the labeled label, and when the labeled label contains noise, the noise in the labeled label is corrected. When the confidence score tends to 0, the pseudo label is the original label that may contain noise, and the labeled result is selected to be trusted.

[0019] The loss function for training the pseudo label is:

[0020]

[0021] is the number of training data, is the generated distribution , and is the divergence between the pseudo label

[0022] Step four, using confidence training loss function confidence Training is carried out.

[0023] The confidence training loss function is calculated as follows:

[0024]

[0025] The hyperparameter is used to control The value size level.

[0026] The loss is divided into two parts: the first part is the generated pseudo label The divergence between the label The divergence between the label The divergence between the label The divergence between the label The second part is to maximize The value, that is, to make the generated The value as large as possible. The value as large as possible.

[0027] Step five, use the trained confidence As a pseudo label distribution ratio, the gradient is returned to the mapping function of the linear mapping module So that a more reasonable and more balanced pseudo label information distribution ratio can be constructed at each stage.

[0028] The advantages of the present application are:

[0029] A confidence-driven pseudo label generation method for noisy labels, by using an additional confidence module and a confidence training algorithm, a reasonable dynamic information distribution weight is generated, so that the generated pseudo label is closer to the real semantic label, and the existing information distribution imbalance problem in the existing pseudo label algorithm can be solved. The influence of noise in the label space is greatly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The present application is a confidence-driven pseudo label generation method for noisy labels;

[0031] Figure 2 The present application is a confidence-driven pseudo label generation method for noisy labels;

[0032] Figure 3 The present application is a confidence-driven pseudo label generation method for noisy labels;

[0033] Figure 4 An introduction confidence pseudo label generation algorithm used in the present application. DETAILED DESCRIPTION

[0034] The present application will be further explained in detail below with reference to the embodiments and the accompanying drawings.

[0035] The present application proposes a confidence-driven pseudo label generation method for noise labels. Through an end-to-end trainable method, i.e., a confidence-driven pseudo label generation algorithm (CDPL), the additional computational overhead required can be negligible, to realize reasonable dynamic allocation of label information in the pseudo label and output information of the neural network model.

[0036] Unlike other pseudo label schemes that combine network output distribution, the present method gradually and adaptively modifies the target, so that the pseudo label distribution is more reasonable and balanced. The core design of the present application comes from the following two reasons: 1) When a deep neural network model starts learning from scratch, the labeled results containing noise are more reliable than the output of the early model, and even if there is a lot of noise in the data labeling, the model will first learn simple patterns to fit most of the labels without remembering the noise. 2) As the number of training iterations increases, the model can acquire more knowledge and make more confident predictions, at which time the model output is more reliable than the label, so the model output is used to modify its semantic label.

[0037] The confidence-driven pseudo label generation method for noise labels, as shown in Figure 1 and Figure 2 , is divided into the following steps:

[0038] Step one, divide the collected original labels into two parts as training data, one part of the training samples is input into the feature extractor, and the other part is labeled to obtain labels containing noise labels ;

[0039] Step two, the feature extractor extracts features from the training samples , and inputs them into the classifier and linear mapping module of the deep neural network model, respectively. The classifier outputs a distribution , and the linear mapping module outputs a confidence value between 0 and 1.

[0040] This embodiment selects to split the ordinary classifier network into a feature extractor and a classifier, and additionally adds a linear mapping module after the feature extractor to form an improved deep neural network model.

[0041] The linear mapping module is used for calculating the confidence of a deep neural network model for a prediction result , which represents the reliability of the model output distribution; when is large, it indicates that the model output is very feasible, and when it is small, it indicates that the model is not sure about the result of its own output.

[0042] The confidence is a simple way to obtain: directly use the maximum probability value of the deep neural network model output distribution , that is ; according to the existing algorithm, the confidence of the model comes from two uncertainties:

[0043] 1) Arbitrary uncertainty (Aleatoric Uncertainty): This uncertainty comes from the ambiguity of the sample, that is, the confidence corresponding to the difficult sample will be lower.

[0044] 2) Cognitive uncertainty (Epistemic Uncertainty): This uncertainty comes from the lack of cognition of the model, mainly caused by insufficient samples.

[0045] At the beginning of training, the reason for the low confidence is basically due to cognitive uncertainty, and the maximum value of the softmax probability of the model output loses part of the arbitrary uncertainty, so it cannot accurately represent the confidence of the model for the output.

[0046] To solve this problem, the embodiment uses an additional sigmoid layer mapping to obtain the confidence of the model, and the confidence performs better than the maximum value of the softmax probability.

[0047] As shown in Figure 3 , wherein the softmax module is used for classification result output, the sigmoid module is used for outputting the confidence, the loss function is only used for training the confidence module of the sigmoid, and the loss function is used for training the complete classification network by using the pseudo label.

[0048] As can be seen, the confidence value is mapped by an additional linear layer, and the layer is The mapping layer is ; is a mapping function from the feature space to the real number interval; Since the confidence value required by the present application is a real value in the interval of 0 to 1, the output real value is obtained by using ​The function maps it to a confidence value between 0 and 1. The function is defined as follows:

[0049]

[0050] Due to its good derivative property, the function is commonly applied in the activation function of neural networks. Ultimately, the sample Confidence about the current relation extraction model value is represented by .

[0051] Step three, use the distribution output by the classifier Combine the confidence value to construct pseudo labels and train the loss; the loss gradient is backpropagated to the feature extractor and the classifier;

[0052] As shown in Figure 4 , the proportion of output and label is dynamically allocated by the additional confidence to construct the pseudo label, and the pseudo label calculation formula is as follows:

[0053]

[0054] Wherein denotes the original label with noisy label, denotes the output distribution of the classifier.

[0055] When the confidence tends to 1, the pseudo label at this time is the output result of the deep neural network model, indicating that the model is very confident about its own output result, so it can be directly used to replace the label, and when the label contains noise, the noisy label will be corrected. When the confidence tends to 0, the pseudo label at this time is the original label (may contain noise), because the model at this time cannot judge whether the label is correct, so it can only choose to believe the label result.

[0056] From the intuitive point of view, compare the noisy label to the possible wrong homework answers, and the deep neural network model to the student. The confidence indicates the student's confidence in his own calculation result, when the student is very confident about the result he calculated, but the answer is different, at this time it can be judged that the answer is wrong, then the pseudo label is the model's own output result. When the student is very unconfident about the result, the pseudo label is still the homework answer.

[0057] As can be seen, the proposed solution can effectively correct noisy labels to correct labels for samples where the model is highly certain about the output, and because the semantic labels can also be corrected, it does not cause overfitting. However, when the model is not very confident about the output, it can fully trust the labeling results to prevent the model from failing to learn difficult samples. For some less certain samples, the confidence level... When the values ​​are neither too high nor too low, the generated pseudo-labels become very smooth. In this case, the loss is relatively small and will not have a significant impact on the model. Therefore, this is equivalent to learning the sample with a lag. The model will learn from the sample again when there are enough iterations and the model has learned more knowledge and has a stronger judgment ability.

[0058] After obtaining the pseudo-labels, the loss function for training on these pseudo-labels is:

[0059]

[0060] For the amount of training data, For the generated distribution With pseudo-tags Between Divergence;

[0061] The loss is basically the same as that of other pseudo-labeling methods, but the pseudo-labels constructed by the CDPL algorithm are closer to the real label distribution, therefore... The gradient generated by the loss is "cleaner". In each iteration, the training algorithm performs a two-stage algorithm update:

[0062] In the first stage, the classifier network outputs an ensemble distribution for constructing pseudo-labels, while the linear mapping module outputs a one-dimensional confidence value in the interval of 0 to 1. Then, based on the distribution and confidence values ​​obtained from the output... Constructing pseudo tags During this stage, losses will also be calculated. , used for Confidence learning in the mapping layer.

[0063] The second stage is the model training stage, where the labels are the pseudo-labels constructed in the first stage. The loss function used is Divergence loss.

[0064] Step 4: Train the loss function using pseudo-labels and confidence scores. Mapping functions for linear mapping modules Conduct training;

[0065] The formula for calculating the confidence training loss function is as follows:

[0066]

[0067] is an hyper-parameter, used to control the value size level.

[0068] The loss is divided into two parts: the first part is the divergence between the generated pseudo label and the one-hot label ; this part represents making the pseudo label closer to the one-hot label . At this time, the label There are two different processing conditions:

[0069] 1) When there is a clean auxiliary data set, the one-hot label is represented by the clean label, and the generated pseudo label is generated by the noisy label of remote supervision; this case does not need the second half, only the divergence between the pseudo label and the clean label is calculated; minimizing this loss means making the generated pseudo label closer to the real label.

[0070] 2) When there is no clean data set, the one-hot label is represented by the noisy label of remote supervision; according to the definition of the pseudo label , if only the first term, the loss minimum condition is , that is, all values equal to 0 can be satisfied. However, this is not the required confidence value of the present application, so in this case, the second part is added: maximizing the value of , that is, making the generated value as large as possible.

[0071] In the loss , the first term makes the pseudo label as close to the label as possible, and the second term makes the generated confidence value as large as possible; then according to the definition of the pseudo label and the loss , the loss minimum condition is when the distribution ​​​​ Proximity to label At that time, confidence level Large value, and distribution When the output is highly unstable, confidence level Smaller values ​​allow for labeling. Replacement pseudo-tags Because the model cannot obtain data containing annotations. Prior knowledge, therefore in Under the combined effect of the two factors, when the model controls its own output distribution... When you are confident, The value will be relatively large, and when the input sample When the pattern has not yet been learned or the semantics are relatively ambiguous, The value will be smaller because the distribution at this time... With annotation The distances between them are likely to differ significantly. This makes the final output... This value simultaneously guarantees both arbitrary uncertainty and cognitive uncertainty. During the training phase, this part of the loss gradient is backpropagated to... The mapping module, feature extractor, and relation classifier will not be trained.

[0072] The final output confidence level The value aligns with the confidence concept required by this invention, that is, the confidence level when the model is highly confident in its output. The value will be higher, and when there is uncertainty about the output. The value will be too low.

[0073] Step 5: Utilize the confidence scores after training As a proportion of pseudo-label assignment, the gradient is back-passed to the mapping function of the linear mapping module. Training is conducted to construct a more reasonable and balanced distribution ratio of pseudo-label information at each stage of training.

[0074] Once the confidence score is trained, the confidence score represents the model's level of confidence in the output.

[0075] By using confidence level as the pseudo-label allocation ratio, as shown in the formula... Compared with the existing fixed proportion allocation method, the method is more smooth and more consistent with the noise correction target. The confidence considers that the amount of information contained in the model will gradually increase with the training. In the early stage of training, the model is not accurate in judging the result, so the confidence will be low, and the information allocation of the pseudo label will tend to the original label. With the increase of the number of training iterations, the confidence will also increase according to the accuracy of the model, and at this time the information allocation will tend to the model output. Thus, a more reasonable and balanced pseudo label information allocation ratio can be constructed at each stage of training.

Claims

1. A confidence-driven pseudo-label generation method for noisy labels, characterized in that, The steps are as follows: First, the collected original labels are divided into two parts as training data, one part is labeled to obtain labels containing noise labels ; the other part of the training sample After the feature extractor extracts the features, it is input into the classifier and the linear mapping module of the deep neural network model respectively, the classifier outputs the distribution , and the linear mapping module outputs the confidence value from 0 to 1 ; the mapping function of the linear mapping module is a mapping function from the feature space to the real interval; i.e. ; mapping function After outputting a real value, it is mapped to a confidence value between 0 and 1 using the function The function is defined as follows: Finally, the samples Regarding the confidence of the current relation extraction model Values are represented by ; Then, the distribution output by the classifier is used In combination with the confidence The value constructs a pseudo label And loss training is carried out; the loss The gradient is returned to the feature extractor and the classifier; The pseudo-label calculation formula is as follows: wherein denotes the original label with noise, denotes the classifier output distribution; The loss function for training the pseudo-label is the number of training data, the generated distribution with pseudo labels between divergence; Finally, using the training loss function The confidence Training, and the trained confidence As a pseudo-label assignment ratio, the gradient is returned to the mapping function of the linear mapping module , So that a more reasonable and more balanced pseudo-label information assignment ratio can be constructed at each stage.

2. The method of claim 1, wherein, The confidence When approaching 1, the pseudo label is used to replace the labeled label, and the labeled label with noise is corrected; when the confidence When approaching 0, the pseudo label is the original label that may contain noise, and the labeled result is selected.

3. The method of claim 1, wherein, The training loss function The calculation formula is as follows: is a hyperparameter, used to control the value size level; The loss is divided into two parts: the first part is the generated pseudo-labels. with annotation Between Divergence; this part indicates that pseudo-labels One-hot labeling that is closer to the label The second part is maximization. of Value, that is: the value that makes the generated The value should be as large as possible.

Images