An adversarial active learning method based on independence enhancement

Abstract

This invention relates to an adversarial active learning method based on independence enhancement. It includes the following steps: 1. Using a generator to reconstruct samples to learn their feature representations, and concatenating these representations with annotation features extracted from the task model; 2. Designing a Batch Loss Prediction Module (BLPM) to re-label the state values ​​of the samples, and training the discriminator based on the new state values; 3. Designing an independent sample selection module based on hierarchical clustering to perform independence selection on the samples, labeling the selected samples, adding the labeled samples to the label pool, and updating and training the model. This invention not only considers the information content of a single sample but also takes into account the independence between samples within a group, while reducing the total labeling cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of active learning, and proposes an adversarial active learning method based on independence enhancement. BACKGROUND

[0002] The purpose of the pool-based method is to query samples with rich information from the sample pool, and then label these samples.

[0003] In the pool-based method, the uncertainty-based method first trains the task model using the existing labeled samples, and then selects the unlabeled samples with the highest uncertainty of the model for labeling in each iteration, that is, the method takes uncertainty as an indicator of the information contained in the sample. Deep active learning not only can handle high-dimensional data, but also effectively reduces the labeling cost. Therefore, some scholars have carried out research work on deep active learning. For example, Yang et al. proposed a multi-class active learning method, which uses the entire pool to evaluate the uncertainty of the data and considers the diversity of the selected samples; Gal et al. applied the Bayesian deep learning method to active learning, developing an effective active learning framework for high-dimensional data; Yang et al. combined fully convolutional layers with active learning, using the uncertainty information provided by the fully convolutional layers to determine the annotation region, thereby significantly reducing the annotation difficulty; recently, Yoo et al. designed a loss prediction module (LPM) that can predict the loss value generated by the task model for the sample, LPM is optimized with the task model and is used to predict the loss of the unlabeled sample, and the sample is selected according to the loss prediction value.

[0004] The pool-based method also includes a distribution-based method that selects samples that can increase the diversity of the labeled pool by analyzing the distribution of labeled samples and unlabeled samples. Nguyen et al. increase the diversity of samples by clustering unlabeled samples and selecting samples in different clusters; Hasan et al. propose a continuous learning framework to identify context-aware information from unlabeled video data. In addition, some scholars use gradients and errors to evaluate the diversity of the distribution. The Variational Adversarial Active Learning (VAAL) method uses a variational autoencoder (VAE) and a discriminator to learn adversarially, which selects samples based on diversity, but the diversity of samples is not directly related to the amount of information contained in the samples. The SRAAL (State-Relabeling Adversarial Active Learning) method introduces a task-related index. Unlike directly assigning the same label to unlabeled samples, the online uncertainty index (OUI) is used to relabel the unlabeled samples to calculate the contribution of the samples to the model, and good performance is obtained. SRAAL uses annotation information and labeled / unlabeled state information to select samples, fully considering the distribution and uncertainty of samples. OUI introduces the maximum probability value and variance of the probability vector in the calculation process, but its application effect is general. For example, suppose the predicted class vector of the sample by the task model is V, and the value of the kth element is the maximum, and the real class of the sample is any one of the classes, and the contribution degree calculated by OUI is the same, which limits the performance of the model.

[0005] The present application considers the uncertainty and diversity of samples and the similarity between samples, and aims to increase the independence between samples. In addition, the present application improves the LPM, and uses the improved LPM to predict the state value of the unlabeled sample. In order to meet the requirement of SRAAL for the accuracy of the state, the present application designs a new loss function for the LPM, and names the improved module as Batch Loss Prediction Module (BLPM), and then uses the BLPM to replace the OUI in SRAAL.

[0006] The mainstream active learning method currently usually selects the samples with the top K information content as an optimal sample combination for labeling, wherein K is the number of samples in the combination. In fact, the total information content of a group of samples is not simply the sum of the information content of the samples in the group, that is, the sample combination selected only according to the information content ranking does not necessarily contain the highest total information content. The reason is that some samples in the group may have high similarity, and the existence of similar samples causes the contribution of some samples in the group to the model to be much smaller than their information content. The present application considers that the total information content of the samples in the group is inversely proportional to the similarity between the samples and is proportional to the independence between the samples. Labeling the samples with high independence will obtain more efficient labeling. Therefore, the present application not only focuses on the information content of a single sample but also takes into account the independence between the samples in the group. In order to increase the independence between the samples in the group, the present application designs an independent sample selection module based on hierarchical clustering. The module selects the optimal sample combination by using the pseudo-labels after hierarchical clustering and the predicted values of the samples by the discriminator. SUMMARY

[0007] The present application comprehensively considers the diversity, uncertainty and independence between the samples, and proposes an active learning method based on independence enhancement to solve the problem of too high similarity between the samples in the active learning method.

[0008] An active learning method based on independence enhancement, characterized in that it comprises the following steps:

[0009] Step 1: using an unsupervised image reconstruction network based on a coding-decoding structure as a generator to reconstruct the samples, learning the feature representation of the samples, and splicing the feature representation of the samples with the annotated features extracted in the task model;

[0010] Step 2: designing a loss prediction module Batch Loss Prediction Module (BLPM) with a new loss function, using the BLPM to relabel the state values of the samples, and training the discriminator according to the new state values;

[0011] Step 3: designing an independent sample selection module based on hierarchical clustering to select the independence of the samples, then labeling the selected samples, adding the labeled samples to the labeled pool, and finally updating and training the model.

[0012] The step 1 comprises the following steps:

[0013] Step 1.1: using the generator G to reconstruct the samples to learn the feature representation of the samples. The encoder encodes the samples to learn the features of the samples, and the decoder is used to reconstruct the samples. The objective function in the coding-decoding process is:

[0014]

[0015]

[0016]

[0017] where is the objective function for unlabeled samples, is the objective function for labeled samples, and z is the feature representation. The generator learns a rich feature representation by reconstructing all samples.

[0018] Step 1.2: The present application splices the sample feature learned by the encoder with the feature vector extracted from the task model, and then takes the spliced vector as the final feature representation of the sample.

[0019] The step 2 includes the following steps:

[0020] Step 2.1: The present application takes the loss prediction value of the sample as the uncertainty indicator. A BLPM with a new loss function is designed, which can compare the loss values of all samples in a batch:

[0021]

[0022]

[0023] where, l i and l j are the real losses of the task model, and are the task losses predicted by the BLPM module. For a pair of samples (x i , x j ), when l i > l j and are simultaneously satisfied, the module has no loss, otherwise the module will have a loss, so that it increases the value of and reduces the value of . The joint loss function of the target model and the BLPM is:

[0024]

[0025] After that, the BLPM will predict the loss value of the unlabeled sample. Finally, the present application normalizes the output of the BLPM to [0, 1], and takes the normalized value as the state value of the unlabeled sample.

[0026] Step 2.2: The objective function of the discriminator D is:

[0027] L D = -E[log(D(qθ (z L |x L )))]-E[log(state(x U ) new -D(q θ (z L |x L )))]

[0028] where state(x U ) new is the state value of unlabeled samples. The objective function of generator G and discriminator D in the process of adversarial learning is:

[0029] L adv =-E[log(D(q θ (z L |x L )))]-E[log(D(q θ (z U |x U )))]

[0030] The total objective function of generator G is defined as follows:

[0031] L Total =λ2L G +λ3L adv

[0032] The step 3 comprises the following steps:

[0033] Step 3.1: In order to reduce the similarity between selected samples and improve the independence between samples, the independent sample selection module based on hierarchical clustering is designed. The class of samples is initialized first, wherein each sample is a class, that is, and the initial cluster number C=N is set, the merging speed is controlled by hyperparameter p, and the total merging proportion is controlled by hyperparameter s; then, the hierarchical clustering of the feature representation of the sample from bottom to top is performed, and the pseudo label y i of x i is constantly updated. In the clustering process, the mean of the Euclidean distance of all samples in two clusters is used as the distance measurement of the two clusters:

[0034]

[0035] where C a and C b are two clusters, v i and v j are the spliced vectors corresponding to the samples in the two clusters, n a and n b correspond to C aand C b the number of samples in the cluster.

[0036] Step 3.2: After the clustering is completed, the present application selects samples according to the descending order of the predicted values, and the selected samples are all from different clusters.

[0037] Step 3.3: Finally, the present application uses Oracle to label the selected top K samples. This process takes into account the amount of information of a single sample and also considers the independence between samples. Finally, the labeled samples are added to the labeled pool and the model is updated and trained.

[0038] Advantages of the present application: At present, mainstream active learning methods evaluate the contribution of samples to the model by calculating the uncertainty or sample distribution of unlabeled samples. Although some research results integrate uncertainty-based methods and distribution-based methods, the application effect is not ideal due to the lack of high independence between samples. Therefore, the present application designs an independent-enhanced adversarial active learning method. First, a generator is used to reconstruct the sample to learn the feature representation of the sample, and the feature representation of the sample is spliced with the annotation feature extracted in the task model; second, a BLPM is designed to re-label the state value of the sample, and the discriminator is trained according to the new state value; then, a hierarchical clustering-based independent sample selection module is designed to select samples for independence, and then the selected samples are labeled. Finally, the labeled samples are added to the labeled pool and the model is updated and trained. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of an independent-enhanced adversarial active learning method;

[0040] Figure 2 is a model structure diagram of the independent-enhanced adversarial active learning method;

[0041] Figure 3 is a performance comparison diagram of the present application and the current mainstream method on two image classification data sets and two image segmentation data sets. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described below with reference to the three specific drawings in the embodiments of the present application.

[0043] Figure 1 is a specific flowchart of the present application, which comprises the following steps:

[0044] Step 1: An unsupervised image reconstruction network based on a coding and decoding structure is used as a generator to reconstruct the sample, learn the feature representation of the sample, and splice the feature representation of the sample with the annotation feature extracted in the task model;

[0045] Step 2: LPM (BLPM) with a new loss function is designed, the state value of the sample is re-labeled using BLPM, and the discriminator is trained according to the new state value;

[0046] Step 3: A hierarchical clustering-based independent sample selection module is designed to select independent samples, then the finally selected samples are labeled, and the labeled samples are added to the label pool, and finally the model is updated and trained.

[0047] The step 1 comprises the following steps:

[0048] Step 1.1: As shown in the figure, the application uses the generator G to reconstruct the sample and learn the feature representation of the sample. The encoder encodes the sample to learn the features of the sample, and the decoder is used to reconstruct the sample. The objective function in the encoding and decoding process is: Figure 2

[0049]

[0050]

[0051]

[0052] Wherein is the objective function of the unlabeled sample, is the objective function of the labeled sample, and z represents the feature representation. The generator learns rich feature representation by reconstructing all samples.

[0053] Step 1.2: The application splices the sample features learned by the encoder with the feature vectors extracted from the task model, and then splices the vectors as the final feature representation of the sample.

[0054] The step 2 comprises the following steps:

[0055] Step 2.1: The application takes the loss prediction value of the sample as the uncertainty index. LPM (BLPM) with a new loss function is designed. Specifically, the loss function of BLPM compares the loss values of all samples in a batch:

[0056]

[0057]

[0058] Wherein l i and l j are the real losses of the task model, and ​The task loss predicted for the BLPM module. For a pair of samples (x i ,x j ), the module has no loss when l i > l j and , otherwise the module has loss, thus it increases the value of l and decreases the value of l . The joint loss function of the target model and the BLPM is:

[0059]

[0060] After that, the BLPM will predict the loss value of the unlabeled samples. Finally, the output of the BLPM is normalized to [0, 1] by the present application, and the normalized value is taken as the state value of the unlabeled sample.

[0061] Step 2.2: The objective function of the discriminator D is:

[0062] L D =-E[log(D(q θ (z L |x L )))]-E[log(state(x U ) new -D(q θ (z L |x L )))]

[0063] Where state(x U ) new represents the state value of the unlabeled sample. The objective function of G and D in the adversarial learning process is:

[0064] L adv =-E[log(D(q θ (z L |x L )))]-E[log(D(q θ (z U |x U )))]

[0065] The total objective function of G is defined as follows:

[0066] L Total =λ2L G +λ3L adv

[0067] The step 3 includes the following steps:

[0068] Step 3.1: As Figure 2As shown, the independent sample selection module based on hierarchical clustering is designed to reduce the similarity between selected samples and improve the independence between samples. The initial clustering number C=N is set, the merging speed is controlled by the hyperparameter p, and the total merging proportion is controlled by the hyperparameter s; then, the feature representation of the sample is subjected to bottom-up hierarchical clustering, and the pseudo-label y of x is constantly updated i i During the clustering process, the mean of the Euclidean distances of all samples in two clusters is used as the distance measurement of the two clusters:

[0069]

[0070] Where C a and C b are two clusters, v i and v j represent the spliced vectors corresponding to the samples in the two clusters, n a and n b correspond to the sample numbers in C a and C b .

[0071] Step 3.2: After the clustering is completed, the samples are selected according to the descending order of the predicted values, and the selected samples are all from different clusters; finally, the selected top K samples are labeled by using the oracle. In summary, this process takes into account the information amount of a single sample while also considering the independence between samples.

[0072] Step 3.3: The selected samples are labeled, and the labeled samples are added to the labeled pool; finally, the model is updated and trained. Figure 3 The performance comparison results of the present application and the current mainstream methods on multiple data sets are shown.​

Claims

1. An independence-enhanced based adversarial active learning method, characterized in that, Comprising the following steps: Step 1: using an unsupervised image reconstruction network based on a coding structure as a generator to reconstruct the samples, learn the feature representation of the samples, and splice the feature representation of the samples with the annotation features extracted in the task model; Step 2: a loss prediction module Batch Loss Prediction Module (BLPM) with a new loss function is designed, and the state value of the sample is relabeled using the BLPM, comprising the following steps: Step 2.1: the loss function compares the loss values of all samples in a batch: wherein l i and l j are the real losses of the task model, and are the predicted task losses of the BLPM module, for a pair of samples (x i , x j ), the module has no loss only when l i > l j and are satisfied, otherwise the module will have a loss, thus it increases the value of l and decreases the value of l , the joint loss function of the target model and the BLPM is: After that, the BLPM predicts the loss value of the unlabeled sample, and the output is normalized to [0, 1], and the normalized value is taken as the state value of the unlabeled sample; Step 2.2: the objective function of the discriminator D is: L D = -E[log(D(q θ (z L |x L )))] - E[log(state(x U ) new -D(q θ (z L |x L )))] where state(x U ) is the state value of the unlabeled sample, and the objective function of the generator G and the discriminator D in the adversarial learning process is: L adv = -E[log(D(q θ (z L |x L )))] - E[log(D(q θ (z U |x U ))] The total objective function of the generator G is defined as follows: L Total = λ2L G + λ3L adv Then take the loss prediction value of the sample as the uncertainty index, and train the discriminator according to the new state value; Step 3: an independent sample selection module based on hierarchical clustering is designed to select independent samples, then the selected samples are labeled, and the labeled samples are added to the labeled pool, and finally the model is updated and trained.

2. The independence-enhanced based adversarial active learning method according to claim 1, characterized in that, The step 1 comprises the following steps: Step 1.1: using the generator G to reconstruct the samples to learn the feature representation of the samples, wherein the encoder encodes the samples to learn the features of the samples, and the decoder is used to reconstruct the samples, and the objective function in the coding and decoding process is: wherein, Jtarget is the objective function for unlabelled samples, Jtarget is the objective function for labelled samples, z represents the feature representation, the generator learns the feature representation by reconstructing all samples; Step 1.2: splice the sample features learned by the encoder with the feature vectors extracted from the task model, and then take the spliced vectors as the final feature representation of the samples.

3. The independence-enhanced based adversarial active learning method according to claim 1, wherein, The step 3 comprises the following steps: Step 3.1: In order to reduce the similarity between the selected samples and improve the independence between the samples, an independent sample selection module based on hierarchical clustering is designed; first, the categories of the samples are initialized, and each sample is a category, that is and the initial cluster number C=N is set, the merging speed is controlled by the hyperparameter p, and the total merging proportion is controlled by the hyperparameter s, then the hierarchical clustering of the feature representation of the samples from bottom to top is performed, and the pseudo-label y i of x i is constantly updated, in the clustering process, the mean of the Euclidean distances of all samples in the two clusters is used as the distance measurement of the two clusters: where C a and C b are two clusters, v i and v j are the concatenated vectors of the sample pairs in the two clusters, n a and n b are the number of samples in C a and C b respectively. Step 3.2: after clustering, the samples are selected in descending order of prediction value, and the selected samples come from different clusters; Step 3.3: finally, the selected K samples are labeled by Oracle, which considers the amount of single sample information and also takes into account the independence between samples, and finally, the labeled samples are added to the labeled pool and the model is updated and trained.

Images