Active learning method based on maximum mean difference adversarial

Abstract

The present invention discloses an active learning method based on maximum mean difference adversarial learning, which obtains an unlabeled image data set; for each image data in the unlabeled image data set, a feature extractor is used to extract a feature map of the image data, and the feature map is respectively sent to a classifier and a discriminator; the uncertainty score of the feature map is determined by the classifier, and the difference score of the feature map is determined by the discriminator; the query score of the feature map is calculated according to the uncertainty score and the difference score; and an unlabeled image is selected based on the query score of each image data; the present invention simultaneously utilizes the uncertainty and difference indicators of active learning to achieve the optimal performance of the machine learning model with minimal manpower cost, and uses a method based on maximum mean difference adversarial learning to perform adversarial training on the discriminator, so that the image has better diversity.

Description

Technical Field

[0001] The present invention belongs to the technical field of image classification, and in particular relates to an active learning method based on maximum mean difference adversarial learning. Background Art

[0002] Image classification is an image processing method that distinguishes different categories of targets based on the different features reflected in the image information. It uses computers to perform quantitative analysis on images and classify the image or each pixel or area in the image into one of several categories to replace human visual interpretation.

[0003] Active learning is a special case of machine learning. Its purpose is to maximize model performance gains while annotating as few examples as possible. For example, before an exam, teachers often spend more time with students reviewing key and difficult content, while spending less time reviewing less important content. Active learning is similar to a teacher selecting key and difficult content during class and reviewing it repeatedly until all the key content is reviewed. Students then complete the remaining less important content independently after class. In technical terms, active learning uses image selection strategies to select the most difficult images to classify—the most representative images—for manual verification, review, and annotation. The manually annotated data is then used for supervised or semi-supervised learning to further improve model performance. Active learning reduces the labor cost of annotation and improves the efficiency of engineering practice. To date, many research results have been published on active learning methods. These methods have good performance and are easy to implement. They can significantly reduce the number of annotated images, which is of great significance in reducing annotation costs.

[0004] Uncertainty sampling is a query method that extracts difficult-to-distinguish images from an image dataset and provides them to business experts or annotators for annotation, thereby rapidly improving the effectiveness of the method. The key to uncertainty sampling methods lies in describing the uncertainty of the image. Common methods include minimum confidence, edge sampling, and entropy.

[0005] The above method relies on the uncertainty of the model to select the most uncertain image. However, the uncertainty-based strategy is easily affected by outliers that belong to a different distribution from the labeled images, which in turn affects the image selection accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide an active learning method based on maximum mean difference adversarial learning to reduce the influence of outliers and select unlabeled images through uncertainty and maximum mean difference to improve the image selection accuracy.

[0007] The present invention adopts the following technical solution: an active learning method based on maximum average difference confrontation, comprising the following steps:

[0008] Obtain an unlabeled image dataset;

[0009] For each image data in the unlabeled image dataset, a feature extractor is used to extract the feature map of the image data, and the feature map is sent to the classifier and discriminator respectively;

[0010] The uncertainty score of the feature map is determined by the classifier, and the difference score of the feature map is determined by the discriminator;

[0011] Calculate the query score of the feature graph based on the uncertainty score and the difference score;

[0012] Unlabeled images are selected based on the query score of each image data.

[0013] Furthermore, determining the uncertainty score of the feature map by the classifier includes:

[0014] Determine the probability value of the feature map belonging to each image category through the classifier;

[0015] The lowest probability value is selected as the uncertainty score of the feature map.

[0016] Furthermore, the query score of the feature graph is calculated based on the uncertainty score and the difference score, including:

[0017] Query score = α*uncertainty score - (1-α)*difference score,

[0018] Among them, α is the score weight.

[0019] Furthermore, the feature extractor, classifier and discriminator are jointly trained using the unlabeled training image dataset and the labeled training image dataset.

[0020] Further, through Train the feature extractor, where α1 is the weight of the feature extractor, θ F is the feature extractor, W is the weight matrix of the feature extractor, (x L ,y L ) is the Lth training sample in the labeled training image dataset, x L is the training image, y L is x L Tags, Represents the mathematical expectation of the difference between the actual output and the expected output of the Lth training sample, represents the normalized sample probability distribution, σ represents the Isomax function, F(x L) represents the features of the labeled training samples after they are input into the feature extractor, and T represents the temperature parameter of the neural network.

[0021] Further, through Train the classifier, where α2 is the weight of the classifier, θ C is the classifier, x i represents the i-th labeled training sample, x i′ represents another labeled training sample that is different from the i-th labeled training sample, y j represents the jth unlabeled training sample, y j′ represents another unlabeled training sample that is different from the jth unlabeled training sample.

[0022] Further, through Train the discriminator, α3 is the weight of the discriminator, θ D is the discriminator.

[0023] Furthermore, the isomax function is used in the classifier to determine the probability value.

[0024] Another technical solution of the present invention: an active learning method based on maximum mean difference confrontation, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned active learning method based on maximum mean difference confrontation is implemented.

[0025] The beneficial effects of the present invention are as follows: the present invention simultaneously utilizes the uncertainty and diversity indicators of active learning to achieve the optimal performance of the machine learning model with minimal manpower cost, and uses a maximum average difference adversarial method to perform adversarial training on the discriminator to make the image have better diversity. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A schematic diagram of a flow chart of an active learning method based on maximum mean difference adversarial learning according to an embodiment of the present invention;

[0027] Figure 2 A flowchart of training and application combined with an embodiment of the present invention;

[0028] Figure 3 This is a comparison chart of the image sample accuracy rates of the various methods selected in the verification embodiment of the present invention. DETAILED DESCRIPTION

[0029] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] For binary or multi-classification models, they are usually able to score each image and determine which category it is more likely to belong to. For example, in a binary classification scenario, two data are predicted by a classifier, and the predicted probabilities for the two categories are: (0.9, 0.1) and (0.51, 0.49). In this case, the probability of the first image being judged as the first category is 0.9, and the probability of the second image being judged as the first category is 0.51. Therefore, the second image is obviously more "difficult" to distinguish and, therefore, more worthy of further annotation. The so-called minimum confidence method is to select those images with the smallest maximum probability for annotation.

[0031] Edge sampling involves selecting images that are easily classified into either class, or where the probability of the images being classified into the two classes is similar. Edge sampling involves selecting images where the difference between the model's highest and second-highest probability predictions is minimal. Specifically, for binary classification problems, the minimum confidence method and edge sampling are equivalent.

[0032] In mathematics, entropy can be used to measure the uncertainty of a system. A higher entropy indicates greater uncertainty, while a lower entropy indicates less uncertainty. Therefore, in binary or multi-class classification scenarios, images with higher entropy can be selected as candidates for labeling. Compared to the minimum confidence method and edge sampling method, the entropy method considers all possible class decisions made by the model for a given image. While the minimum confidence method only considers the highest probability, the edge sampling method considers both the highest and the second highest probability.

[0033] Ian J. Goodfellow first proposed the adversarial network model in 2014. In just two years, this model has achieved remarkable results in the field of deep learning generative models. It proposes a new framework for estimating generative models using an adversarial process. Compared to previous methods, this model can be considered a breakthrough in unsupervised representation learning. Adversarial networks consist of two models: a generative model and a discriminative model. The generative model learns the joint distribution of observed images, while the discriminative model learns the distribution of unobserved variables given the observed variable x. During adversarial training, the discriminator continuously learns to distinguish between real data (i.e., images), while the generator continuously learns to create fake data (i.e., images) that resembles real data (i.e., images) in order to deceive the discriminator. The final training result is that the discriminator can effectively distinguish between real and fake data, while the generator creates fake data that closely resembles real data. In active learning, instead of using real and fake data, labeled and unlabeled images are used for discrimination to identify those unlabeled images that are most dissimilar to the labeled ones.

[0034] This paper proposes an active learning method based on maximum mean discrepancy adversarial learning. Using a maximum mean discrepancy adversarial network, this method implicitly selects the most valuable images, eliminating the need for manually pre-setting criteria for image judgment. This method exploits the uncertainty and variability of images in an adversarial manner, combining the advantages of two approaches in active learning.

[0035] By adopting the maximum mean difference (MMD), a dissimilarity metric suitable for high-dimensional data, to measure the difference between the distributions of unlabeled and labeled images, and training the discriminator with the MMD loss, the accuracy of active learning is improved. Compared with the most common H-divergence, MMD can better capture the diversity of images.

[0036] Machine learning methods can still fall short in providing useful estimates of predictive uncertainty. A major branch of active learning involves iteratively selecting the most valuable images to annotate based on uncertainty. This approach relies on the classification probabilities predicted by the task model to determine the model's uncertainty about the image. However, these probabilities provided by the task model are unreliable because neural networks using a softmax classifier cannot recognize out-of-distribution images. Therefore, replacing softmax with isomax, a classifier that is more sensitive to out-of-distribution images, can improve the accuracy of active learning uncertainty measures.

[0037] Meanwhile, active learning aims to develop efficient methods by querying for the most representative images, annotated by human annotators. Current active learning techniques either rely on model uncertainty to select the most uncertain samples, or use clustering or reconstruction to select the most diverse unlabeled examples. Our method exploits image uncertainty and diversity in a score calculation, combining the strengths of both approaches in the active learning literature.

[0038] A specific embodiment of the present invention discloses an active learning method based on maximum average difference adversarial learning, comprising the following steps: obtaining an unlabeled image dataset; for each image data in the unlabeled image dataset, using a feature extractor to extract a feature map of the image data, and sending the feature map to a classifier and a discriminator respectively; determining the uncertainty score of the feature map by the classifier, and determining the difference score of the feature map by the discriminator; calculating the query score of the feature map based on the uncertainty score and the difference score; and selecting an unlabeled image based on the query score of each image data.

[0039] The present invention simultaneously utilizes the uncertainty and diversity indicators of active learning to achieve the optimal performance of the machine learning model with minimal manpower cost, and uses a maximum average difference adversarial method to perform adversarial training on the discriminator to make the image have better diversity.

[0040] Active learning methods are divided into two phases: network parameter optimization and query batch selection for image sample queries. The network training phase is a min-max optimization problem. More specifically, it involves maximizing a critical function to distinguish between labeled and unlabeled empirical distributions using a loss calculated as the maximum mean difference, while the feature extractor function has the opposite goal of minimizing the empirical distribution. Furthermore, because this method utilizes both labeled and unlabeled samples to jointly train and generate feature maps, whereas previous active learning methods only used labeled data to generate features, these unlabeled samples are also used to train the task model itself to improve classification accuracy. This is a semi-supervised process.

[0041] In the query stage of selecting samples, two criteria are used. The first criterion is the similarity with existing labeled data, and the second criterion is the minimum confidence. Combining the two criteria, the unlabeled data is ranked, which is called the label degree score. This can guide the query strategy function not only to query those image samples that look dissimilar to those seen before, but also to rank the most uncertain image samples and select the corresponding image samples for manual labeling.

[0042] In another embodiment of the present invention, the feature extractor, the classifier and the discriminator are jointly trained by the unlabeled training image dataset and the labeled training image dataset.

[0043] Specifically, through Train the feature extractor, where α1 is the weight of the feature extractor, θ F is the feature extractor, W is the weight matrix of the feature extractor, (x L ,y L ) is the Lth training sample in the labeled training image dataset, x L is the training image, y L is x L Tags, Represents the mathematical expectation of the difference between the actual output and the expected output of the Lth training sample, represents the normalized sample probability distribution, σ represents the Isomax function, F(x L ) represents the features of the labeled training samples after they are input into the feature extractor, T represents the temperature parameter of the neural network, and the T in the upper right corner of W is the transpose.

[0044] In addition, through Train the classifier, where α2 is the weight of the classifier, θ C is the classifier, x i represents the i-th labeled training sample, x i′represents another labeled training sample that is different from the i-th labeled training sample, y j represents the jth unlabeled training sample, y j′ represents another unlabeled training sample that is different from the jth unlabeled training sample.

[0045] More specifically, through Train the discriminator, α3 is the weight of the discriminator, θ D is the discriminator.

[0046] Regarding the loss function of the classifier, the standard cross entropy loss is used, that is, Optimize. Use a cosine similarity based classifier, denoted as C, and weight matrix It takes the normalized features as input and maps them to K class prototype vectors [w1,w2,…,w K ], where K is the total number of classes in the dataset. The output of the classifier is converted to a probability value using the isomax function (σ) Isomax is a classification loss function that is more sensitive to outlier samples. Its main formula is as follows:

[0047]

[0048] Among them, f θ (x) represents the high-level features (embedding) associated with sample x, Represents the prototype associated with class j, E s is a constant. Isomax is sensitive to outliers without affecting the performance of the classifier.

[0049] Regarding the discriminator, the maximum mean difference is abbreviated as MMD. This is an integral probability metric that uses a class of witness functions to distinguish between distributions P and Q and selects the function with the largest expected difference from P and Q:

[0050]

[0051] The basic assumption of maximum mean difference is that if, for all functions f with the sample space generated by the distributions as input, the means of the corresponding images on f generated by enough samples from two distributions are equal, then the two distributions can be considered to be the same. It is now commonly used to measure the similarity between two distributions.

[0052] Specifically, the statistical test method based on the maximum mean difference refers to the following method: based on samples from two distributions, by finding a continuous function f in the sample space, finding the mean of the function values of samples with different distributions on f, and by taking the difference between the two means, the average difference of the two distributions corresponding to f can be obtained. Find an f that makes this average difference have a maximum value, and the maximum average difference is obtained. Finally, the maximum average difference is taken as the test statistic to determine whether the two distributions are the same. If this value is small enough, the two distributions are considered to be the same, otherwise they are considered to be different. At the same time, this value is also used to judge the degree of similarity between the two distributions. If F represents a set of continuous functions in the sample space, then the maximum average difference can be expressed by the following formula:

[0053]

[0054] In this method, the labeled dataset and the unlabeled dataset are mapped to the reproducing kernel Hilbert space, which is a kernel learning method. i ,y j , x i′ ,y j′ , by finding the mapping function k on the sample space, finding the mean of the function value of samples with different distributions on k, and by taking the difference between the two means, we can get the mean difference of the two distributions corresponding to k. Find a k that makes this mean difference have the maximum value, and we get the maximum average difference. Finally, take the maximum average difference as the test statistic to determine whether the two distributions are the same. If this value is small enough, the two distributions are considered to be the same, otherwise they are considered to be different. A simpler understanding is: find the distance between the means of two piles of data in high-dimensional space. Its loss function is as follows. The maximum average difference loss is as follows shown.

[0055] Based on the above, the loss can be extended to the deep representation learning scenario. Inspired by transfer learning and adversarial generative networks, the maximum-minimum optimization problem is used to train the network model. The maximum-minimum formula is as follows:

[0056]

[0057] Among them, θ F ,θ C and θ D are the parameters corresponding to the feature extractor, task predictor, and distribution discriminator; R is the loss function of the task predictor, and MMD is the (min-max) loss of the maximum mean difference discriminator.

[0058] During the training phase, all observed data is used to optimize the network parameters. Labeled and unlabeled samples are passed through the feature extractor simultaneously. Labeled samples are used to train the feature extractor and task predictor, while unlabeled samples are used to train the distribution discriminator.

[0059] The sampling strategy is shown below. The key idea of the method is to select samples for labeling using two criteria, namely, calculating the query score of the feature map based on the uncertainty score and the difference score. Specifically, the following are included:

[0060] Query score = α*uncertainty score - (1-α)*difference score,

[0061] Among them, α is the score weight.

[0062] Specifically, the uncertainty score uses the probability associated with the discriminator's prediction as a score to rank samples according to their differences, which can be interpreted as how similar they are to the labeled data. The closer the probability is to 0, the more confident the discriminator is that it came from the unlabeled pool. That is, determining the uncertainty score of a feature map by a classifier includes: determining the probability value of the feature map belonging to each image category by the classifier; selecting the lowest probability value as the uncertainty score of the feature map. The unlabeled data with the lowest prediction confidence method used by the classifier is used as a measure of uncertainty scoring. The higher the probability of the lowest predicted label, the less confident the task predictor is in its prediction.

[0063] Applying this method to image classification tasks, we first use a relevant image dataset, feed it into a feature extractor, and perform feature extraction and normalization. Next, we use the cross-entropy loss to train the feature extractor and classifier, and the maximum mean difference loss to train the distribution discriminator. During the query phase, we use the trained classifier and adversarial discriminator to query image samples using the aforementioned query strategy. The most informative images are selected, manually annotated, and added to the annotated data pool until the annotation budget is reached.

[0064] In summary, if Figure 1 As shown, the present invention first feeds labeled and unlabeled samples into a feature extractor for feature extraction. A classifier is then used to maximize the entropy of the unlabeled data, thereby selecting images for which the machine learning model has the least uncertainty. A discriminator uses the features generated by the feature extractor to predict which pool each sample belongs to (i.e., labeled or unlabeled image data), thereby selecting image samples that differ most from the labeled image samples. Finally, a hybrid query strategy is used to combine the uncertainty and diversity of the samples.

[0065] In addition, if Figure 2As shown, the active learning method in the embodiment of the present invention is an iterative interactive training process, which mainly consists of five core parts, including: an unlabeled sample pool, an active learning sample selection strategy, users, a training image sample dataset, and a classifier model.

[0066] Active learning combines these steps into a single process. First, the dataset is loaded into a feature extractor network. Common feature extractors include VGG-16 and Resnet-18. This extractor extracts features from both labeled and unlabeled samples. Next, the features are normalized for further processing and fed into a classifier and discriminator. The classifier identifies the image samples that the machine learning model is most uncertain about, while the discriminator identifies the samples that differ most from the pool of labeled image data.

[0067] The training process of the entire network using the relevant loss is as follows:

[0068] Input: Samples and labels (x L ,y L ), the samples of the unlabeled sample pool U (x U ), the initial feature extractor model θ F , the initial classifier model θ C , the initial discriminator model θ D , hyper parameters: Epoch (number of iterations), α1, α2, α3.

[0069] 1: For the initial Epoch round n = 1 to Epoch; 2: Calculate L CE ; 3: Calculation 4: 5: 6: 7:n=n+1. Output: θ after training and updating in the manner of 4, 5, and 6 F ,θ C ,θ D .

[0070] Finally, we use active learning as a query strategy to select image samples that require annotation by annotators in related fields. The uncertainty score is the probability of the classifier's prediction, using a minimum confidence approach. The difference score is the probability of the discriminator's prediction.

[0071] Select b image samples with the largest total scores, where b is the number of samples set for each batch. After selecting the image samples, update the model performance, the unlabeled sample pool, and the labeled data pool in an iterative training method until the target model reaches the preset performance or no labeled data is provided. In each iteration, the number of labeled samples increases, and the performance of the model also improves accordingly (ideally). In practical applications, the accuracy of the annotator should be guaranteed as much as possible to alleviate the situation where the model learns bias in the early stages of training (specifically caused by incorrectly labeled samples).

[0072] To verify the effectiveness of the present invention, the method was tested on the public dataset Cifar-10. Cifar-10 consists of 60,000 32×32 images with ten categories. The training and test sets contain 50,000 and 10,000 images, respectively. In Cifar-10, the initial size of the classification experiment was selected to be 1000, and the budget scale of active learning was 1000. For the query strategy, α was set to 0.1. According to the experimental results, the present method significantly improves the performance of few-shot classification tasks.

[0073] Regarding the experimental results, such as Figure 3 As shown, the horizontal axis represents the number of labeled image samples, the vertical axis represents the classification accuracy, and a~f, etc. represent the various methods marked on the figure. It can be seen from the figure that the selection accuracy corresponding to the method of the present invention is better than other methods in selecting the correct image accuracy. Experiments have shown that this method has good accuracy in image classification tasks compared with more classic or currently popular methods.

[0074] The present invention also discloses an active learning method based on maximum average difference confrontation, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned active learning method based on maximum average difference confrontation is implemented.

[0075] The aforementioned device may be a computing device such as a desktop computer, laptop, PDA, or cloud server. The device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the device may include more or fewer components, or a combination of certain components, or different components, and may also include, for example, input / output devices, network access devices, and the like.

[0076] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0077] In some embodiments, the memory may be an internal storage unit of the device, such as a hard disk or memory of the device. In other embodiments, the memory may also be an external storage device of the device, such as a plug-in hard disk equipped on the device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card, etc. Furthermore, the memory may include both an internal storage unit of the device and an external storage device. The memory is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory may also be used to temporarily store data that has been output or is about to be output.

[0078] It should be noted that the specific content of the above-mentioned device is based on the same concept as the embodiment of the method of the present invention. Its specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

Claims

1. An active learning method based on maximum mean difference adversarial learning, characterized in that: The following steps are involved: Obtain an unlabeled image dataset; For each image data in the unlabeled image dataset, a feature extractor is used to extract a feature map of the image data, and the feature map is respectively fed into a classifier and a discriminator; Determining an uncertainty score of the feature map by the classifier, and determining a difference score of the feature map by the discriminator; Calculating a query score for the feature graph based on the uncertainty score and the difference score; selecting an unlabeled image based on a query score for each of the image data; Determining the uncertainty score of the feature map by the classifier includes: Determining, by the classifier, a probability value that the feature map belongs to each image category; Selecting the lowest probability value as the uncertainty score of the feature map; Calculating the query score of the feature graph according to the uncertainty score and the difference score includes: Query score = α * Uncertainty score - (1- α ) * difference score, in, α is the score weight; The feature extractor, classifier and discriminator are jointly trained using an unlabeled training image dataset and a labeled training image dataset; pass The feature extractor is trained, wherein is the weight of the feature extractor, is the feature extractor, , is the weight matrix of the feature extractor, ( , ) is the first L training samples, For training images, for Tags, Indicates the L The mathematical expectation of the difference between the actual output and the expected output of the training samples, represents the normalized sample probability distribution, represents the isomax function, Represents the features of the labeled training samples after inputting the feature extractor, T represents the temperature parameter of the neural network; pass ( + ) trains the classifier, wherein, is the weight of the classifier, is the classifier, , represents the mapping function to the regenerated Hilbert space, N represents the number of images in the labeled training image dataset, M represents the number of images in the unlabeled training image dataset, Indicates the i labeled training samples, Indicates another i labeled training samples, different labeled training samples, Indicates the j unlabeled training samples, Indicates another j unlabeled training samples are different unlabeled training samples.

2. The active learning method based on maximum mean difference confrontation as claimed in claim 1, characterized in that: pass Train the discriminator, is the weight of the discriminator, is the discriminator.

3. The active learning method based on maximum mean difference adversarial learning as claimed in claim 1, characterized in that: The isomax function is used in the classifier to determine the probability value.

4. An active learning device based on maximum mean difference adversarial learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the processor implements the active learning method based on maximum mean difference adversarial learning as described in any one of claims 1 to 3.

Images