An active learning method with uncertainty and representativeness

Abstract

The application discloses an active learning framework with uncertainty and representation, which is characterized in that: firstly, a self-supervised pre-task learner based on contrast learning is set to sort and batch process samples of the whole data pool; then, the batched samples are subjected to contrast clustering under a dual framework of instance level and cluster level; then, the abnormal value scores of all samples are calculated based on a K nearest neighbor algorithm, and the outlier samples in all clusters are sampled. The application can better classify unlabeled data, realize clustering under a dual contrast framework of instance level and cluster level, has a better effect, shows strong advantages in the first iteration, and can solve the cold start problem of active learning.

Description

Technical Field

[0001] This invention relates to an active learning method that combines uncertainty and representativeness, applicable to fields in computer vision that require large amounts of labeled data. Background Technology

[0002] In recent years, the continuous advancement of deep learning methods has led to significant success in the field of computer vision, with remarkable progress in tasks such as image classification, object detection, and semantic segmentation. However, traditional supervised learning relies on large-scale data annotation, which, for specific tasks such as medical image classification and lesion segmentation, agricultural image classification, and industrial defect detection, requires annotation by industry experts. This significantly increases the time and financial costs of model training. Active learning, as a weakly supervised method, can achieve optimal model performance by annotating a subset of all samples. This subset is generated by selecting the samples that best improve model performance within a fixed labeling budget.

[0003] Traditional active learning methods are divided into two categories: representativeness-based and uncertainty-based. Representativeness-based active learning tends to select samples that represent the entire data distribution, but may overlook special cases or boundary samples within the distribution. Uncertainty-based methods, on the other hand, focus on selecting samples with high uncertainty, thus rapidly improving model performance, but suffer from drawbacks such as repeated sampling and difficulty in fitting the overall data distribution. Therefore, researching a hybrid selection strategy for active learning that better combines the advantages of both methods and overcomes the problems of single sampling methods is of great significance. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of existing active learning methods, this invention provides an active learning method that combines uncertainty and representativeness, with low labeling cost, good effect and easy use.

[0005] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0006] An active learning method that combines uncertainty and representativeness first establishes a self-supervised pre-task learner based on contrastive learning. For the entire data pool The samples are sorted and processed in batches. Then, the batched samples are compared and clustered under a dual framework of instance level and cluster level. Then, the outlier scores of all samples are calculated based on the K-nearest neighbor algorithm, and outlier samples in all clusters are sampled. The present invention includes the following steps:

[0007] S1. A self-supervised pre-task learner using contrastive learning. For unlabeled data pools In Sample Perform random enhancement to obtain a sample pair , , Indicates unlabeled data pool The first in One sample, and Indicates the sample Two samples obtained by performing random augmentation;

[0008] S2, take the sample and The inputs are respectively fed into the feature extraction encoder. , to obtain the feature vector and ;

[0009] S3, The feature vector and The inputs are fed into the multilayer perceptron. To obtain the perception vector and ;

[0010] S4. Use the contrastive loss function calculate Contrast loss First, adjust the unlabeled data pool according to the magnitude of the contrast loss. In The samples are sorted in reverse order, and then arranged in order. The samples were divided into groups based on quantity. Groups, each group contains The nth sample, the th The samples in each group are represented as follows: , , , , This represents the floor function. Indicates the first The first group One sample;

[0011] S5. Use contrastive clustering outlier samplers to... Perform data augmentation, the first Groups were obtained a sample pair , and Indicates the sample Two samples obtained after data augmentation;

[0012] S6, will and The inputs are respectively fed into the feature extraction encoder. , to obtain the feature vector and ;

[0013] S7. Transform the feature vector and The inputs are fed into the multilayer perceptron. To obtain the perception vector and ;

[0014] S8, regarding the first In each group a sample pair Contrastive clustering is performed using both instance-level and cluster-level contrastive heads to achieve contrastive clustering under both instance-level and cluster-level frameworks. Instance-level contrastive loss and cluster-level contrastive loss are calculated.

[0015] S9. Summing the instance-level contrast loss and the cluster-level contrast loss yields the overall loss;

[0016] S10. For clusters obtained by comparison clustering using cluster-level comparison head, use the K-nearest neighbor algorithm to calculate the Euclidean distance from each sample in the cluster to the cluster center, filter out the outlier samples of each cluster, and calculate the outlier score of each outlier sample.

[0017] S11. Sort all outlier samples selected from each cluster in reverse order based on their outlier scores, and select the outlier samples with the lowest outlier scores from each cluster. Outlier samples are labeled as target samples that possess both uncertainty and representativeness.

[0018] Specifically, in step S1, the unlabeled data pool... In Sample Two methods were used for random augmentation: random cropping, rotation, coloring, and blurring. The total number of samples after random augmentation was [number missing]. One, including One negative sample.

[0019] Specifically, in step S4, the comparison loss function is... Represented as:

[0020]

[0021] in: Indicates the use of the contrastive loss function calculate The comparative loss, Represents the cosine similarity function. This represents the temperature parameter.

[0022] Specifically, in step S5, a contrastive clustering outlier sampler is used to sample... Two data augmentation methods were used: ResizedCrop, ColorJitter, Grayscale, HorizontalFlip, and GaussianBlur. The total number of samples after data augmentation was [number missing]. One, including One negative sample.

[0023] Specifically, this active learning method and self-supervised pre-task learner All use ResNet18 as the backbone network; in the self-supervised pre-task learner During training, the batch size was set to 128, the initial learning rate was set to 0.1, the momentum encoder size was set to 0.9, and the weight decay value was set to 5e-4.

[0024] Specifically, in step S8, the instance-level comparison header is used to compare the first... In each group Perform comparative clustering on each sample pair, set the dimension of the row space to 128, and include instance-level temperature parameters. Set to a fixed value of 0.5, and use cosine distance to measure the similarity between sample pairs:

[0025]

[0026] The sample loss under different data augmentation methods is calculated using the following formula:

[0027]

[0028]

[0029] The instance-level contrastive loss is calculated using the following formula:

[0030]

[0031] in: express and Cosine similarity between them; Indicates a data augmentation method The sample loss, Indicating another data augmentation method The sample loss; Indicates the first In each group Instance-level contrastive loss for each sample pair.

[0032] Specifically, in step S8, the cluster-level comparison head is used to compare the first... In each group When performing comparative clustering on pairs of samples, first... one sample Mapping to a dimension equals the number of clusters In the space ,Will one sample It is also mapped into this space to obtain In this space, each element in the feature vector of a sample is considered as the probability that it belongs to the corresponding cluster, thus assigning a soft label to each cluster; simultaneously, in this space, the dimension of the column space is set to the number of clusters. Cluster-level temperature parameters Set a fixed value of 1.0, use the Adam optimizer, and set the initial learning rate to 0.0003; measure the similarity between sample pairs using cosine distance:

[0033]

[0034] The sample loss under different data augmentation methods is calculated using the following formula:

[0035]

[0036]

[0037] The cluster-level contrast loss is calculated using the following formula:

[0038]

[0039] in: express and Cosine similarity between them; Indicates a data augmentation method The sample loss, Indicating another data augmentation method The sample loss; Indicates the first In each group Cluster-level contrastive loss for each sample pair The entropy for cluster assignment probability is set to avoid assigning too many samples to the same cluster.

[0040] Specifically, in step S9, the overall loss is expressed as .

[0041] Specifically, in step S10, the outlier score of an outlier sample calculated using the K-nearest neighbor algorithm is the reciprocal of the Euclidean distance of the outlier sample. The larger the Euclidean distance, the lower the outlier score, and the more likely it is to be an outlier sample.

[0042] The active learning method provided by this invention, which combines uncertainty and representativeness, includes a self-supervised pre-task learner. and outlier sampler; among which, self-supervised pre-task learner Responsible for the data pool The samples are sorted and divided into small batches of data pools; the outlier sampler is responsible for comparative clustering and sampling of the sorted small batches of data pools. Specifically, the self-supervised pre-task learner First, the data pool The data pool is then randomly augmented, the contrast loss of the augmented samples is calculated, and finally, the data is allocated to the pool based on the magnitude of the contrast loss. Divided into Each small batch of data pools; the outlier sampler first performs data augmentation on the samples in the small batch data pool, then applies instance-level data headers and cluster-level data headers to perform comparative clustering on the samples, and then uses the K-nearest neighbor algorithm to calculate the outlier score of the samples and sample outlier samples.

[0043] Beneficial effects: The active learning method provided by this invention, which combines uncertainty and representativeness, adopts a sample collection strategy that combines a self-supervised pre-task learner with an outlier sampler compared to existing technologies. This strategy can better classify unlabeled data and achieve clustering under a dual comparison framework at the instance and cluster levels, resulting in superior performance. It demonstrates a strong advantage in the first iteration and can solve the cold start problem of active learning. This invention also has excellent sample selection capabilities and a wide range of applications. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the implementation process of the present invention;

[0045] Figure 2 This is a flowchart illustrating the self-supervised pre-task learner in this invention.

[0046] Figure 3 This is a schematic diagram of the outlier sampler of the present invention. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0048] An active learning approach that combines uncertainty and representativeness includes a self-supervised pre-task learner. and outlier sampler; among which, self-supervised pre-task learner Responsible for the data pool The samples are sorted and divided into small batches of data pools; the outlier sampler is responsible for comparative clustering and sampling of the sorted small batches of data pools. Specifically, the self-supervised pre-task learner First, the data pool The data pool is then randomly augmented, the contrast loss of the augmented samples is calculated, and finally, the data is allocated to the pool based on the magnitude of the contrast loss. Divided into Each small batch of data pools; the outlier sampler first performs data augmentation on the samples in the small batch data pool, then applies instance-level data headers and cluster-level data headers to perform comparative clustering on the samples, and then uses the K-nearest neighbor algorithm to calculate the outlier score of the samples and sample outlier samples.

[0049] In this case, active learning methods and self-supervised pre-task learners are used. All use ResNet18 as the backbone network; in the self-supervised pre-task learner During training, the batch size was set to 128, the initial learning rate was set to 0.1, the momentum encoder size was set to 0.9, and the weight decay value was set to 5e-4.

[0050] like Figure 1 The diagram shown is a flowchart of the implementation process of this case. The specific steps will be explained in detail below.

[0051] S1. A self-supervised pre-task learner using contrastive learning. For unlabeled data pools In Sample Perform random enhancement to obtain a sample pair , , Indicates unlabeled data pool The first in One sample, and Indicates the sample Two samples obtained by performing random augmentation.

[0052] like Figure 2 As shown, a self-supervised pre-task learner Includes a random enhancement module and a feature extraction encoder. and multilayer perceptron .

[0053] For unlabeled data pools In Sample Two methods were used for random augmentation: random cropping, rotation, coloring, and blurring. The total number of samples after random augmentation was [number missing]. One, including One negative sample.

[0054] S2, take the sample and The inputs are respectively fed into the feature extraction encoder. , to obtain the feature vector and .

[0055] S3, The feature vector and The inputs are fed into the multilayer perceptron. To obtain the perceptual vector and .

[0056] S4. Use the contrastive loss function calculate Contrast loss First, adjust the unlabeled data pool according to the magnitude of the contrast loss. In The samples are sorted in reverse order, and then arranged in sequence. The samples were divided into groups based on quantity. Groups, each group contains The nth sample, the th The samples in each group are represented as follows: , , , , This represents the floor function. Indicates the first The first group One sample.

[0057] Contrast loss function Represented as:

[0058]

[0059] in: Indicates the use of the contrastive loss function calculate The comparative loss, Represents the cosine similarity function. This represents the temperature parameter.

[0060] S5. Use contrastive clustering outlier samplers to... Perform data augmentation, the first Groups were obtained a sample pair , and Indicates the sample Two samples obtained after data augmentation.

[0061] like Figure 3 As shown, the outlier sampler includes a data augmentation module, an instance-level comparison head module, and a cluster-level comparison head module.

[0062] Outlier sampler using contrastive clustering Two data augmentation methods were used: ResizedCrop, ColorJitter, Grayscale, HorizontalFlip, and GaussianBlur. The total number of samples after data augmentation was [number missing]. One, including One negative sample.

[0063] S6, will and The inputs are respectively fed into the feature extraction encoder. , to obtain the feature vector and .

[0064] S7. Transform the feature vector and The inputs are fed into the multilayer perceptron. To obtain the perceptual vector and .

[0065] S8, regarding the first In each group a sample pair Contrastive clustering is performed using both instance-level and cluster-level contrastive heads, achieving contrastive clustering within a dual framework of instance-level and cluster-level. Instance-level contrastive loss and cluster-level contrastive loss are then calculated.

[0066] Use instance-level comparison headers to compare the first In each group Perform comparative clustering on each sample pair, set the dimension of the row space to 128, and include instance-level temperature parameters. Set to a fixed value of 0.5, and use cosine distance to measure the similarity between sample pairs:

[0067]

[0068] The sample loss under different data augmentation methods is calculated using the following formula:

[0069]

[0070]

[0071] The instance-level contrastive loss is calculated using the following formula:

[0072]

[0073] in: express and Cosine similarity between them; Indicates a data augmentation method The sample loss, Indicating another data augmentation method The sample loss; Indicates the first In each group Instance-level contrastive loss for each sample pair.

[0074] Use cluster-level comparison head to compare the first In each group When performing comparative clustering on pairs of samples, first... one sample Mapping to a dimension equals the number of clusters In the space ,Will one sample It is also mapped into this space to obtain In this space, each element in the feature vector of a sample is considered as the probability that it belongs to the corresponding cluster, thus assigning a soft label to each cluster; simultaneously, in this space, the dimension of the column space is set to the number of clusters. Cluster-level temperature parameters Set a fixed value of 1.0, use the Adam optimizer, and set the initial learning rate to 0.0003; measure the similarity between sample pairs using cosine distance:

[0075]

[0076] The sample loss under different data augmentation methods is calculated using the following formula:

[0077]

[0078]

[0079] The cluster-level contrast loss is calculated using the following formula:

[0080]

[0081] in: express and Cosine similarity between them; Indicates a data augmentation method The sample loss, Indicating another data augmentation method The sample loss; Indicates the first In each group Cluster-level contrastive loss for each sample pair The entropy for cluster assignment probability is set to avoid assigning too many samples to the same cluster.

[0082] S9. Summing the instance-level contrast loss and the cluster-level contrast loss yields the overall loss. .

[0083] S10. For clusters obtained by comparative clustering using cluster-level comparison heads, the K-nearest neighbor algorithm is used to calculate the Euclidean distance from each sample in the cluster to the cluster center, outlier samples in each cluster are screened out, and outlier scores of each outlier sample are calculated.

[0084] The outlier score of an outlier sample calculated using the K-nearest neighbors algorithm is the reciprocal of the Euclidean distance to that outlier sample. The larger the Euclidean distance, the lower the outlier score, and the more likely it is to be an outlier sample.

[0085] S11. Sort all outlier samples selected from each cluster in reverse order based on their outlier scores, and select the outlier samples with the lowest outlier scores from each cluster. Outlier samples are labeled as target samples that possess both uncertainty and representativeness.

[0086] S12. Perform iterative training by sampling in multiple batches and test on the entire test set until the label budget is reached.

[0087] This proposed solution was compared with six different baselines, covering random sampling, uncertainty-based sampling, and representative sampling methods, to evaluate its effectiveness on three challenging image classification benchmark datasets. The six different baselines are:

[0088] 1. Random: A very simple sample selection method, which randomly samples K samples from the unlabeled data pool in each round for iterative training.

[0089] 2. CoreSet: An active learning method based on core sets, which selects the most representative samples from the entire unlabeled data pool. It is a classic active learning method based on representativeness.

[0090] 3. CoreGCN: One of the most advanced representative active learning methods, it uses the high-order graph node information of CoreSet for a new sampling technique by introducing latent spatial distance, and uses risk minimization between core sets on GCN features.

[0091] 4. UncertainGCN: One of the most advanced uncertainty-based active learning methods, it utilizes GCN and applies sampling techniques to graph-induced node embedding, sampling based on confidence scores.

[0092] 5. VAAL: A pool-based semi-supervised active learning algorithm that uses variational autoencoders (VAEs) and adversarial networks to learn a latent space to distinguish between unlabeled and labeled data.

[0093] 6. Lloss: A simple yet efficient active learning method that adds a loss prediction module to the target network to predict the target loss of unlabeled inputs.

[0094] In the comparative experiments, the effectiveness of the proposed method was evaluated by mixed sampling on three publicly available datasets from the internet, including two RGB datasets, CIFAR10 and SVHN, and a grayscale image dataset, FashionMnist. These image datasets are commonly used for active learning training. Each dataset was sampled 10 times, with a fixed budget of 1000 samples per sample, resulting in a final labeled sample count of 10000 for all experiments. The clustering module was trained for 100 epochs per sampling, while the main task learner was trained for 200 epochs.

[0095] The same hyperparameters were used for all experiments and datasets. All experiments were run on Ubuntu 18.04 with an Intel(R) Xeon(R) E5-2686 v4 CPU, 64GB of RAM, and an NVIDIA GeForce RTX4090 GPU with 24GB of VRAM. PyTorch version 1.12.1 was used for all experiments. The experimental results are shown in the following tables.

[0096] Table 1. Performance of different methods on the CIFAR10 dataset

[0097]

[0098] Table 2 shows the performance of different methods on the FashionMnist dataset.

[0099]

[0100] Table 3 Performance of different methods on the SVHN dataset

[0101]

[0102] As can be seen from the various tables, this implementation demonstrated a strong advantage in the first iteration and solved the cold start problem of active learning. The experimental results were in a leading position in most iterations.

[0103] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. An active learning method that combines uncertainty and representativeness, characterized in that: Includes the following steps: S1. A self-supervised pre-task learner using contrastive learning. For unlabeled data pools In Sample Perform random enhancement to obtain a sample pair , , Indicates unlabeled data pool The first in One sample, and Indicates the sample Two samples obtained by performing random augmentation, wherein the samples are images; S2, take the sample and The inputs are respectively fed into the feature extraction encoder. , to obtain the feature vector and ; S3, The feature vector and The inputs are fed into the multilayer perceptron. To obtain the perception vector and ; S4. Use the contrastive loss function calculate Contrast loss First, adjust the unlabeled data pool according to the magnitude of the contrast loss. In The samples are sorted in reverse order, and then arranged in order. The samples were divided into groups based on quantity. Groups, each group contains The nth sample, the th The samples in each group are represented as follows: , , , , This represents the floor function. Indicates the first The first group One sample; S5. Use contrastive clustering outlier samplers to... Perform data augmentation, the first Groups were obtained a sample pair , and Indicates the sample Two samples obtained after data augmentation; S6, will and The inputs are respectively fed into the feature extraction encoder. , to obtain the feature vector and ; S7. Transform the feature vector and The inputs are fed into the multilayer perceptron. To obtain the perceptual vector and ; S8, regarding the first In each group a sample pair Contrastive clustering is performed using both instance-level and cluster-level contrastive heads to achieve contrastive clustering under both instance-level and cluster-level frameworks. Instance-level contrastive loss and cluster-level contrastive loss are calculated. S9. Summing the instance-level contrast loss and the cluster-level contrast loss yields the overall loss; S10. For clusters obtained by comparison clustering using cluster-level comparison head, use the K-nearest neighbor algorithm to calculate the Euclidean distance from each sample in the cluster to the cluster center, filter out the outlier samples of each cluster, and calculate the outlier score of each outlier sample. S11. Sort all outlier samples selected from each cluster in reverse order based on their outlier scores, and select the outlier samples with the lowest outlier scores from each cluster. Outlier samples are labeled as target samples that possess both uncertainty and representativeness.

2. The active learning method with both uncertainty and representativeness according to claim 1, characterized in that: In step S1, the unlabeled data pool In Sample Two methods were used for random augmentation: random cropping, rotation, coloring, and blurring. The total number of samples after random augmentation was [number missing]. One, including One negative sample.

3. The active learning method with both uncertainty and representativeness according to claim 1, characterized in that: In step S4, the comparison loss function Represented as: ， in: Indicates the use of the contrastive loss function calculate The comparative loss, Represents the cosine similarity function. This represents the temperature parameter.

4. The active learning method combining uncertainty and representativeness according to claim 1, characterized in that: In step S5, the outlier sampler of contrastive clustering is used to... Two data augmentation methods were used: ResizedCrop, ColorJitter, Grayscale, HorizontalFlip, and GaussianBlur. The total number of samples after data augmentation was [number missing]. One, including One negative sample.

5. The active learning method combining uncertainty and representativeness according to claim 1, characterized in that: This active learning method and self-supervised pre-task learner All use ResNet18 as the backbone network; in the self-supervised pre-task learner During training, the batch size was set to 128, the initial learning rate was set to 0.1, the momentum encoder size was set to 0.9, and the weight decay value was set to 5e-4.

6. The active learning method with both uncertainty and representativeness according to claim 1, characterized in that: In step S8, the instance-level comparison header is used to compare the first... In each group Perform comparative clustering on each sample pair, set the dimension of the row space to 128, and include instance-level temperature parameters. Set to a fixed value of 0.5, and use cosine distance to measure the similarity between sample pairs: ， The sample loss under different data augmentation methods is calculated using the following formula: ， ， The instance-level contrast loss is calculated using the following formula: ， in: express and Cosine similarity between them; Indicates a data augmentation method The sample loss, Indicating another data augmentation method The sample loss; Indicates the first In each group Instance-level contrastive loss for each sample pair.

7. The active learning method combining uncertainty and representativeness according to claim 1, characterized in that: In step S8, the cluster-level comparison head is used to compare the first... In each group When performing comparative clustering on pairs of samples, first... one sample Mapping to a dimension equals the number of clusters In the space ,Will one sample It is also mapped into this space to obtain ; In this space, each element in the feature vector of each sample is considered as the probability that it belongs to the corresponding cluster, thus assigning a soft label to each cluster; At the same time, in this space, the dimension of the column space is set to the number of clusters. Cluster-level temperature parameters Set a fixed value of 1.0, use the Adam optimizer, and set the initial learning rate to 0.0003; measure the similarity between sample pairs using cosine distance: ， The sample loss under different data augmentation methods is calculated using the following formula: ， ， The cluster-level contrast loss is calculated using the following formula: ， in: express and Cosine similarity between them; Indicates a data augmentation method The sample loss, Indicating another data augmentation method The sample loss; Indicates the first In each group Cluster-level contrastive loss for each sample pair The entropy for cluster assignment probability is set to avoid assigning too many samples to the same cluster.

8. The active learning method with both uncertainty and representativeness according to claim 1, characterized in that: In step S9, the overall loss is expressed as .

9. The active learning method with both uncertainty and representativeness according to claim 1, characterized in that: In step S10, the outlier score of an outlier sample calculated using the K-nearest neighbor algorithm is the reciprocal of the Euclidean distance of the outlier sample. The larger the Euclidean distance, the lower the outlier score, and the more likely it is to be an outlier sample.

Images