Debiasing learning method and system for new-old category competition in generalized category discovery

Abstract

The application discloses a debiased learning method and system for new-old category competition in generalized category discovery, and belongs to the technical field of artificial intelligence, comprising: extracting standardized features for all samples, setting a classification branch to output basic category prediction results, and independently setting an auxiliary debiased branch; constructing and maintaining two types of storage libraries, constructing a semantic calibration loss based on a neighborhood supervision signal; obtaining balanced prediction representation by using dynamic fusion weight, constructing a sample similarity matrix based on the balanced prediction representation, and constructing a debiased contrastive loss relying on same-class sample labels or different-class sample labels; generating sample pseudo labels based on debiased category representation, and calculating category alignment loss only for sample pseudo labels with a confidence higher than a threshold; obtaining a debiased total loss by weighted fusion, combining the debiased total loss and the original loss of the classification branch itself as a whole optimization target, and iteratively training all network units. The application also provides a debiased learning system, which ensures the performance balance of new-old category competition.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a bias-reduction learning method and system for competition between new and old categories in generalized category discovery. Background Technology

[0002] Generalized category discovery (GCD) is a core task in open-world visual understanding. Its goal is to: given a labeled dataset containing annotations of old classes (OldClasses)... (The tag belongs to the old category collection) ) and an unlabeled dataset that contains both old and new classes (NewClasses). ( ), using the supervision signal of the old class to simultaneously apply the supervision signal to all categories in the unlabeled data ( , including The identification and clustering are performed, and the new set of classes is defined as follows: .

[0003] Existing methods are mainly divided into two categories: clustering-based methods (such as semi-supervised K-means) and representation learning-based methods (such as SimGCD). The former tends to overfit to old classes, while the latter overfits to new classes due to excessive reliance on features of new classes in unlabeled data. Both clustering-based and representation learning-based methods face task competition between old and new classes.

[0004] Specifically, during training, the accuracy of the old category initially increased and then decreased, with the peak time closely coinciding with the time when the accuracy of the new category began to significantly improve. Confidence distribution analysis revealed that the confidence distribution of the new category (lower negative entropy, i.e., higher confidence) was systematically higher than that of the old category, revealing the root cause: the parameterized GCD classifier exhibited systematic overconfidence in the new category, leading to prediction bias in the new category encroaching on the feature distribution space of the old category.

[0005] It is evident that existing entropy calibration methods all assume that the label spaces of known / unknown classes are consistent, which is not applicable to the setting of coexistence and competition between new and old classes in GCD. Existing GCD methods (SimGCD, CiPR, etc.) have not identified and addressed the above-mentioned competition problem, but only optimized in one direction (new class or old class).

[0006] Therefore, there is an urgent need for bias-free learning methods and systems to address the competition between new and old categories in generalized category discovery, in order to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a bias-free learning method and system for the competition between new and old categories in generalized category discovery, so as to solve the aforementioned technical problems.

[0008] To achieve the above objectives, the present invention provides the following technical solution: Bias-free learning methods for addressing competition between new and old categories in generalized category discovery include: S1. Use a shared feature extraction unit to extract standardized features from all samples, set up a basic generalized category to discover classification branches, and output the basic category prediction results. Set up an auxiliary debiasing branch independently to map the standardized features to a dedicated representation space and output the debiased category prediction results. S2. Construct and maintain two types of synchronously dynamically updated repositories. The first repository stores the standardized features of all samples, and the second repository stores the basic category prediction results corresponding to all samples. Based on the auxiliary debiased branch output representation, retrieve the nearest neighbor samples in the first repository. The second repository extracts the prediction information of the nearest neighbor samples to generate a neighborhood supervision signal. Construct a semantic calibration loss based on the neighborhood supervision signal. S3. Dynamic fusion weights are generated by random distribution sampling. The basic category prediction results and the biased category prediction results are weighted and fused using the dynamic fusion weights to obtain a balanced prediction representation. A sample similarity matrix is ​​constructed based on the balanced prediction representation. A biased contrast loss is constructed based on the labels of similar or dissimilar samples. S4. Generate sample pseudo-labels based on the biased class prediction results, set a confidence screening threshold, and calculate the class alignment loss only for sample pseudo-labels with confidence scores higher than the threshold to constrain the stability of the old class feature distribution. S5. The semantic calibration loss, debiased contrast loss, and category alignment loss are weighted and fused to obtain the total debiased loss. The total debiased loss is then combined with the original loss of the basic generalized category discovery classification branch as the overall optimization objective. All network units are trained iteratively to simultaneously and evenly improve the accuracy of old category recognition and new category mining.

[0009] Furthermore, in S2, both the first and second repositories are updated synchronously using a batch-level momentum sliding mechanism, and the update expression is: ; ; In the formula, The first repository for storing standardized features; The second repository is predicted to be the storage base; , For momentum hyperparameters, , ; Standardized features for the current batch Predict the base category for the current batch.

[0010] Furthermore, in S2, retrieving nearest neighbor samples in the first repository based on the auxiliary debiased branch output representation includes: S21. Take the exclusive representation output by the auxiliary bias-removing branch of the current batch, and perform a temperature-scaled inner product operation with all stored features in the first repository to generate a distance matrix of the sample dimension of the batch repository. S22. For each distance vector corresponding to the current batch of samples in the distance matrix, sort them by distance value from smallest to largest, and select the samples with the smallest distance. For each sample, generate the nearest neighbor index set corresponding to that sample; S23. Based on the nearest neighbor index set for each sample, retrieve the base class prediction probability corresponding to each nearest neighbor index from the second repository, and then... The arithmetic mean of the predicted probabilities of each nearest neighbor is taken to obtain the mean of the semantic supervision of the neighborhood corresponding to the sample. S24. Based on the basic category prediction probability of the current batch of samples and the mean of neighborhood semantic supervision, calculate the KL divergence between the basic category prediction probability and the mean of neighborhood semantic supervision as the semantic calibration loss.

[0011] Furthermore, the formula for calculating the distance matrix is ​​as follows: ; ; in, This is a temperature scaling factor used to scale the similarity representation and smooth distribution differences. Batch size; For exclusive representation; The index filtering formula for the nearest neighbor index set is: ; in, The total number of all samples stored in the first repository; The formula for calculating the mean of the neighborhood semantic supervision is as follows: ; in, For the sample The The storage index corresponding to each nearest neighbor; The base predicted probability of this nearest neighbor index is stored in the second repository; The formula for calculating the semantic calibration loss is as follows: ; in, The total number of samples in the unlabeled dataset. This represents the total number of samples in the labeled old category dataset.

[0012] Furthermore, S3 also includes: The dynamic fusion weights follow a Gamma distribution. Distribution hyperparameters ; The formula for the balanced prediction representation of the fusion of the basic category prediction result and the biased category prediction result is as follows: ; in, Based on category prediction, To correct biased class prediction, This is the mixing coefficient.

[0013] Furthermore, S3 also includes: the sample similarity matrix is ​​calculated using a Softmax normalized inner product, and the formula for calculating the sample similarity matrix is: ; in, Temperature parameters are calculated for similarity. The total number of samples in the unlabeled dataset. This represents the total number of samples in the labeled old category dataset; True similarity based on one-hot encoding (Positive sample pairs are 1, negative sample pairs are 0) is the supervision signal. The bias-reduced contrastive loss is calculated using the following formula: .

[0014] Furthermore, in S4, the confidence screening threshold is set to 0.95. Samples below the threshold are filtered out using an indicator function, and the class alignment loss in the form of cross-entropy is calculated only for high-confidence samples. The formula for calculating the class alignment loss is as follows: ; in, Cross-entropy constraints are incorporated into the prediction.

[0015] Furthermore, the formula for calculating the total debiasing loss in S5 is as follows: ; in, For semantic calibration loss, Bias-free contrast loss, For category alignment loss, , For weight hyperparameters.

[0016] Furthermore, the basic generalized category discovery classification branch is the SimGCD or CiPR mainstream GCD model, and the auxiliary bias correction branch is an external plug-in module that can be directly compatible with any basic generalized category discovery classification branch without modifying the basic branch network structure and initial parameters.

[0017] This application also provides a debiasing learning system for competition between new and old categories in generalized category discovery, which is used to execute a debiasing learning system for competition between new and old categories in generalized category discovery, including a feature extraction module, a basic GCD classification module, an auxiliary debiasing branch module, a dual dynamic storage module, a semantic calibration calculation module, a debiasing contrast calculation module, a distribution alignment constraint module, and a joint optimization module; The feature extraction module is used to input image samples and output standardized features; The basic GCD classification module receives standardized features and outputs basic category prediction results. The auxiliary debiasing branch module includes an MLP projection unit and an auxiliary classification unit. The MLP projection unit receives standardized features and outputs a specific representation. The auxiliary classification unit outputs a debiased category prediction result based on the specific representation. The dual dynamic repository module includes a first feature repository and a second prediction repository. The first feature repository is used to store standardized features of all samples, and the second prediction repository is used to store the basic class prediction results and update the data in the repository through batch momentum sliding. The semantic calibration calculation module, in conjunction with the dual dynamic repository module, is used for nearest neighbor retrieval, generating neighborhood supervision signals and outputting semantic calibration loss. The debiased comparison calculation module is used for Gamma distribution sampling, dynamic fusion of weights, fusion of bi-branch predictions, construction of sample similarity matrix, and output of debiased comparison loss. The distribution alignment constraint module is used to filter high-confidence pseudo-label samples and calculate the class alignment loss in the form of cross-entropy. The joint optimization module is used to weight and fuse the three types of losses to obtain the debiased total loss, and combine it with the original loss of the basic GCD classification module to complete the network iterative training.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This application extracts standardized features through a shared feature extraction unit and sets up a basic branch and an auxiliary debiasing branch. This maps the original features to different representation spaces, providing a multi-dimensional prediction perspective for subsequent debiasing processing and avoiding overfitting of a single branch to the new category. By constructing and maintaining two synchronously dynamically updated repositories and retrieving nearest neighbor samples based on the representation of the auxiliary debiasing branch, it can generate supervision signals using semantic information within the neighborhood. Semantic calibration loss constrains model predictions, enabling the model to refer to the semantic consistency of the neighborhood when processing samples, mitigating prediction bias caused by overconfidence in the new category. By using random distribution sampling to generate dynamic fusion weights, the basic prediction and debiasing prediction are weighted and fused, dynamically balancing the prediction contributions of different branches. Based on this, a sample similarity matrix is ​​constructed, and debiasing contrastive loss strengthens the clustering of similar samples and the separation of dissimilar samples, eliminating competition and interference between new and old categories at the feature space level. By generating pseudo-labels based on biased prediction and setting a confidence threshold, the class alignment loss is calculated only for high-confidence samples, which can specifically constrain the stability of the old class feature distribution. By merging the semantic calibration loss, biased contrast loss, and class alignment loss with the original loss of the basic branch, an overall optimization objective is constructed, which can achieve collaborative training of network units. This ensures that while improving the accuracy of new class mining, the model can effectively maintain the recognition performance of old classes, thereby ensuring the performance balance of competition between new and old classes. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation of the invention. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating the implementation of the bias correction learning method provided in this application embodiment; Figure 2 A flowchart illustrating the implementation of retrieving nearest neighbor samples provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the implementation process of the bias removal learning method provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the bias correction learning system provided in the embodiments of this application; Figure 5 A graph showing the competition curves of the accuracy of new and old categories during the SimGCD training process of the bias-reduction learning method provided in the embodiments of this application. Figure 6 A comparison of the confidence distribution of new and old classes during the SimGCD training process using the bias-reduction learning method provided in this application embodiment; Figure 7 Weight hyperparameters provided for embodiments of this application and Sensitivity analysis plot; Figure 8 A visualization comparison diagram in the t-SNE feature space provided for embodiments of this application; Figure 9 The confidence distribution change diagram after the application of the bias correction learning method provided in the embodiments of this application; Figure 10 The accuracy training curve is shown after applying the bias-reduction learning method provided in the embodiments of this application. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] See Figures 1 to 10 As shown, the debiasing learning method for the competition between new and old categories in generalized category discovery includes: S1, using a shared feature extraction unit to extract standardized features from all samples, setting a basic generalized category discovery classification branch and outputting the basic category prediction results, and independently setting an auxiliary debiasing branch to map the standardized features to a dedicated representation space and output the debiased category prediction results; S2, constructing and maintaining two synchronously dynamically updated repositories, the first repository storing the standardized features of all samples, and the second repository storing the basic category prediction results corresponding to all samples, and retrieving nearest neighbor samples in the first repository based on the output representation of the auxiliary debiasing branch, the second repository extracting the prediction information of nearest neighbor samples to generate neighborhood supervision signals, and constructing a semantic calibration loss based on the neighborhood supervision signals; S3, using a random distribution sampling method. A dynamic fusion weight is generated, and the basic category prediction results and the biased category prediction results are weighted and fused using the dynamic fusion weight to obtain a balanced prediction representation. A sample similarity matrix is ​​constructed based on the balanced prediction representation, and a biased contrast loss is constructed based on the labels of similar or dissimilar samples. S4. Sample pseudo-labels are generated based on the biased category prediction results, and a confidence screening threshold is set. The category alignment loss is calculated only for sample pseudo-labels with confidence scores higher than the threshold to constrain the stability of the old category feature distribution. S5. The semantic calibration loss, biased contrast loss, and category alignment loss are weighted and fused to obtain the total biased loss. The total biased loss is merged with the original loss of the basic generalized category discovery classification branch as the overall optimization objective. All network units are iteratively trained to simultaneously and evenly improve the accuracy of old category recognition and new category mining.

[0022] In the above scheme, this application extracts standardized features through a shared feature extraction unit and sets up a basic branch and an auxiliary debiasing branch. This maps the original features to different representation spaces, providing a multi-dimensional prediction perspective for subsequent debiasing processing and avoiding overfitting of a single branch to the new category. By constructing and maintaining two synchronously dynamically updated repositories and retrieving nearest neighbor samples based on the representation of the auxiliary debiasing branch, it can generate supervision signals using semantic information within the neighborhood. The model prediction is constrained by semantic calibration loss, enabling the model to refer to the semantic consistency of the neighborhood when processing samples, thus mitigating prediction bias caused by overconfidence in the new category. By using random distribution sampling to generate dynamic fusion weights, the basic prediction and debiasing prediction are weighted and fused, which dynamically balances the prediction contributions of different branches. Based on this, a sample similarity matrix is ​​constructed, and the debiasing contrastive loss strengthens the clustering of similar samples and the separation of dissimilar samples, eliminating competition and interference between new and old categories at the feature space level. By generating pseudo-labels based on biased prediction and setting a confidence threshold, the class alignment loss is calculated only for high-confidence samples, which can specifically constrain the stability of the old class feature distribution. By merging the semantic calibration loss, biased contrast loss, and class alignment loss with the original loss of the basic branch, an overall optimization objective is constructed. This enables the collaborative training of network units to ensure that the model improves the accuracy of new class mining while effectively maintaining the recognition performance of old classes, thereby ensuring the performance balance of competition between new and old classes.

[0023] In S2, both the first and second repositories are updated synchronously using a batch-level momentum sliding mechanism. The update expression is: ； ; In the formula, The first repository for storing standardized features; A second repository for storing the basic category prediction results corresponding to the full set of samples; 、 For momentum hyperparameters, , ; Standardized features for the current batch Predict the base category for the current batch.

[0024] This application covers the full data. Two momentum update memories are constructed and maintained separately. The prediction probability of the base class prediction results in the second memory is calculated using the following formula: ; in, Indicates standardized features, preferred. .

[0025] The first repository stores standardized features extracted from all samples. Its purpose is to provide a stable and comprehensive feature space for use by the bias-free branch when retrieving nearest neighbor samples. It can be a large matrix or tensor, where the number of rows corresponds to the number of stored samples and the number of columns corresponds to the feature dimension; alternatively, it can be a hash table or dictionary where the keys are sample IDs and the values ​​are the corresponding standardized feature vectors. The first repository uses momentum parameters... Applying a moving average to the standardized features means that with each update, the new standardized features... It will not completely replace the old storage features Instead, The weights retain historical information, so as to The weights are incorporated into the current batch information, ensuring the stability of the feature distribution within the retrieval space and effectively avoiding nearest neighbor retrieval bias caused by noise or instantaneous changes in a single batch of data, making the retrieved nearest neighbor samples more representative.

[0026] The second repository is specifically used to store the base class prediction results for all samples. These base class prediction results are typically probability distributions after Softmax activation, reflecting the confidence of the base generalized class discovery branch for each sample's class. Its purpose is to provide basic prediction information for generating neighborhood supervision signals. It can be a matrix or tensor, with the number of rows corresponding to the number of stored samples and the number of columns corresponding to the number of classes; alternatively, it can be a list or array, where each element is a predicted probability vector for a sample. The second repository uses momentum parameters... The basic category prediction results are updated using a sliding method to ensure that the basic category predictions for the current batch are accurate. by The weights are incorporated into historical prediction information, enabling neighborhood supervision signals to be generated based on more representative historical prediction averages, rather than solely relying on the prediction results of the current batch.

[0027] The step S2, which involves retrieving nearest neighbor samples in the first repository based on the auxiliary bias-removal branch output representation, includes: S21, taking the exclusive representation of the current batch after auxiliary bias-removal branch output, and performing a temperature-scaled inner product operation with all stored features in the first repository to generate a distance matrix in the sample dimension of the batch repository; S22, for each distance vector corresponding to the current batch sample in the distance matrix, sorting them by distance value from smallest to largest, and selecting the nearest neighbor sample with the smallest distance. For each sample, generate a set of nearest neighbor indices corresponding to that sample; S23, based on the set of nearest neighbor indices for each sample, retrieve the basic category prediction probabilities corresponding to each nearest neighbor index from the second repository, and perform... S24. Calculate the mean of the neighborhood semantic supervision for the sample by taking the arithmetic mean of the predicted probabilities of the basic category and the mean of the neighborhood semantic supervision based on the basic category predicted probability and the mean of the neighborhood semantic supervision for the current batch of samples. KL Divergence is used as the semantic calibration loss.

[0028] In the above scheme, the specific representation refers to the low-dimensional vector representation with specific semantic information generated for the current batch of samples after auxiliary debiasing branch processing. Its role is to capture the core features of the samples in the debiasing learning task and map them to a specific space that is conducive to distinguishing between new and old categories. This specific representation is obtained by performing a nonlinear transformation on the standardized features through a multilayer perceptron (MLP), or it can be an enhanced feature vector obtained after processing with a self-attention mechanism.

[0029] A distance vector is a row (or column) in a distance matrix corresponding to a sample in the current batch. It contains distance information between that sample and all samples in the first repository. Sort by distance value in ascending order is a common sorting algorithm used to find the sample most similar to the current sample (with the smallest distance).

[0030] The base class prediction probabilities corresponding to each nearest neighbor index are retrieved from the second repository, which stores the base class prediction results for all samples. Using the nearest neighbor index set, the base class prediction probability distribution for the corresponding nearest neighbor samples can be accurately retrieved from the second repository. This base class prediction probability distribution reflects the class judgment of the base classifier for these nearest neighbor samples. Taking the arithmetic mean of the predicted probabilities of the nearest neighbors is a method for aggregating neighborhood information, by... Averaging the predicted probabilities of the nearest neighbors yields a smooth category distribution that represents neighborhood consensus, reducing the impact of prediction errors from individual nearest neighbor samples and providing a more robust semantic supervision signal.

[0031] The base category prediction probability of the current batch of samples refers to the original category prediction result output by the base generalized category discovery classification branch for the current batch of samples, which is usually a probability distribution vector. KL Kullback-Leibler Divergence is an asymmetric measure of the difference between two probability distributions.

[0032] The semantic calibration loss is calculated using KL divergence, and its goal is to minimize the difference between the current sample's own prediction and the mean of the neighborhood semantic supervision. By optimizing the semantic calibration loss, the model can be forced to adjust its base class prediction to align with the consensus of semantically similar neighborhood samples, thereby achieving semantic calibration of the base classifier's prediction results and mitigating the prediction bias caused by competition between new and old classes.

[0033] The formula for calculating the distance matrix is: ; ; It is used to quantify the similarity between the specific representation of the current batch of samples and all the standardized features stored in the first repository.

[0034] in, This is a temperature scaling factor used to scale the similarity representation and smooth distribution differences. It can be a preset fixed value, such as 0.07 or 0.1, or it can be a hyperparameter that is dynamically adjusted by the optimization algorithm during training. Batch size; The specific representation is obtained after the auxiliary debiasing branch extracts features from the current batch of samples, and represents the semantic information of the sample in a specific representation space.

[0035] The index filtering formula for the nearest neighbor index set is: ; in, The total number of samples stored in the first repository is used as the index for filtering. This indexing method efficiently identifies the local neighborhoods semantically closest to the current sample from a large-scale sample repository, providing accurate contextual information for subsequent semantic supervision signal generation. For example, It can be an empirically set integer, such as 5, 10, or 15, or it can be selected based on the characteristics of the dataset or the performance of the model for cross-validation.

[0036] The formula for calculating the mean of the neighborhood semantic supervision is as follows: ; It is used to calculate samples Neighborhood semantic supervision mean .in, For the sample The The storage index corresponding to each nearest neighbor; The base predicted probability for this nearest neighbor index is stored in the second repository.

[0037] In obtaining samples of Nearest neighbor index set Then, the base class prediction probabilities corresponding to these nearest neighbor samples are retrieved from the second repository. Through this Nearest basic category prediction probability By performing an arithmetic average, a supervisory signal that aggregates local semantic information can be obtained. This not only suppresses noise or bias that may exist in the prediction of individual nearest neighbor samples, but also provides a more stable and representative neighborhood semantic consensus. In addition to the arithmetic average, a weighted average can also be used, for example, assigning different weights based on the distance between neighbors.

[0038] The formula for calculating the semantic calibration loss is as follows: ; The semantic calibration loss is used to measure the base class prediction probability of the current batch of samples. Mean of semantic supervision in its neighborhood ,in, The total number of samples in the unlabeled dataset. This represents the total number of samples in the labeled old category dataset. Divergence is an asymmetric measure of the difference between two probability distributions. This is achieved by minimizing the semantic calibration loss. The model is guided to adjust the samples. Basic category prediction This enables basic category prediction. Mean of semantic supervision in its local neighborhood Convergence. Local semantic consistency is used to calibrate the model's predictions, thereby constraining the distribution of new and old classes in the feature space and mitigating the encroachment of overconfidence in the new class on the feature space of the old class. Besides... KL Divergence can also be measured using Jensen-Shannon divergence or cross-entropy loss to measure distributional differences.

[0039] This implementation method verifies the effectiveness under different datasets, and focuses on retaining the main effect table and the necessary implementation parameter table, as shown in Tables 1-2 below. Table 1 compares experimental results on the CUB-based fine-grained dataset. Table 1 shows that DeGCD+SimGCD provides a 2.9% improvement across all classes and a 2.8% improvement over the old classes in the CUB dataset compared to the suboptimal method; DeGCD+SimGCD provides an 11% improvement across all classes and a 3.3% improvement over the old classes in StanfordCars compared to the suboptimal method; and DeGCD+SimGCD provides a 2.9% improvement across all classes in Herbarium19 compared to the suboptimal method.

[0040] Table 2 compares experimental results on a CUB-based general dataset. Table 2 shows that on CIFAR-100, DeGCD+SimGCD improves performance by 3.9% across all classes and 4.0% for new classes compared to the second-best baseline CiPR; on ImageNet-100, it improves performance by 2.2% across all classes and 3.3% for new classes.

[0041] S3 further includes: The dynamic fusion weights follow a Gamma distribution. Distribution hyperparameters ; The formula for the balanced prediction representation of the fusion of the basic category prediction result and the biased category prediction result is as follows: ; in, Basic category prediction typically involves discovering the classification branch output from the basic generalized category, and tends to have a higher recognition ability for older categories. For debiased category prediction, it is output by an auxiliary debiased branch, which aims to alleviate overconfidence in new categories. The mixing coefficient. (The mixing coefficient is used to...) Linear weighting is applied to the base class prediction and the biased class prediction to generate an balanced prediction representation. For example, when When the value is high, the base category prediction Dominant in integration; when When the value is low, biased class prediction The influence of this is enhanced to ensure the final equilibrium prediction characterization. It can comprehensively consider the advantages of classification branches and de-biasing branches, thereby effectively improving the ability to discover new categories while maintaining the stability of old category recognition.

[0042] S3 further includes: the sample similarity matrix is ​​calculated using a Softmax normalized inner product, and the formula for calculating the sample similarity matrix is: ; in, Temperature parameters are calculated for similarity. The total number of samples in the unlabeled dataset. This represents the total number of samples in the labeled old category dataset. True similarity based on one-hot encoding (Positive sample pairs are 1, negative sample pairs are 0) is the supervision signal. The bias-reduced contrastive loss is calculated using the following formula: .

[0043] This application's sample similarity matrix is ​​used to quantify the similarity between any two samples in a dataset within the feature space. By using Softmax to normalize the inner product calculation, the original inner product similarity can be converted into a probability distribution, ensuring that the similarity values ​​are between 0 and 1, and that the sum of the similarities for all sample pairs is 1. The sample similarity matrix can be implemented by directly calculating the inner product of all sample pairs using matrix multiplication, and then applying the Softmax function to normalize the result; or by calculating the inner product of each pair of samples, and then inputting these inner product values ​​into the Softmax function to obtain the normalized similarity. Temperature parameter Used to adjust the sharpness of the output distribution of the Softmax function.

[0044] When temperature parameter When the temperature parameter is small, the probability distribution of the Softmax output will be more concentrated on the samples with the highest similarity, making the model more sensitive to differences in similarity; when the temperature parameter is small... When the similarity is large, the probability distribution output by Softmax becomes smoother, reducing its sensitivity to differences in similarity. Its purpose is to smooth the similarity distribution, avoiding gradient vanishing or overfitting problems caused by overly sharp predictions, thus allowing the model to focus more on the relative relationships between samples when constructing the similarity matrix.

[0045] True similarity of one-hot encoding This is a method of converting category labels into binary vectors. In this application, it represents the true similarity between sample pairs; if two samples belong to the same category (positive sample pair), then... If the sample is in a different category (negative sample pair), then the result is 1. It is 0.

[0046] Bias-free contrast loss This is a loss function used for contrastive learning, aiming to teach the model a feature representation where features of similar samples are close to each other, while features of dissimilar samples are far apart. It achieves this by encoding true similarity in a one-hot manner. As a supervisory signal, and combined with the form of cross-entropy, it can explicitly introduce known old category label information into the contrastive learning framework, thereby correcting the model's bias in handling new and old categories and suppressing the encroachment of new categories on the feature space of old categories.

[0047] In S4, the confidence threshold is set to 0.95. Samples below the threshold are filtered out using an indicator function. The class alignment loss in the form of cross-entropy is calculated only for high-confidence samples. The formula for calculating the class alignment loss is as follows: ; in, Cross-entropy constraints are incorporated into the prediction.

[0048] In this application, the confidence screening threshold is a preset value used to judge the reliability of the model's prediction results. Setting this threshold to 0.95 aims to ensure that only samples for which the model has extremely high certainty about its predictions are further processed, thereby improving the quality of pseudo-labels. This threshold can be used to determine whether a sample is screened by comparing the class probability value output by the model with this threshold, or by calculating the prediction entropy value and comparing it with the entropy threshold based on a 0.95 confidence level. In the training code, the confidence level of all samples can be calculated first. Then for each sample Perform threshold judgment and remove samples that do not meet the conditions. The terms are set to 0, and finally all terms are summed or averaged; or the indicator function logic is directly integrated into the loss function module, so that the loss calculation process automatically ignores low confidence samples, ensuring that the model always uses high-quality samples as a reference during the optimization process, thereby more accurately aligning and anchoring the old categories in the feature space.

[0049] This application applies a consistency constraint to the high-confidence output of the biased classifier, ensuring that the old class maintains a stable distribution within its dominant high-confidence region, effectively preventing the expansion and erosion of the new class. In this embodiment... The implementation effect was verified in the basic scenarios of CUB and CIFAR-100, and the main effect table and the necessary implementation parameter table were retained as the focus, as shown in Table 3 below. Table 3 Comparison of effects in basic scenarios of CUB and CIFAR-100. Table 3 shows that, when the number of categories is unknown, DeGCD improves by 3.9% across all categories of CUB compared to the baseline SimGCD, by 3.1% across all categories of CIFAR-100 compared to the baseline SimGCD, by 5.7% across new categories of CIFAR-100 compared to the baseline SimGCD, and by 0.3% across old categories of CIFAR-100 compared to the baseline SimGCD.

[0050] The formula for calculating the total debiasing loss in S5 is: ; in, For semantic calibration loss, To eliminate biased contrast loss, For category alignment loss, , For weight hyperparameters.

[0051] Total loss after partial correction This is used to guide model training, enabling it to simultaneously improve the accuracy of old category recognition and new category discovery in generalized category discovery tasks. By integrating semantic calibration loss, bias-reduction contrastive loss, and category alignment loss functions, it ensures that the model is effectively constrained across different dimensions.

[0052] Among them, semantic calibration loss It is responsible for constraining the semantic consistency of the feature space through neighborhood supervision signals, ensuring that the model's semantic understanding of samples is consistent with neighborhood information; bias-free contrastive loss. Responsible for enhancing inter-class discrimination through the sample similarity matrix, making the features of samples of the same class more similar and the features of samples of different classes more dissimilar; class alignment loss. By using high-confidence pseudo-labels to constrain the stability of the old class feature distribution, the bias in new class predictions can be prevented from eroding the old class. This is achieved by setting weight hyperparameters. and This allows for the adjustment of the importance of different loss functions in the overall optimization process, enabling the semantic calibration loss, bias-reduction contrast loss, and category alignment loss functions to meet the actual needs during the training phase. It also dynamically balances the influence of each constraint on the model parameter update, ensuring that the model can effectively mine new category features while preventing the degradation of old category recognition accuracy during iterative training.

[0053] The basic generalized category discovery classification branch is the SimGCD or CiPR mainstream GCD model, and the auxiliary bias correction branch is an external plug-in module that can be directly compatible with any basic generalized category discovery classification branch without modifying the basic branch network structure and initial parameters.

[0054] In the above scheme, the basic generalized category discovery classification branch, such as the SimGCD or CiPR model, is a widely used basic model in the current field of generalized category discovery. Its core function is to extract features from input samples and output corresponding category prediction results. The auxiliary debiasing branch is configured as an external plug-in module, which independently receives and processes standardized features from the shared feature extraction unit and outputs its own debiased category prediction results. This allows the auxiliary debiasing branch to be integrated into the existing generalized category discovery framework in a non-intrusive manner, thereby avoiding performance degradation or additional training costs that may be caused by structural adjustments or parameter resets.

[0055] In practice, the basic generalized category discovery classification branch can be a SimGCD model built on a ResNet-50 backbone network, where the shared feature extraction unit is the convolutional layer of ResNet-50, and the classification head is a fully connected layer. The auxiliary bias-reduction branch can be an independent network consisting of two or three fully connected layers (MLP), with each layer potentially containing a ReLU activation function and a BatchNormalization layer. The input to this auxiliary bias-reduction branch is directly taken from the normalized feature vector output by the ResNet-50 backbone network, and its output is a probability distribution vector corresponding to the number of categories, serving as the bias-reduction category prediction result. In actual deployment, this auxiliary bias-reduction branch module can be loaded as an independent Python class or function library and run in parallel with the basic SimGCD model during the training loop. For example, in the PyTorch framework, the SimGCD model can be instantiated, and then the auxiliary bias-reduction branch module can be instantiated, with both sharing the output of the same feature extractor. When configuring the optimizer, the parameters of the SimGCD model and the auxiliary bias-reduction branch module can be optimized simultaneously without modifying or re-initializing the initial weights of the SimGCD model's ResNet-50 structure or its classification head. This overcomes the limitations of traditional methods that require large-scale structural reconstruction or parameter resetting of existing generalized class discovery models when introducing bias-reduction mechanisms.

[0056] Example 1: Scene discovery using CUB-200-2011 fine-grained bird cosmological classification The fine-grained visual recognition standard dataset CUB-200-2011 was used. The dataset contains 200 different bird species, and each image is accompanied by local key points of the bird and species subdivision attributes. Labeled datasets The top 100 bird species were selected as the known old categories. For each bird category, 50% of all images were randomly selected as labeled samples with precise bird category labels. These samples were used only to provide supervision signals for the old categories. Unlabeled datasets All remaining images were grouped into the unlabeled set, which contained the remaining images of the aforementioned 100 old categories, as well as 100 new bird species that had never been seen before (a new category set). The total number of samples in the unlabeled set was 4,500.

[0057] The visual feature extraction backbone uses the ViT-B / 16 visual Transformer model based on DINO unsupervised pre-training: Image input to ViT-B / 16 backbone network, extraction of 768-dimensional [CLS] global feature vector from encoder output, and processing of features. Normalization yields a standardized characterization ; Standardized features are input into the existing mainstream GCD main classifier SimGCD. The main classifier outputs the original class prediction logits, which are then normalized by Softmax to obtain the base class prediction. The original SimGCD network structure and loss function are completely preserved without any modifications. Standardized features are fed in parallel into an external de-biasing head, which consists of two MLP projection layers and an auxiliary classifier connected in series. The MLP maps the 768-dimensional original features to an independent low-dimensional projection space, outputting a dedicated projection representation. The auxiliary classifier outputs debiased prediction logits based on the projected representation, which are then normalized to obtain the debiased prediction probability. ; The two branches run in complete parallel, the feature extractors share weights, and the debiasing module does not share parameters with the main classifier, existing as an independent additional module.

[0058] Build two independent repositories, and perform momentum sliding updates after each training batch is completed: ； ； Momentum hyperparameter unified setting = With a weight of 0.99, historical global sample information is retained with extremely high weight, while fresh features and prediction results of the current batch are absorbed with only a small weight, smoothing the representation fluctuations caused by noise in a single batch and ensuring the global semantic stability of the memory retrieval.

[0059] Take the current batch of de-biased outputs and outputs a unique representation. , with the first repository M Perform temperature-scaled inner product operations on all stored features to generate a distance matrix. ; Sort the distance vectors of each sample in ascending order, and select the samples with the smallest distances. K Each sample is used as a Top-K nearest neighbor, and a nearest neighbor index set is extracted. From the second repository based on the nearest neighbor index Read the corresponding basic prediction probability from the middle, and then... K The arithmetic mean of the probabilities of each nearest neighbor is used to generate the mean of semantic supervision in the neighborhood. ; Calculate the prediction probability of the current sample by the main classifier. Mean of semantic supervision in the neighborhood The semantic calibration loss is obtained by averaging the KL divergence over all samples. Semantic calibration loss Local sample predictions tend to be consistent, suppressing the continuous propagation of isolated high-confidence false labels of new categories and mitigating the accumulation of bias.

[0060] Setting distribution hyperparameters , from Real-time random sampling mixing coefficients in the distribution ; Dynamic weights are used to fuse predictions from the main classifier and the bias-reduced classifier. ; This neutralizes the dual shortcomings of the main classifier being overconfident in new classes and underfitting the partial branch; Based on the fusion equilibrium prediction The self-similarity matrix of the entire sample is calculated using the Softmax inner product with temperature scaling, and the temperature parameter is used to calculate the similarity. =0.1; Using real-class or dissimilar pairings of samples as supervision signals, the bias-reduced contrastive loss is calculated. It brings similar samples closer together and pushes away dissimilar samples, thus correcting the prediction bias of the similarity matrix.

[0061] Temperature scaling is applied to the predicted probabilities output by the biased classifier, and the pseudo-label is generated by taking the maximum probability value of each sample. Set the confidence threshold to 0.95 and filter samples using an indicator function: only high-reliability samples with a maximum confidence of ≥0.95 are retained for loss calculation, while low-confidence ambiguous samples are discarded directly; The cross-entropy loss is calculated for the selected high-confidence samples to obtain the alignment constraint loss. By stabilizing the category distribution boundary of high-confidence regions, it prevents new category features from continuously encroaching on the old category feature space, thus alleviating the ebb and flow of category competition during training.

[0062] Total bias loss weighted combination formula: ； Weight hyperparameters =0.35, weight hyperparameter =0.2.

[0063] Overall model optimization objective = SimGCD native training loss + weighted total debiasing loss The AdamW optimizer is used to backpropagate and update all trainable parameters (ViT last layer Transformer, SimGCD main classifier, debiased MLP and auxiliary classifier), and the training is carried out in a complete iterative training for 200 epochs. During the training process, the recognition accuracy of old categories and the clustering accuracy of new categories are optimized simultaneously.

[0064] In fine-grained bird recognition scenarios, the accuracy of this implementation surpasses that of mainstream comparison algorithms such as NGCN and SimGCD for both all categories and existing categories, validating the unique advantages of the proposed solution in scenarios where class distinctions are difficult to make. Specifically, the overall accuracy (ACC) of this implementation reaches 64.2%, for existing categories 68.4%, and for new categories 62.1%. This achieves state-of-the-art (SOTA) accuracy for both all categories and existing categories, with the accuracy for existing categories improving by 2.8% compared to the SimGCD baseline (65.6%), and the accuracy for new categories improving by 4.4% compared to the SimGCD baseline (62.1%).

[0065] This application also provides a bias-reduction learning system for competition between new and old classes in generalized class discovery. The system includes a feature extraction module, a basic GCD classification module, an auxiliary bias-reduction branch module, a dual dynamic storage module, a semantic calibration calculation module, a bias-reduction contrast calculation module, a distribution alignment constraint module, and a joint optimization module. The feature extraction module takes image samples as input and outputs standardized features. The basic GCD classification module receives standardized features and outputs basic class prediction results. The auxiliary bias-reduction branch module includes an MLP projection unit and an auxiliary classification unit. The MLP projection unit receives standardized features and outputs a specific representation, while the auxiliary classification unit outputs a bias-reduction class prediction result based on the specific representation. The dual dynamic storage module includes a... The system comprises a feature repository and a prediction repository. The first feature repository stores standardized features of all samples, and the second prediction repository stores basic class prediction results. Data within the repository is updated via batch momentum sliding. A semantic calibration calculation module, linked to the dual dynamic repository module, is used for nearest neighbor retrieval, generating neighborhood supervision signals, and outputting semantic calibration loss. A debiasing contrast calculation module is used for dynamically fusing weights using Gamma distribution sampling, fusing bi-branch predictions, constructing a sample similarity matrix, and outputting debiasing contrast loss. A distribution alignment constraint module is used to filter high-confidence pseudo-label samples and calculate class alignment loss in the form of cross-entropy. A joint optimization module is used to weightedly fuse the three class losses to obtain the total debiasing loss, and combines this with the original loss from the basic GCD classification module to complete iterative network training.

[0066] In the above scheme, the feature extraction module serves as the foundation of the system. By standardizing the input samples, it provides a unified and stable feature input for subsequent classification and representation learning. The basic GCD classification module is responsible for outputting the basic category prediction results, which serve as the core reference for measuring the current model's ability to recognize old classes.

[0067] The auxiliary debiasing branch module introduces MLP projection units and auxiliary classification units to map standardized features to a dedicated representation space. It can capture debiased class prediction results independently of the basic branch, thus providing a differentiated prediction perspective for subsequent debiasing strategies and effectively avoiding the overconfidence of a single branch in new categories.

[0068] The dual dynamic repository module, through the setting of a first feature repository and a second prediction repository, combined with a batch momentum sliding update mechanism, enables real-time dynamic maintenance of the features and prediction information of the entire sample, providing reliable global context information for subsequent nearest neighbor retrieval and semantic calibration.

[0069] The semantic calibration calculation module uses information from the dual dynamic repository to perform nearest neighbor retrieval. By generating neighborhood supervision signals and calculating semantic calibration loss, it can effectively correct the prediction bias of the model on local samples and enhance the semantic consistency of the feature space.

[0070] The debiased contrast calculation module generates dynamic fusion weights through Gamma distribution sampling, weights and fuses the basic predictions and debiased predictions, and constructs a sample similarity matrix. Through the constraint of the debiased contrast loss, the model can maintain a more balanced discrimination ability when dealing with new and old categories, and avoid the encroachment of new category features on the old category space.

[0071] The distribution alignment constraint module filters high-confidence samples by setting a confidence threshold and calculates the class alignment loss, which can specifically constrain the stability of the old class feature distribution and prevent the degradation of the old class recognition performance during the process of mining new classes.

[0072] The joint optimization module weighted and fused the category alignment loss, bias correction contrast loss, semantic calibration loss, and the original loss of the basic GCD classification module. Through iterative training of the overall optimization objective, it achieved coordinated updates of network parameters, thereby improving the accuracy of new category mining while effectively ensuring the stability of the accuracy of old category recognition, and realizing the coordinated optimization of new and old categories.

[0073] 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 present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A bias-reduction learning method for competition between new and old categories in generalized category discovery, characterized in that: include: S1. Use a shared feature extraction unit to extract standardized features from all samples, set up a basic generalized category to discover classification branches, and output the basic category prediction results. Set up an auxiliary debiasing branch independently to map the standardized features to a dedicated representation space and output the debiased category prediction results. S2. Construct and maintain two types of synchronously dynamically updated repositories. The first repository stores the standardized features of all samples, and the second repository stores the basic category prediction results corresponding to all samples. Based on the auxiliary debiased branch output representation, retrieve the nearest neighbor samples in the first repository. The second repository extracts the prediction information of the nearest neighbor samples to generate a neighborhood supervision signal. Construct a semantic calibration loss based on the neighborhood supervision signal. S3. Dynamic fusion weights are generated by random distribution sampling. The basic category prediction results and the biased category prediction results are weighted and fused using the dynamic fusion weights to obtain a balanced prediction representation. A sample similarity matrix is ​​constructed based on the balanced prediction representation. A biased contrast loss is constructed based on the labels of similar or dissimilar samples. S4. Generate sample pseudo-labels based on the biased class prediction results, set a confidence screening threshold, and calculate the class alignment loss only for sample pseudo-labels with confidence scores higher than the threshold to constrain the stability of the old class feature distribution. S5. The semantic calibration loss, debiased contrast loss, and category alignment loss are weighted and fused to obtain the total debiased loss. The total debiased loss is then combined with the original loss of the basic generalized category discovery classification branch as the overall optimization objective. All network units are trained iteratively to simultaneously and evenly improve the accuracy of old category recognition and new category mining.

2. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 1, characterized in that, In S2, both the first and second repositories are updated synchronously using a batch-level momentum sliding mechanism, and the update expression is: ； ； In the formula, The first repository for storing standardized features; The second repository is predicted to be the storage base; , For momentum hyperparameters, , ; Standardized features for the current batch Predict the base category for the current batch.

3. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 1, characterized in that, The S2 step of retrieving nearest neighbor samples in the first repository based on the auxiliary bias-removal branch output representation includes: S21. Take the exclusive representation output by the auxiliary bias-removing branch of the current batch, and perform a temperature-scaled inner product operation with all stored features in the first repository to generate a distance matrix of the sample dimension of the batch repository. S22. For each distance vector corresponding to the current batch of samples in the distance matrix, sort them by distance value from smallest to largest, and select the samples with the smallest distance. For each sample, generate the nearest neighbor index set corresponding to that sample; S23. Based on the nearest neighbor index set for each sample, retrieve the base class prediction probability corresponding to each nearest neighbor index from the second repository, and then... The arithmetic mean of the predicted probabilities of each nearest neighbor is taken to obtain the mean of the semantic supervision of the neighborhood corresponding to the sample. S24. Based on the base class prediction probability and the mean of neighborhood semantic supervision of the current batch of samples, calculate the relationship between the base class prediction probability and the mean of neighborhood semantic supervision. KL Divergence is used as the semantic calibration loss.

4. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 3, characterized in that, The formula for calculating the distance matrix is: ， ； in, This is a temperature scaling factor used to scale the similarity representation and smooth distribution differences. Batch size; For exclusive representation; The index filtering formula for the nearest neighbor index set is: ； in, The total number of all samples stored in the first repository; The formula for calculating the mean of the neighborhood semantic supervision is as follows: ； in, For the sample The The storage index corresponding to each nearest neighbor; The base predicted probability of this nearest neighbor index is stored in the second repository; The formula for calculating the semantic calibration loss is as follows: ； in, The total number of samples in the unlabeled dataset. This represents the total number of samples in the labeled old category dataset.

5. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 1, characterized in that, S3 further includes: The dynamic fusion weights follow a Gamma distribution. Distribution hyperparameters ; The formula for the balanced prediction representation of the fusion of the basic category prediction result and the biased category prediction result is as follows: ； in, Based on category prediction, To correct biased class prediction, This is the mixing coefficient.

6. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 1, characterized in that, S3 further includes: the sample similarity matrix is ​​calculated using a Softmax normalized inner product, and the formula for calculating the sample similarity matrix is: ； in, Temperature parameters are calculated for similarity. The total number of samples in the unlabeled dataset. This represents the total number of samples in the labeled old category dataset; True similarity based on one-hot encoding (Positive sample pairs are 1, negative sample pairs are 0) is the supervision signal. The bias-reduced contrastive loss is calculated using the following formula: 。 7. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 1, characterized in that, In S4, the confidence threshold is set to 0.

95. Samples below the threshold are filtered out using an indicator function. The class alignment loss in the form of cross-entropy is calculated only for high-confidence samples. The formula for calculating the class alignment loss is as follows: ； in, Cross-entropy constraints are incorporated into the prediction.

8. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 1, characterized in that, The formula for calculating the total de-biasing loss in S5 is as follows: ； in, For semantic calibration loss, Bias-free contrast loss, For category alignment loss, , For weight hyperparameters.

9. The bias-reduction learning method for competition between new and old categories in generalized category discovery according to claim 1, characterized in that, The basic generalized category discovery classification branch is the SimGCD or CiPR mainstream GCD model, and the auxiliary bias correction branch is an external plug-in module that can be directly compatible with any basic generalized category discovery classification branch without modifying the basic branch network structure and initial parameters.

10. A bias-reduction learning system for competition between new and old categories in generalized category discovery, characterized in that: The method for performing the debiased learning method according to any one of claims 1-9 includes a feature extraction module, a basic GCD classification module, an auxiliary debiased branch module, a dual dynamic storage module, a semantic calibration calculation module, a debiased contrast calculation module, a distribution alignment constraint module, and a joint optimization module. The feature extraction module is used to input image samples and output standardized features; The basic GCD classification module receives standardized features and outputs basic category prediction results. The auxiliary debiasing branch module includes an MLP projection unit and an auxiliary classification unit. The MLP projection unit receives standardized features and outputs a specific representation. The auxiliary classification unit outputs a debiased category prediction result based on the specific representation. The dual dynamic repository module includes a first feature repository and a second prediction repository. The first feature repository is used to store standardized features of all samples, and the second prediction repository is used to store the basic class prediction results and update the data in the repository through batch momentum sliding. The semantic calibration calculation module, in conjunction with the dual dynamic repository module, is used for nearest neighbor retrieval, generating neighborhood supervision signals and outputting semantic calibration loss. The debiased comparison calculation module is used for Gamma distribution sampling, dynamic fusion of weights, fusion of bi-branch predictions, construction of sample similarity matrix, and output of debiased comparison loss. The distribution alignment constraint module is used to filter high-confidence pseudo-label samples and calculate the class alignment loss in the form of cross-entropy. The joint optimization module is used to weight and fuse the three types of losses to obtain the debiased total loss, and combine it with the original loss of the basic GCD classification module to complete the network iterative training.

Images