Geometric calibration method and system for unsupervised long-tail data

Abstract

The present application provides a kind of geometric calibration method and system for unsupervised long tail data, it is related to self-supervised long tail learning technical field, including: geometric balance measurement step: predefine optimal geometric balance structure, measure the balance of feature space on coarse granularity;Proxy label assignment step: dynamic proxy label assignment is carried out to calibrate the distribution of head and tail data in feature space, and gradually realize class balance based on double-layer optimization training mode.The present application introduces geometric balance structure, calibrates long tail data representation by geometric label redistribution, further strengthens the class balance and migratability of representation, and achieves significant effect improvement in a variety of complex skewed distribution upstream and downstream scenarios.

Description

Technical Field

[0001] This invention relates to the field of self-supervised long-tail learning technology, and more specifically, to a geometric calibration method and system for unsupervised long-tail data. Background Technology

[0002] In recent years, self-supervised learning algorithms have made significant progress in computer vision and natural language understanding, demonstrating their excellent ability to mine general representations and their applicability to a wide range of downstream fields. This rapid progress is mainly attributed to the efficient training on unlabeled data, which can be collected on a large scale. However, real-world natural resources often exhibit a long-tailed distribution, and learning representations directly on this data can lead to a bias problem in the representation space, where the majority of samples dominate the feature space, while a minority of samples are compressed. Therefore, especially as fairness becomes increasingly important in machine learning, it is necessary to focus on the problem of self-supervised long-tailed learning.

[0003] Supervised long-tail learning algorithms have flourished, while self-supervised long-tail learning urgently needs further research. A search reveals that BCL (a paper published at the 2022 International Conference on Machine Learning (ICML): Contrastive Learning with Boosted Memorization, publicly available at: https: / / proceedings.mlr.press / v162 / zhou22l.html) drives instance-level data augmentation based on the memory differences between head and tail categories in deep neural networks. SDCLR (a paper published at the 2021 International Conference on Machine Learning (ICML): Self-Damaging Contrastive Learning, publicly available at: http: / / proceedings.mlr.press / v139 / jing21a.html) finds, by comparing feature encoders and their pruned versions, that the augmented hard samples mainly cover tail samples, thus implicitly enhancing the learning preference for tail samples. DnC (paper published at the 2021 International Conference on Computer Vision (ICCV): Divide and Contrast: Self-Supervised Learning From Uncurated Data, publicly available at: https: / / openaccess.thecvf.com / content / ICCV2021 / html / Tian_Divide_and_Contrast_Self-Supervised_Learning_From_Uncurated_Data_ICCV_2021_paper.html) employs a divide-and-conquer approach to mitigate the inherent heterogeneity of the data and avoid representation collapse of tail categories. rwSAM (paper published at the 2022 International Conference on Representation Learning (ICLR): Self-supervised Learning is More Robust to Dataset Imbalance, publicly available at: https: / / openreview.net / forum?id=4AZz9osqrar¬eId=vwXjA-Os4AW) uses a data-aware sharpness-perceived minimization scheme, implicitly regularizing tail samples during the optimization process. However, to date, few studies have taken into account the inherent limitations of the widely adopted contrastive learning loss and designed corresponding calibration mechanisms to improve class balance in representation learning.

[0004] In summary, the drawback of existing technologies is that they fail to consider the inherent flaw of self-supervised contrastive loss, thus failing to fundamentally solve the representation bias problem induced by long-tail data. Specifically, the goal of self-supervised representation learning is to uniformly embed data representations throughout the feature space, achieving sample balance. Under balanced data, sample balance ensures that each class is evenly distributed within the feature space, equivalent to class balance; however, under long-tail data, sample balance can lead to overexpansion of head classes and forced collapse of tail classes. However, an ideal, more generalizable representation space requires each class to evenly divide the entire feature space, achieving class balance, which is difficult for existing self-supervised representation learning methods to achieve. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a geometric calibration method and system for unsupervised long-tail data.

[0006] According to the present invention, a geometric calibration method and system for unsupervised long-tail data are provided, the scheme of which is as follows:

[0007] Firstly, a geometric calibration method for unsupervised long-tailed data is provided, the method comprising:

[0008] Geometric equilibrium measurement steps: Predefine the optimal geometric equilibrium structure and measure the equilibrium of the feature space at a coarse-grained level;

[0009] Proxy label allocation steps: Dynamic proxy label allocation is performed to calibrate the distribution of head and tail data in the feature space, and class balance is gradually achieved based on a two-layer optimization training method.

[0010] Preferably, the geometric equilibrium measurement step includes:

[0011] Introducing a geometrically balanced topology In this structure, any two vectors satisfy the property that the angle between them is equal, i.e., U i U j =C, i≠j and C is a certain constant, this definition implicitly implies the property of class balance;

[0012] If the geometric equilibrium structure U is fixed as the classifier during the training process, and the true label y of the sample is known, then each sample is directly learned through cross-entropy loss, so that the sample representation of each category approaches the geometric equilibrium class center of the corresponding category.

[0013] Preferably, the proxy label assignment step includes: generating proxy labels to analyze and correct the biased representation space;

[0014] The following geometric correction loss is proposed to generate surrogate labels:

[0015]

[0016] in, This represents the proxy labels generated for images 1 through N in the training dataset; Represents the geometric correction loss function; Represents the training dataset; x i q represents the i-th image in the training dataset; i Represents the geometric label of the i-th image in the training dataset; This represents the proxy label generated for the i-th image in the training dataset; Represents a vector of dimension N. This represents a vector of dimension K, where π represents the regularization of the category distribution of the agent label;

[0017] In the above formula, during the coarse-grained measurement stage, the distribution information π calculated based on the geometric equilibrium structure U for the feature space is used as a surrogate label. Marginal distribution constraints are used to adaptively and dynamically estimate the true distribution information of the data; in the fine-grained correction stage, instance-level clustering is performed to optimize the correlation between geometric labels and true labels;

[0018] By integrating geometric correction loss into the existing contrastive learning paradigm, and iteratively solving the model parameters and surrogate labels through a corresponding two-layer optimization method, the class balance is gradually reconstructed from the sample balance, thereby achieving long-tail self-balancing of the representation space.

[0019] Secondly, a geometric calibration system for unsupervised long-tail data is provided, the system comprising:

[0020] Geometric equilibrium measurement module: predefines the optimal geometric equilibrium structure to measure the equilibrium of the feature space at a coarse-grained level;

[0021] Proxy label allocation module: performs dynamic proxy label allocation to calibrate the distribution of head and tail data in the feature space, and gradually achieves class balance based on a two-layer optimization training method.

[0022] Preferably, the geometric equilibrium measurement module includes:

[0023] Introducing a geometrically balanced topology In this structure, any two vectors satisfy the property that the angle between them is equal, i.e., U i U j =C, i≠j and C is a certain constant, this definition implicitly implies the property of class balance;

[0024] If the geometric equilibrium structure U is fixed as the classifier during the training process, and the true label y of the sample is known, then each sample is directly learned through cross-entropy loss, so that the sample representation of each category approaches the geometric equilibrium class center of the corresponding category.

[0025] Preferably, the proxy label allocation module includes: generating proxy labels to analyze and correct the biased representation space;

[0026] The following geometric correction loss is proposed to generate surrogate labels:

[0027]

[0028] in, This represents the proxy labels generated for images 1 through N in the training dataset; Represents the geometric correction loss function; Represents the training dataset; x i q represents the i-th image in the training dataset; i Represents the geometric label of the i-th image in the training dataset; This represents the proxy label generated for the i-th image in the training dataset; Represents a vector of dimension N. This represents a vector of dimension K, where π represents the regularization of the category distribution of the agent label;

[0029] In the above formula, during the coarse-grained measurement stage, the distribution information π calculated based on the geometric equilibrium structure U for the feature space is used as a surrogate label. Marginal distribution constraints are used to adaptively and dynamically estimate the true distribution information of the data; in the fine-grained correction stage, instance-level clustering is performed to optimize the correlation between geometric labels and true labels;

[0030] By integrating geometric correction loss into the existing contrastive learning paradigm, and iteratively solving the model parameters and surrogate labels through a corresponding two-layer optimization method, the class balance is gradually reconstructed from the sample balance, thereby achieving long-tail self-balancing of the representation space.

[0031] Thirdly, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the geometric calibration method for unsupervised long-tail data.

[0032] Fourthly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps in the geometric calibration method for unsupervised long-tail data.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention has a wide range of applications and can be combined with almost all existing self-supervised contrastive learning algorithms to significantly improve the representation balance of the algorithm on self-supervised long-tail learning problems;

[0035] 2. This invention is simple to deploy, low in cost, and introduces very little additional computation, which is negligible compared to the training of deep neural networks;

[0036] 3. The solution of this invention has the theoretical guarantee of convergence to class balance, which can effectively suppress the representation bias caused by the long tail of data;

[0037] 4. This invention achieves excellent self-supervised representation learning performance on multiple benchmark long-tail datasets through long-tail robust geometric calibration technology, and still achieves outstanding results on large-scale network datasets and fine-grained datasets.

[0038] Other beneficial effects of the present invention will be explained in detail through the introduction of specific technical features and technical solutions in specific embodiments. Those skilled in the art should be able to understand the beneficial technical effects brought about by these technical features and technical solutions through the introduction of these technical features and technical solutions. Attached Figure Description

[0039] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0040] Figure 1 The flowchart shows the geometric calibration algorithm based on self-supervised representation learning of long-tail data.

[0041] Figure 2 This is the overall structure of a geometric calibration algorithm based on self-supervised representation learning of long-tail data. Detailed Implementation

[0042] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0043] This invention provides a geometric calibration method for unsupervised long-tail data, referring to... Figure 1 As shown, the method specifically includes:

[0044] Geometric equilibrium measurement steps: Predefine the optimal geometric equilibrium structure and measure the equilibrium of the feature space at the coarse-grained level.

[0045] This step specifically includes: introducing a geometrically balanced topology. In this structure, any two vectors satisfy the property that the angle between them is equal, i.e., U i U j =C, i≠j and C is a certain constant, this definition implicitly implies the property of class balance;

[0046] If the geometric equilibrium structure U is fixed as the classifier during the training process, and the true label y of the sample is known, then each sample is directly learned through cross-entropy loss, so that the sample representation of each category approaches the geometric equilibrium class center of the corresponding category.

[0047] Proxy label allocation steps: Dynamic proxy label allocation is performed to calibrate the distribution of head and tail data in the feature space, and class balance is gradually achieved based on a two-layer optimization training method.

[0048] This step specifically includes: generating surrogate labels to analyze and correct the biased representation space;

[0049] The following geometric correction loss is proposed to generate surrogate labels:

[0050]

[0051] in, This represents the proxy labels generated for images 1 through N in the training dataset; Represents the geometric correction loss function; Represents the training dataset; x i q represents the i-th image in the training dataset; i Represents the geometric label of the i-th image in the training dataset; This represents the proxy label generated for the i-th image in the training dataset; Represents a vector of dimension N. This represents a vector of dimension K, where π represents the regularization of the category distribution of the agent label;

[0052] In the above formula, during the coarse-grained measurement stage, the distribution information π calculated based on the geometric equilibrium structure U for the feature space is used as a surrogate label. Marginal distribution constraints are used to adaptively and dynamically estimate the true distribution information of the data; in the fine-grained correction stage, instance-level clustering is performed to optimize the correlation between geometric labels and true labels;

[0053] By integrating geometric correction loss into the existing contrastive learning paradigm, and iteratively solving the model parameters and surrogate labels through a corresponding two-layer optimization method, the class balance is gradually reconstructed from the sample balance, thereby achieving long-tail self-balancing of the representation space.

[0054] The present invention also provides a geometric calibration system for unsupervised long-tail data. The geometric calibration system for unsupervised long-tail data can be implemented by executing the process steps of the geometric calibration method for unsupervised long-tail data. That is, those skilled in the art can understand the geometric calibration method for unsupervised long-tail data as a preferred embodiment of the geometric calibration system for unsupervised long-tail data.

[0055] Geometric equilibrium measurement module: predefines the optimal geometric equilibrium structure to measure the balance of the feature space at a coarse-grained level.

[0056] This module specifically includes: introducing a geometrically balanced topology. In this structure, any two vectors satisfy the property that the angle between them is equal, i.e., U i U j =C, i≠j and C is a certain constant, this definition implicitly implies the property of class balance;

[0057] If the geometric equilibrium structure U is fixed as the classifier during the training process, and the true label y of the sample is known, then each sample is directly learned through cross-entropy loss, so that the sample representation of each category approaches the geometric equilibrium class center of the corresponding category.

[0058] Proxy label allocation module: performs dynamic proxy label allocation to calibrate the distribution of head and tail data in the feature space, and gradually achieves class balance based on a two-layer optimization training method.

[0059] This module specifically includes: generating surrogate labels to analyze and correct the biased representation space;

[0060] The following geometric correction loss is proposed to generate surrogate labels:

[0061]

[0062] in, This represents the proxy labels generated for images 1 through N in the training dataset; Represents the geometric correction loss function; Represents the training dataset; x i q represents the i-th image in the training dataset; i Represents the geometric label of the i-th image in the training dataset; This represents the proxy label generated for the i-th image in the training dataset; Represents a vector of dimension N. Let π represent a vector of dimension K, and let π represent the regularization of the category distribution of the agent label.

[0063] In the above formula, during the coarse-grained measurement stage, the distribution information π calculated based on the geometric equilibrium structure U for the feature space is used as a surrogate label. Marginal distribution constraints are used to adaptively and dynamically estimate the true distribution information of the data; in the fine-grained correction stage, instance-level clustering is performed to optimize the correlation between geometric labels and true labels.

[0064] By integrating geometric correction loss into the existing contrastive learning paradigm, and iteratively solving the model parameters and surrogate labels through a corresponding two-layer optimization method, the class balance is gradually reconstructed from the sample balance, thereby achieving long-tail self-balancing of the representation space.

[0065] The present invention will now be described in more detail.

[0066] This invention provides a geometric calibration method for unsupervised long-tail data, referring to... Figure 1 As shown, the method specifically includes: geometric equilibrium measurement and proxy label allocation.

[0067] In the geometric equilibrium measurement section, an optimal geometric equilibrium structure is predefined to measure the coarse-grained equilibrium of the feature space. Specifically, we first apply a contrastive loss function to perform preliminary self-supervised training on the model, using a predefined geometric equilibrium structure U, and based on samples x... i The initial representation of the geometric position outputs the geometric label q. i Furthermore, the class distribution information π of the geometric labels in the training data is statistically analyzed to achieve a geometric balance measurement in the feature space. If the geometrically balanced structure u is regarded as a classifier, this process can be considered a standard representation classification process.

[0068] In the proxy label assignment part, we explicitly compress the overexpansion of the head class and limit the passive collapse of the tail class by implementing fine-grained instance-level calibration. Specifically, we fine-grainedly adjust the assignment based on the geometric label q of each sample. i Perform geometric correction loss optimization and generate corresponding surrogate labels. The geometric correction loss is a surrogate label. With geometric label q i A variant of cross-entropy loss generates surrogate labels constrained by the class distribution π, thereby correcting and optimizing the initially obtained noisy geometric labels. The calibrated surrogate labels effectively alleviate the phenomenon of excessive expansion of the head of the representation space and passive collapse of the tail classes.

[0069] Simultaneously, a two-layer optimization approach is proposed, involving additional optimal transport optimization to dynamically pursue class balance. Specifically, the geometric correction part of this invention involves the optimization process of surrogate labels and model parameters. The objective function of both optimization processes is the geometric correction loss function, but the optimization objects are the surrogate labels and model parameters, respectively. Surrogate label optimization obtains higher-quality, corrected geometric labels, while model parameter optimization updates the sample representation, making the sample representation tend towards the corrected surrogate labels. The geometric labels before correction are the representation space learned by contrastive loss, reflecting sample balance, while the corrected surrogate labels can effectively alleviate the phenomenon of excessive expansion of the head of the representation space and passive collapse of the tail classes, thereby achieving class balance. In the framework of two-layer optimization, we first perform preliminary optimization of the model, and then iteratively optimize the surrogate labels and model parameters in each training generation.

[0070] This invention is an improvement on the standard self-supervised contrastive learning algorithm. The learning objective of self-supervised contrastive loss is to narrow the distance between different perspectives of positive samples in the feature space, while widening the distance between negative samples. However, this loss is prone to inducing representation bias in long-tailed data. This invention performs geometric correction at the feature space level, introducing class balance regularization through dynamic geometric equalization calibration. The overall structure is shown below. Figure 2 Overall structure of a geometric calibration algorithm based on self-supervised representation learning of long-tail data.

[0071] This invention is based on the observation that basic self-supervised contrastive training techniques, under long-tailed data distributions, ultimately optimize to sample balance, leading to representation bias. This phenomenon results in a significant difference in the spatial size of the feature space trained by self-supervised contrastive learning across the first and last classes, limiting the downstream generalization of the representation. Therefore, after analyzing the representation bias problem under long-tailed distributions, this invention proposes a novel geometric calibration technique. This technique uses a predefined optimal geometric balance structure to perform coarse-grained distribution estimation of the feature space and dynamically assigns surrogate labels to calibrate the distribution of first and last data in the feature space. Based on a two-layer optimization training method, class balance is gradually achieved to counteract the negative impact of long-tailed data on representation learning.

[0072] In terms of implementation, this invention first introduces a geometrically balanced topology. In this structure, any two vectors satisfy the property that the angle between them is equal, i.e., U i U j=C, i≠j, and C is a constant. Thus, this definition implicitly contains the property of class balance. If we fix the geometrically balanced structure U as the classifier during training, assuming the true label y of the sample is known, we can directly learn each sample through cross-entropy loss, making the sample representations of each class approach the geometrically balanced class center of the corresponding class. Because the class centers satisfy the property of equal subtended angles, we can constrain the feature space to approximate sample balance.

[0073] However, the technical challenge here lies in the lack of label information supervision in self-supervised scenarios, thus requiring the generation of reliable surrogate labels to analyze and correct the biased representation space. To address this challenge, we propose the following geometric correction loss to generate surrogate labels:

[0074]

[0075] In the above equation, during the coarse-grained measurement stage, we use the distribution information π calculated for the feature space based on the geometric equilibrium structure U as a surrogate label. Marginal distribution constraints are used to adaptively and dynamically estimate the true distribution information of the data. In the fine-grained correction stage, instance-level clustering is performed to optimize the correlation between geometric labels and true labels. This invention can flexibly integrate geometric correction loss into existing contrastive learning paradigms, iteratively solving model parameters and surrogate labels through a corresponding two-layer optimization method, gradually reconstructing class balance from sample balance, and achieving long-tail self-balancing of the representation space. Specifically, this process involves the optimization of surrogate labels and model parameters. The objective function of both optimization processes is the geometric correction loss function, but the optimization objects are surrogate labels and model parameters, respectively. Surrogate label optimization obtains higher-quality, corrected geometric labels, and model parameter optimization updates the sample representation, making the sample representation tend towards the corrected surrogate labels. The geometric labels before correction are the representation space obtained by contrastive loss learning, reflecting sample balance, while the corrected surrogate labels can effectively alleviate the phenomenon of excessive expansion of the head of the representation space and passive collapse of the tail classes, thereby achieving class balance. In the framework of two-layer optimization, we first perform preliminary optimization of the model, and then iteratively optimize surrogate labels and model parameters in each training generation.

[0076] The theoretical motivation behind the algorithm design of this invention is the recent neural collapse theory for dealing with long-tailed data in supervised learning. This significantly improves upon the technical shortcomings of ordinary self-supervised contrastive learning. Geometric calibration techniques can effectively suppress representation bias caused by long-tailed data and achieve class balance. Class balance can effectively improve the generalization performance of representations when transferred downstream, thus theoretically guaranteeing the effectiveness of geometric calibration techniques in solving the self-supervised long-tail problem.

[0077] This invention employs self-supervised pre-training on various upstream datasets and is tested on different downstream datasets and tasks. It achieves state-of-the-art performance compared to various self-supervised long-tail learning algorithms, while also improving representation balance. The upstream datasets include five self-supervised long-tail benchmark datasets and a large-scale uncorrected network dataset. The downstream datasets include long-tail benchmark datasets and fine-grained classification datasets. Downstream tasks include classification, semantic detection, and instance segmentation. All experiments demonstrate the superior performance of this invention for self-supervised long-tail representation learning and its effectiveness in suppressing representation bias.

[0078] This invention provides a geometric calibration method and system for unsupervised long-tail data. It mainly focuses on the design of a geometric calibration algorithm for self-supervised representation learning under long-tail data. Based on the introduction of a geometric equilibrium structure, the long-tail data representation is calibrated by geometric label redistribution, which further enhances the class balance and transferability of the representation. Significant performance improvements have been achieved in various complex skewed distribution upstream and downstream scenarios.

[0079] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0080] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A geometric calibration method for unsupervised long-tail data, characterized in that, include: Geometric equilibrium measurement steps: Predefine the optimal geometric equilibrium structure and measure the equilibrium of the feature space at a coarse-grained level; Proxy label allocation steps: Perform dynamic proxy label allocation to calibrate the distribution of head and tail data in the feature space, and gradually achieve class balance based on a two-layer optimization training method; The geometric equilibrium measurement steps include: Introducing a geometrically balanced topology In this structure, any two vectors satisfy the property that their subtended angles are equal, i.e. and As a certain constant, this definition implicitly implies the property of class balance; Geometric equilibrium structure If the classifier is fixed during the training process, and the true label y of the sample is known, then each sample is directly learned through cross-entropy loss, so that the sample representation of each category approaches the geometric equilibrium class center of the corresponding category. The proxy label assignment step includes: generating proxy labels to analyze and correct the biased representation space; The following geometric correction loss is proposed to generate surrogate labels: in, Indicates the first in the training dataset Proxy labels generated from the image; Represents the geometric correction loss function; This represents the training dataset; Represents the first in the training dataset Zhang Image; Indicates the first in the training dataset Geometric labels for the images; Indicates the first in the training dataset Proxy labels generated from the image; Represents a vector of dimension N. Represents a vector of dimension K. This indicates a regularization for the category distribution of proxy tags; In the above equation, during the coarse-grained measurement stage, a geometrically balanced structure is utilized. Distribution information calculated from the feature space As a proxy label Marginal distribution constraints are used to adaptively and dynamically estimate the true distribution information of the data; in the fine-grained correction stage, instance-level clustering is performed to optimize the correlation between geometric labels and true labels; By integrating geometric correction loss into the existing contrastive learning paradigm, and iteratively solving the model parameters and surrogate labels through a corresponding two-layer optimization method, the class balance is gradually reconstructed from the sample balance, thereby achieving long-tail self-balancing of the representation space.

2. A geometric calibration system for unsupervised long-tail data, characterized in that, include: Geometric equilibrium measurement module: predefines the optimal geometric equilibrium structure to measure the equilibrium of the feature space at a coarse-grained level; Proxy label allocation module: performs dynamic proxy label allocation to calibrate the distribution of head and tail data in the feature space, and gradually achieves class balance based on a two-layer optimization training method; The geometric equilibrium measurement module includes: Introducing a geometrically balanced topology In this structure, any two vectors satisfy the property that their subtended angles are equal, i.e. and As a certain constant, this definition implicitly implies the property of class balance; Geometric equilibrium structure If the classifier is fixed during the training process, and the true label y of the sample is known, then each sample is directly learned through cross-entropy loss, so that the sample representation of each category approaches the geometric equilibrium class center of the corresponding category. The proxy label allocation module includes: generating proxy labels to analyze and correct the biased representation space; The following geometric correction loss is proposed to generate surrogate labels: in, Indicates the first in the training dataset Proxy labels generated from the image; Represents the geometric correction loss function; This represents the training dataset; Represents the first in the training dataset Zhang Image; Indicates the first in the training dataset Geometric labels for the images; Indicates the first in the training dataset Proxy labels generated from the image; Represents a vector of dimension N. Represents a vector of dimension K. This indicates a regularization for the category distribution of proxy tags; In the above equation, during the coarse-grained measurement stage, a geometrically balanced structure is utilized. Distribution information calculated from the feature space As a proxy label Marginal distribution constraints are used to adaptively and dynamically estimate the true distribution information of the data; in the fine-grained correction stage, instance-level clustering is performed to optimize the correlation between geometric labels and true labels; By integrating geometric correction loss into the existing contrastive learning paradigm, and iteratively solving the model parameters and surrogate labels through a corresponding two-layer optimization method, the class balance is gradually reconstructed from the sample balance, thereby achieving long-tail self-balancing of the representation space.

3. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the geometric calibration method for unsupervised long-tailed data as described in claim 1.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the geometric calibration method for unsupervised long-tailed data as described in claim 1.

Images