Small sample continuous learning method and system

Abstract

The invention provides a small sample continuous learning method and system, and the method comprises the steps: obtaining a plurality of pieces of newly-added sample data comprising picture samples and text samples, and obtaining guide features based on the newly-added sample data; performing enhanced sampling on the picture sample to obtain a plurality of enhanced samples, and extracting enhanced features of the enhanced samples; and combining the guiding feature and the plurality of enhanced features into a training set, and constructing a continuous differential optimizer to perform continuous optimization on model parameters of the continuous differential optimizer. By obtaining the enhanced sample, diversified training data can be constructed on the premise of ensuring category consistency, and the situation that the recognition accuracy is reduced due to insufficient training data in the continuous learning process of small samples is improved; according to the method, model parameters are continuously optimized through a plurality of training sets, a parameter updating process is modeled as a continuous dynamic system, smooth and stable parameter optimization is performed according to a continuous dynamic rule, a historical knowledge structure is effectively maintained, and disastrous forgetting is remarkably relieved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a few-shot continuous learning method and system. Background Technology

[0002] With the development of deep learning technology, significant progress has been made in tasks such as image classification and object recognition. However, traditional deep learning models typically rely on large-scale, static, and fully labeled datasets, and obtain fixed model parameters through one-time training. When the model encounters new categories or data distributions in subsequent stages, it often needs to be retrained or fine-tuned, making it difficult for the model to adapt to the continuously emerging new knowledge in real-world scenarios. Furthermore, directly updating the model on new tasks can easily lead to "catastrophic forgetting," meaning that the model's ability to recognize already learned categories significantly decreases when learning new categories.

[0003] To address this issue, continuous learning has gradually gained attention. Among the many continuous learning scenarios, few-shot continuous learning (FSCL) is a more challenging task. This task requires the model to have relatively sufficient basic class data in the initial stage, but only a very small number of samples for each class in subsequent incremental sessions. This makes it necessary for the model to extract effective features from the limited new samples while maintaining old knowledge.

[0004] However, existing small-sample continuous learning still has the following drawbacks: 1. The model relies on discrete batch optimization processes in incremental sessions, which cannot characterize the continuous evolution of learning behavior, resulting in unstable optimization trajectories and thus causing historical knowledge to be forgotten; 2. The number of new session samples is extremely small, and repeated use of a small amount of data for training will result in a single training signal and a decrease in recognition accuracy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a few-shot continuous learning method and system, which solves the technical problems of existing technologies where few-shot continuous learning relies on discrete batch optimization processes in incremental sessions, fails to characterize the continuous evolution of learning behavior, leads to the forgetting of historical knowledge, and has extremely few new session samples, resulting in insufficient training data.

[0006] To achieve the above objectives, firstly, embodiments of this application provide a small-sample continuous learning method, comprising the following steps: Acquire several new sample data that arrive in the order of the session. The new sample data includes image samples and text samples corresponding to the image samples. Based on the new sample data, obtain guidance features. The image sample is augmented to obtain several augmented samples, and the augmentation features of the augmented samples are extracted. The augmentation features correspond to the guidance features. The guidance features and several enhancement features are combined into a training set, and a continuous differential optimizer is constructed. The model parameters of the continuous differential optimizer are continuously optimized based on several training sets, so that the continuous differential optimizer has the ability to recognize several new samples.

[0007] Furthermore, the step of obtaining guiding features based on the newly added sample data includes: Obtain the image features corresponding to the image sample, and obtain the text features corresponding to the text sample; The image features and text features are combined into guidance features based on the condition generator.

[0008] Furthermore, the step of performing augmented sampling on the image samples to obtain several augmented samples includes: Extract the original latent space features of the image samples, and apply random perturbation to the original latent space features to obtain updated latent space features; The guiding features are used as the first conditional signal of the noise prediction network model, and the updated latent space features are used as the input values ​​of the noise prediction network model to obtain the predicted noise features. Enhanced samples are obtained based on the updated latent space features and the predicted noise features.

[0009] Furthermore, the step of continuously optimizing the model parameters of the continuous differential optimizer based on several training sets includes: The guiding features of the training set are used as the second condition signal of the continuous differential optimizer, and the enhancement features of the training set are used as the input value of the continuous differential optimizer, so as to obtain the dynamic law of the model parameters of the continuous differential optimizer. The model parameters are continuously updated based on the dynamic pattern to complete a single stage update and obtain the stage model parameters.

[0010] Furthermore, after the step of continuously updating the model parameters based on the dynamic law to complete a single stage update and obtaining the stage model parameters, the method further includes: The model parameters of the previous stage are used as the model parameters of the next training set to repeat the single stage update and complete continuous optimization to obtain the final model parameters.

[0011] Furthermore, the step of using the enhanced features of the training set as input values ​​for the continuous differential optimizer to obtain the dynamic laws governing the model parameters of the continuous differential optimizer includes: Obtain the predicted loss for each of the enhanced features in the training set, and obtain the average loss based on the predicted losses; Backpropagation is performed on the average loss to obtain the average gradient; The dynamic patterns corresponding to the model parameters are obtained based on the average gradient.

[0012] Furthermore, the formula for obtaining the prediction loss is: , in, This represents the prediction loss of the i-th augmented feature in the t-th training set at the m-th time point. This represents the i-th augmented feature in the t-th training set. Indicates guiding characteristics, This represents the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. This represents the true label corresponding to the t-th training set. This indicates that prediction is performed using a continuous differential optimizer. Represents the loss function; The expression for the dynamic law is: , in, This represents the dynamic behavior of the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. Indicates the learning rate. Let represent the average gradient of the t-th training set at the m-th time point.

[0013] Furthermore, the formula for obtaining the stage model parameters is as follows: , in, This indicates that the t-th training set is in the t-th position. The stage model parameters of the continuous differential optimizer at each time point This represents the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. This represents the dynamic behavior of the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. Indicates the time step.

[0014] Secondly, embodiments of this application provide a few-shot continuous learning system applied to the few-shot continuous learning method described in the first aspect above, the system comprising: The identification module is used to acquire several new sample data that arrive in the order of the session. The new sample data includes image samples and text samples corresponding to the image samples. Guidance features are acquired based on the new sample data. An extension module is used to perform augmented sampling on the image samples to obtain several augmented samples, and to extract the augmented features of the augmented samples, wherein the augmented features correspond to the guidance features; The execution module is used to combine the guidance features and several enhancement features into a training set, and construct a continuous differential optimizer. Based on the several training sets, the model parameters of the continuous differential optimizer are continuously optimized so that the continuous differential optimizer has the ability to recognize several new samples.

[0015] Thirdly, embodiments of this application provide a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the few-sample continuous learning method as described in the first aspect above.

[0016] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the few-sample continuous learning method as described in the first aspect above.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: by acquiring the enhanced samples, diverse training data can be constructed under the premise of ensuring category consistency with a small number of image samples, thereby improving the reduction in recognition accuracy caused by insufficient training data during continuous learning of small samples; by continuously optimizing the model parameters through several training sets, the parameter update process is modeled as a continuous dynamic system, and smooth and stable parameter optimization is performed with continuous dynamic laws, effectively maintaining the historical knowledge structure and significantly alleviating catastrophic forgetting. Attached Figure Description

[0018] Figure 1 This is a flowchart of the few-sample continuous learning method in the first embodiment of the present invention; Figure 2 This is a visualization comparison of the confusion matrix of the recognition results of the few-shot continuous learning method and the traditional method in the first embodiment of the present invention, where a is the traditional method and b is the method in the first embodiment; Figure 3 This is a structural block diagram of the few-shot continuous learning system in the second embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0020] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Please see Figure 1 The first embodiment of the present invention provides a few-shot continuous learning method, which includes the following steps: S10: Obtain several newly added sample data that arrive in the order of the session. The newly added sample data includes image samples and text samples corresponding to the image samples. Obtain guidance features based on the newly added sample data. The newly added sample data refers to the category that needs to be newly learned. The image sample and the text sample represent the same content. For example, the image sample is a picture of a dog, and the text sample is a picture of a dog. The purpose of continuous learning with few samples is to enable the ability to recognize pictures of dogs after training is completed.

[0023] Step S10 includes: S110: Obtain the image features corresponding to the image sample, and obtain the text features corresponding to the text sample; S120: Based on the condition generator, combine the image features and the text features into a guidance feature; In this embodiment, the image sample is converted into the image feature by an image encoder, and the text sample is converted into the text feature by a text encoder. Then, the two are combined into the guidance feature by the condition generator. The guidance feature includes semantic-level features and image-level features, forming a comprehensive representation of the newly added sample data.

[0024] S20: Perform augmented sampling on the image sample to obtain several augmented samples, extract the augmented features of the augmented samples, and the augmented features correspond to the guidance features; In few-shot continuous learning, the number of image samples in the newly added sample data is relatively small. Therefore, to enable more accurate recognition and classification capabilities, it is necessary to augment the data based on a small number of image samples to ensure sufficient training data. Step S20 includes: S210: Extract the original latent space features of the image sample, and apply random perturbation to the original latent space features to obtain updated latent space features; The random perturbation refers to adding a small amount of noise to the original latent space features to distinguish them from the original latent space features.

[0025] S220: Using the guiding features as the first conditional signal of the noise prediction network model, and using the updated latent space features as the input value of the noise prediction network model, to obtain the predicted noise features; The noise prediction network model is used to predict, based on the first conditional signal, the noise that needs to be removed to make the image with updated latent space features clearer, i.e., the predicted noise features.

[0026] S230: Obtain enhanced samples based on the updated latent space features and the predicted noise features; The guiding feature is set as the first conditional signal to ensure that the output of the noise prediction network model allows the subsequently acquired updated latent space features to be semantically faithful to the original latent space features, but with differences from the image samples at the pixel level. In this way, richer and more diverse data can be provided to train the continuous differential optimizer while ensuring semantic alignment, thus ensuring its recognition accuracy. Understandably, by subtracting the predicted noise features from the updated latent space features, enhanced space features are obtained. These enhanced space features are then input into the decoder to be transformed into pixel space, forming the enhanced sample.

[0027] S30: Combine the guidance features and several enhancement features into a training set, and construct a continuous differential optimizer. Continuously optimize the model parameters of the continuous differential optimizer based on several training sets, so that the continuous differential optimizer has the ability to recognize several new samples. It should be noted that each of the newly added sample data corresponds to one of the training sets.

[0028] Step S30 includes: S310: Using the guiding features of the training set as the second condition signal of the continuous differential optimizer, and using the enhancement features of the training set as the input value of the continuous differential optimizer, to obtain the dynamic law of the model parameters of the continuous differential optimizer. Specifically, the predicted loss of each of the enhanced features in the training set is obtained, and an average loss is obtained based on several of the predicted losses; The formula for obtaining the predicted loss is: , in, This represents the prediction loss of the i-th augmented feature in the t-th training set at the m-th time point. This represents the i-th augmented feature in the t-th training set. Indicates guiding characteristics, This represents the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. This represents the true label corresponding to the t-th training set. This indicates that prediction is performed using a continuous differential optimizer. This represents the loss function.

[0029] Backpropagation is performed on the average loss to obtain the average gradient; After obtaining the average loss, starting from the average loss, the chain rule is used to perform a reverse chain of derivatives to obtain the average gradient.

[0030] The dynamic patterns corresponding to the model parameters are obtained based on the average gradient. The expression for the dynamic law is: , in, This represents the dynamic behavior of the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. Indicates the learning rate. Let represent the average gradient of the t-th training set at the m-th time point.

[0031] S320: Based on the dynamic law, the model parameters are continuously updated to complete a single stage update and obtain the stage model parameters; The formula for obtaining the stage model parameters is: , in, This indicates that the t-th training set is in the t-th position. The stage model parameters of the continuous differential optimizer at each time point This represents the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. This represents the dynamic behavior of the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. Indicates the time step.

[0032] Preferably, step S30 further includes: S330: Use the model parameters of the stage as the model parameters of the next training set to repeat the single stage update and complete continuous optimization to obtain the final model parameters; That is, when processing several training sets (new sample data), each time a new identifiable category is learned, the parameters are updated based on the model parameters after the previous learning to ensure its ability to continuously memorize old knowledge.

[0033] By acquiring the enhanced samples, even with a small number of image samples, diverse training data can be constructed while ensuring category consistency, thus improving the reduced recognition accuracy caused by insufficient training data during continuous learning with small samples. By continuously optimizing the model parameters through several training sets, the parameter update process is modeled as a continuous dynamic system, and smooth and stable parameter optimization is performed with continuous dynamic laws, effectively maintaining the historical knowledge structure and significantly alleviating catastrophic forgetting.

[0034] This application is validated on commonly used datasets for classification tasks, such as CIFAR100, CUB200, and miniImageNet, and compared with other learning methods. Specifically: For the CIFAR100 and miniImageNet datasets, all training categories are first divided into 60 base categories and 40 new categories. In the base stage, all base class samples are used to train the initial model. Then, the 40 new categories are divided into eight consecutive sessions, each using a 5-way 5-shot input of new category samples to simulate a few-shot incremental learning environment. On the CUB200 dataset, all 200 classes are divided into 100 base categories and 100 new categories. The new categories are introduced in ten 10-way 5-shot sessions to verify the method's generalization ability in fine-grained scenarios. Regarding network structure, this embodiment uses ResNet20 as the base feature extraction network for CIFAR100 and ResNet18 as the feature extraction network for miniImageNet and CUB200 to ensure a stable and consistent encoding method across different datasets. The continuous differential optimizer is constructed from three layers of multilayer perceptrons to continuously model the gradient flow during model parameter updates, ensuring a smooth and stable evolution trajectory of model parameters across different sessions. This embodiment uses a pre-trained StableDiffusionV1.4 as the augmented sampling model, freezing all its parameters. Conditional control is achieved through the prompt "The image depicts itself" to construct a stable reconstruction prior from the latent space. The balance coefficient γ and learning rate λ are set to 0.4 and 0.6, respectively, and the continuous differential solution time step M is set to 10. Augmented sampling strategies include random cropping, scaling, and horizontal flipping. During training, the system first uses basic classes to train and obtain initial model parameters. Then, a limited number of new class samples are input in each incremental session. After training, the system calculates three metrics—Avg, ΔFinal, and KR—based on the Top-1 accuracy of each session. Avg represents the average performance across sessions, ΔFinal represents the performance difference between the final and initial sessions, and KR = AccT / Acc1 represents the knowledge retention rate. The experimental results are shown in Tables 1 (miniImageNet), 2 (CIFAR100), and 3 (CUB200). Table 1 , Table 2 , Table 3 , As shown in the table above, the method in this application achieves superior performance compared to existing technologies on various datasets, demonstrating that the synergistic mechanism of continuous differential optimizer and augmented sampling can significantly improve the stability and generalization ability of continuous learning with few samples.

[0035] Ablation experiments were conducted on the continuous differential optimizer and enhanced sampling respectively. The experiments were performed under the same small-sample continuous learning setting, maintaining complete consistency in network structure, incremental session partitioning, and training strategy to ensure fair comparison. The results are shown in Table 4 below: Table 4 , Ablation analysis results show that when only the basic continuous learning model is used, the model performance declines significantly with the number of sessions, especially in later sessions where it struggles to maintain the ability to recognize old categories. Adding a continuous differential optimizer (CDO) significantly improves both average performance and knowledge retention, indicating that continuous parameter evolution effectively mitigates the oscillations and knowledge forgetting caused by discrete optimization, enabling the model to integrate new information more stably when new categories arrive. Furthermore, adding augmented sampling (ECS) alone also yields significant benefits. The model achieves optimal performance when both are enabled simultaneously. Its average performance and knowledge retention are significantly higher than those achieved with either module enabled alone, demonstrating a complementary relationship between continuous differential optimization and augmented sampling: the former provides a stable parameter evolution trajectory, while the latter provides greater data diversity; their synergy maximizes the reduction of forgetting and improves the efficiency of new knowledge absorption.

[0036] Please see Figure 2 To further analyze the differences in the accuracy of new and old category recognition in this application, a confusion matrix visualization comparison was performed on the final session classification results of the basic few-shot continuous learning and the few-shot continuous learning in this application. The diagonal region of the confusion matrix reflects the classification accuracy, while the off-diagonal region represents the degree of confusion. The last few categories in the matrix correspond to the new categories introduced during the incremental process. From the comparison of the confusion matrices, it can be observed that the method of this invention exhibits a more obvious, clear, and concentrated diagonal structure in the new category region, indicating that it can accurately identify new category samples under conditions of very few samples, with a concentrated prediction distribution and clear classification boundaries. In contrast, the basic model shows obvious scattering in this region, with a large number of off-diagonal elements, indicating that it does not learn the new category sufficiently and has a high confusion rate. In the old category region, this application also shows stronger stability. The confusion matrix shows that this application can maintain the topological structure of the old category feature space, enabling it to maintain a high accuracy in the final session, with a significant reduction in off-diagonal misclassification. Overall, the confusion matrix shows that this application has higher discriminative power and lower confusion in both new and old category regions, proving that this application can simultaneously improve the adaptability to new categories and the ability to retain knowledge of old categories, further verifying the stability and reliability of the few-sample continuous learning method in this application.

[0037] Please see Figure 3The second embodiment of the present invention provides a few-shot continuous learning system, which is applied to the few-shot continuous learning method described in the above embodiments, and will not be repeated hereafter. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0038] The system includes: The identification module 10 is used to acquire several newly added sample data that arrive in the order of the session. The newly added sample data includes image samples and text samples corresponding to the image samples. Based on the newly added sample data, guidance features are acquired. The identification module 10 includes: The first unit is used to obtain image features corresponding to the image sample and text features corresponding to the text sample; The second unit is used to combine the image features and the text features into guidance features based on a condition generator; The extension module 20 is used to perform enhanced sampling on the image sample to obtain several enhanced samples, and to extract the enhanced features of the enhanced samples, wherein the enhanced features correspond to the guidance features; The expansion module 20 includes: The third unit is used to extract the original latent space features of the image samples and apply random perturbation to the original latent space features to obtain updated latent space features; The fourth unit is used to obtain the predicted noise features by using the guiding features as the first conditional signal of the noise prediction network model and the updated latent space features as the input values ​​of the noise prediction network model. The fifth unit is used to obtain enhanced samples based on the updated latent space features and the predicted noise features; The execution module 30 is used to combine the guidance features and several enhancement features into a training set, and construct a continuous differential optimizer. Based on several training sets, the model parameters of the continuous differential optimizer are continuously optimized so that the continuous differential optimizer has the ability to recognize several new samples. The execution module 30 includes: The sixth unit is used to take the guiding features of the training set as the second condition signal of the continuous differential optimizer and the enhancement features of the training set as the input value of the continuous differential optimizer, so as to obtain the dynamic law of the model parameters of the continuous differential optimizer. The sixth unit is specifically used to obtain the prediction loss of each of the enhanced features in the training set, obtain the average loss based on several prediction losses, perform backpropagation on the average loss to obtain the average gradient, and obtain the dynamic pattern corresponding to the model parameters based on the average gradient. The seventh unit is used to continuously update the model parameters based on the dynamic law to complete a single stage update and obtain the stage model parameters. The execution module 30 further includes: The eighth unit is used to take the stage model parameters as the model parameters of the next training set, repeat the single stage update, and complete continuous optimization to obtain the final model parameters.

[0039] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the few-sample continuous learning method as described in the above technical solutions.

[0040] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the small-sample continuous learning method as described in the above technical solutions.

[0041] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0042] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A few-shot continuous learning method, characterized in that, Includes the following steps: Acquire several new sample data that arrive in the order of the session. The new sample data includes image samples and text samples corresponding to the image samples. Based on the new sample data, obtain guidance features. The image sample is augmented to obtain several augmented samples, and the augmentation features of the augmented samples are extracted. The augmentation features correspond to the guidance features. The guidance features and several enhancement features are combined into a training set, and a continuous differential optimizer is constructed. The model parameters of the continuous differential optimizer are continuously optimized based on several training sets, so that the continuous differential optimizer has the ability to recognize several new samples.

2. The few-shot continuous learning method according to claim 1, characterized in that, The step of obtaining guiding features based on the newly added sample data includes: Obtain the image features corresponding to the image sample, and obtain the text features corresponding to the text sample; The image features and text features are combined into guidance features based on the condition generator.

3. The few-shot continuous learning method according to claim 1, characterized in that, The step of performing augmented sampling on the image samples to obtain a plurality of augmented samples includes: Extract the original latent space features of the image samples, and apply random perturbation to the original latent space features to obtain updated latent space features; The guiding features are used as the first conditional signal of the noise prediction network model, and the updated latent space features are used as the input values ​​of the noise prediction network model to obtain the predicted noise features. Enhanced samples are obtained based on the updated latent space features and the predicted noise features.

4. The few-shot continuous learning method according to claim 1, characterized in that, The step of continuously optimizing the model parameters of the continuous differential optimizer based on several training sets includes: The guiding features of the training set are used as the second condition signal of the continuous differential optimizer, and the enhancement features of the training set are used as the input value of the continuous differential optimizer, so as to obtain the dynamic law of the model parameters of the continuous differential optimizer. The model parameters are continuously updated based on the dynamic pattern to complete a single stage update and obtain the stage model parameters.

5. The few-shot continuous learning method according to claim 4, characterized in that, After the step of continuously updating the model parameters based on the dynamic law to complete a single stage update and obtaining the stage model parameters, the method further includes: The model parameters of the previous stage are used as the model parameters of the next training set to repeat the single stage update and complete continuous optimization to obtain the final model parameters.

6. The few-shot continuous learning method according to claim 4, characterized in that, The step of using the enhanced features of the training set as input values ​​for the continuous differential optimizer to obtain the dynamic laws of the model parameters of the continuous differential optimizer includes: Obtain the predicted loss for each of the enhanced features in the training set, and obtain the average loss based on the predicted losses; Backpropagation is performed on the average loss to obtain the average gradient; The dynamic patterns corresponding to the model parameters are obtained based on the average gradient.

7. The few-shot continuous learning method according to claim 6, characterized in that, The formula for obtaining the predicted loss is: ， in, This represents the prediction loss of the i-th augmented feature in the t-th training set at the m-th time point. This represents the i-th augmented feature in the t-th training set. Indicates guiding characteristics, This represents the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. This represents the true label corresponding to the t-th training set. This indicates that prediction is performed using a continuous differential optimizer. Represents the loss function; The expression for the dynamic law is: ， in, This represents the dynamic behavior of the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. Indicates the learning rate. Let represent the average gradient of the t-th training set at the m-th time point.

8. The few-shot continuous learning method according to claim 4, characterized in that, The formula for obtaining the stage model parameters is as follows: ， in, This indicates that the t-th training set is in the t-th position. The stage model parameters of the continuous differential optimizer at each time point This represents the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. This represents the dynamic behavior of the model parameters of the continuous differential optimizer corresponding to the t-th training set at the m-th time point. Indicates the time step.

9. A few-shot continuous learning system, applied to the few-shot continuous learning method as described in any one of claims 1 to 8, characterized in that, The system includes: The identification module is used to acquire several new sample data that arrive in the order of the session. The new sample data includes image samples and text samples corresponding to the image samples. Guidance features are acquired based on the new sample data. An extension module is used to perform augmented sampling on the image samples to obtain several augmented samples, and to extract the augmented features of the augmented samples, wherein the augmented features correspond to the guidance features; The execution module is used to combine the guidance features and several enhancement features into a training set, and construct a continuous differential optimizer. Based on the several training sets, the model parameters of the continuous differential optimizer are continuously optimized so that the continuous differential optimizer has the ability to recognize several new samples.

10. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the few-shot continuous learning method as described in any one of claims 1 to 8.

Images