Class incremental learning efficient playback method based on pre-training model

By using efficient playback method based on pretrained models in class incremental learning, combining the high performance of the large model and the industrial friendliness of the playback idea, the problem of performance maintenance and cost in class incremental learning is solved, and the problem of high performance in class incremental learning is achieved. Efficient learning training with low computational cost and low historical sample retention.

CN118964979APending Publication Date: 2024-11-15UNIV OF SCI & TECH OF CHINA

Patent Information

Application Number
CN202410979485.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to maintain the performance of old tasks in class incremental learning. At the same time, due to the large size of the pre-trained model, the training cost of traditional playback methods is high and cannot be directly applied to standard deep learning industrial practices.

Method used

The class incremental learning efficient playback method based on pretrained models is adopted. By acquiring the pretrained model, class incremental learning is performed on the target data set, and the approximate playback algorithm is used to achieve the low computational cost and low historical sample retention conditions. Classification prediction model training without catastrophic forgetting.

Benefits of technology

Under the conditions of pre-trained models, efficient training speed and performance are achieved, the historical sample demand is reduced, and the performance loss and high computational cost in traditional methods are avoided, which is suitable for industrial implementation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964979A_ABST
    Figure CN118964979A_ABST
Patent Text Reader

Abstract

The invention provides a class incremental learning efficient playback method based on a pre-training model, and belongs to the technical field of artificial intelligence, and the method comprises the steps: obtaining a data pre-training model; based on the obtained data pre-training model, performing class incremental learning on a target data set; and predicting a historical learning task according to the data pre-training model and a class incremental learning result. According to the method, the pre-training model and the replay thought are combined, under the condition of the pre-training model, training can be greatly accelerated while performance is not lost, and only low-limit historical sample information is needed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an efficient replay method for incremental learning based on a pre-training model. Background Art

[0002] Class incremental learning is an important form of incremental learning, which focuses on the model's learning ability when faced with data that arrives at different times and belongs to the same task but different categories. In this scenario, the model needs to keep the memory of old categories while learning new categories to avoid catastrophic forgetting.

[0003] At present, in the existing deep learning training framework, under the optimization algorithm with back-propagation algorithm, gradient descent and derivative algorithm as the core, if we want to achieve quasi-incremental learning of image, text, audio, video and other data, there are mainly two ideas.

[0004] One idea is sample replay. The idea of ​​sample replay is generally believed to be derived from the theory of complementary learning systems in neuroscience, and replay memory is considered to be one of the most effective ways to learn new tasks without forgetting old knowledge. Sample replay can be divided into direct replay type methods and indirect replay type methods. Direct replay aims to enable a part of historical samples to directly participate in the training of new tasks to maintain the performance of the model on old tasks. Indirect replay aims to reduce the volume of historical samples involved in replay while maintaining performance. Taking the most mainstream replay based on knowledge distillation as an example, the latest model is copied and retained as the teacher model. While the main model is training new tasks, a part of the historical samples are used for knowledge distillation. On this basis, different expansion methods can be used. For example, the model correction method will standardize the weights of the classification layer, and the dynamic network method will expand multiple network structures.

[0005] Another idea is the pre-training model developed in recent years. Thanks to the ultra-high model generalization performance generated by large transformer models combined with large data sets, continuous learning methods based on pre-training models are proposed, such as L2P (Learning to Prompt for Continuous Learning), which uses large model prompt learning technology in the natural language field.

[0006] However, traditional algorithms that use the idea of ​​sample replay cannot avoid the problem that they require a certain scale of historical samples to achieve better results, and it is difficult to keep up with the performance of algorithms based on pre-trained large models. In order to reduce the cost of historical samples, the training cycle and computing resources of the algorithm design are relatively large. Methods based on pre-trained models require additional training processes or components and cannot be directly applied in current standard deep learning industrial practices. They are not hardware-friendly, and because pre-trained models are usually larger in size, the training cost is unaffordable if ordinary replay methods are used. Summary of the invention

[0007] In view of the above problems, in order to overcome or alleviate the shortcomings of the prior art, the present invention provides an efficient replay method for class incremental learning based on a pre-trained model, which fully considers the advantages of replay and pre-trained models in continuous learning, combines the high performance of large models with the industrial friendliness of the replay idea, and achieves the purpose of efficient class incremental learning and training with low computing cost and low historical sample retention.

[0008] On the one hand, the present invention provides an efficient replay method for incremental learning based on a pre-trained model, comprising:

[0009] Get data pre-training model;

[0010] Based on the acquired data pre-training model, performing class incremental learning on the target data set;

[0011] According to the data pre-training model and the class incremental learning results, historical learning tasks are predicted.

[0012] As a further improvement of the present invention, the class incremental learning of the target data set based on the acquired data pre-training model includes using the acquired data pre-training model as an encoder, performing class incremental learning on the target data set on the original decoder, and obtaining a classification prediction model without catastrophic forgetting of historical tasks.

[0013] As a further improvement of the present invention, predicting historical learning tasks based on the data pre-training model and class incremental learning results includes using the classification prediction model as a decoder, and jointly predicting historical learning tasks with the encoder and the trained decoder.

[0014] As a further improvement of the present invention, the data pre-training model is a model pre-trained using a pre-training data set or a third-party pre-training model.

[0015] As a further improvement of the present invention, the data pre-training model is obtained by pre-training with a pre-training data set, including training a transformer model with the pre-training data set.

[0016] As a further improvement of the present invention, the performing class incremental learning on the target data set includes performing class incremental learning on the target data set using an approximate replay algorithm on the original decoder.

[0017] As a further improvement of the present invention, the step of performing class incremental learning on the target data set using an approximate replay algorithm on the original decoder includes:

[0018] Initialize the decoder y=d(r) parameters and freeze the encoder r=e(x) parameters;

[0019] Initialize the memory space of task b, receive the data stream of task b, and prepare to execute the incremental learning of the b-th phase;

[0020] For main training, the data stream sample x is forward inferred through the encoder e to obtain the encoded feature r, which is input into the decoder d to obtain the classification prediction y, and the cross entropy loss is calculated Cross Entropy(y, label), and back propagation is completed to update the decoder parameters;

[0021] Perform auxiliary training and retrieve the memory space information from task 1 to task b-1 {m 1 ,…,m b-1}, input the memory space information into the decoder d to obtain the classification prediction y, perform cross entropy loss calculation Cross Entropy(y,[1,2,…,b-1]), and complete back propagation to update the decoder parameters;

[0022] Update the memory space information of task b, by {m 1 ,…,m b-1} becomes {m 1 ,…,m b};

[0023] Return to perform incremental learning of class b+1 stage.

[0024] As a further improvement of the present invention, in the main training, if the main training gradient constraint α is set, the gradient clipping coefficient of this round of main training is set to α; in the auxiliary training, if the auxiliary training gradient constraint β is set, the gradient clipping coefficient of this round of auxiliary training is set to β.

[0025] As a further improvement of the present invention, the memory space m of task b b The calculation formula is:

[0026]

[0027] Where N b is the number of training instances, r i is the encoding feature and Ω is the importance function.

[0028] As a further improvement of the present invention, the auxiliary training gradient constraint β of task b is adjusted to change the performance of task b.

[0029] As a further improvement of the present invention, the main training gradient constraint α is 0.5.

[0030] As a further improvement of the present invention, for the class incremental learning in stage b, there are:

[0031]

[0032] Among them, w o represents the weight parameter after standard training on the datasets from task 1 to task b, w b represents the weight of the model after completing the b-1 stage training on the dataset of task b, Δw b Indicates that under normal circumstances, other task data participate in the gradient update generated by standard training.

[0033] As a further improvement of the present invention,

[0034]

[0035] represents the back-propagation update on w for all samples from task 1 to task b-1.

[0036] On the other hand, the present invention provides an efficient replay method for incremental learning based on a pre-trained model, comprising:

[0037] Get the image data pre-training model;

[0038] Based on the acquired image data pre-training model, performing class incremental learning on the target image data set;

[0039] According to the image data pre-training model and the class incremental learning results, historical learning tasks are predicted.

[0040] The present invention improves the existing incremental learning method by providing an efficient replay method for incremental learning based on a pre-trained model, combining the pre-trained model with the replay idea. Under the condition of the pre-trained model, the training can be greatly accelerated without losing performance, and only a minimum amount of historical sample information is required. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the steps of an efficient replay method for incremental learning based on a pre-trained model in an embodiment of the present invention.

[0042] Figure 2 It is a schematic diagram of the implementation process of the efficient replay method of incremental learning based on a pre-trained model according to an embodiment of the present invention.

[0043] Figure 3 It is a structural diagram of a pre-trained model-based incremental learning efficient replay device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following is combined with specific embodiments and appendix Figure 1-3 The invention is described in detail so that those skilled in the art can more fully understand the purpose, features and effects of the invention.

[0045] Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as commonly understood by those skilled in the art to which the present invention belongs. When the definition of a term in the present invention conflicts with the meaning commonly understood by those skilled in the art to which the present invention belongs, the definition described in the present invention shall prevail.

[0046] The present invention is based on the pre-trained model of the quasi-incremental learning efficient replay method, which aims to solve the problem of deep neural network model training optimization in the quasi-incremental learning scenario in the continuous learning problem. The continuous learning problem originates from the catastrophic forgetting phenomenon in deep learning training: directly training new tasks on the model will significantly affect the performance of old tasks. This requires people to collect a large amount of training data in advance, which undoubtedly limits the flexibility of artificial intelligence in the real world. On the one hand, due to security and privacy issues, the original data may not be preserved for a long time. On the other hand, real-world data is usually generated dynamically, and people may not always be able to collect all data in advance. Continuous learning technology here refers to the use of a certain method to enable model training to be carried out in steps, that is, first use a part of the data for training, and then use another part of the data for training, and try to maintain the performance on all tasks. The quasi-incremental learning scenario is one of the situations that is closer to the real world, and aims to continuously build a complete classifier network based on a continuous data stream. It has been found that the high generalization of large models can greatly help continuous learning, but it usually requires algorithm design outside the standard training process, and due to the large volume of the pre-trained model, ordinary replay methods cannot be used.

[0047] The present invention improves the training speed and performance and reduces the demand for historical samples by providing an efficient replay method for class incremental learning based on a pre-trained model.

[0048] Embodiment 1

[0049] As a specific embodiment of the present invention, this embodiment provides an efficient replay method for class incremental learning based on a pre-trained model, referring to Figure 1 , Figure 2 , the specific steps are as follows:

[0050] S100, obtaining a data pre-training model;

[0051] S200, performing class incremental learning on the target data set based on the acquired data pre-training model;

[0052] S300: predict historical learning tasks based on the data pre-training model and class incremental learning results.

[0053] The data includes image data, text data, audio data, video data and other classifiable data, and the target data set includes a target image data set, a target text data set, a target audio data set and a target video data set. The efficient replay method for class incremental learning based on a pre-trained model of this embodiment can be used for class incremental learning of image data, and can also be used for class incremental learning of text data, audio data, video data and other classifiable data.

[0054] Specifically,

[0055] In S200, the data pre-training model obtained in S100 is used as an encoder, the encoder can encode the original data into appropriate intermediate features, and the target data set is subjected to class incremental learning on the original decoder to train a classification prediction model without catastrophic forgetting of historical tasks, and the classification prediction model is used as a decoder. The class incremental learning adopts an approximate replay algorithm to approximate all samples under the same category of tasks as a unit of information for replay.

[0056] The traditional replay type algorithm aims to incrementally learn a classification network from scratch, and its characteristic is to adapt to standard training. However, no matter what kind of variant algorithm, it is limited by the size of the historical sample set, that is, too small historical sample information retention directly limits the performance upper limit of the algorithm, and too large historical sample information retention is not much different from standard training. The class incremental learning algorithm based on the pre-trained model breaks through this limitation. Thanks to the advance input of a large amount of data, the high generalization of the pre-trained large model enables the class incremental learning to obtain very high algorithm performance with very small historical sample information. However, the volume of the pre-trained large model determines that if the traditional replay method is used, the computational cost will be very huge. Therefore, the existing class incremental learning algorithm based on the pre-trained model usually uses a non-standard training process, which increases the complexity of training. This embodiment circumvents the shortcomings of both by using approximate replay on the pre-trained model, while obtaining the high algorithm performance brought by the pre-trained large model and the simplicity brought by the replay method.

[0057] In S300, the encoder and the decoder trained by class incremental learning are jointly used to predict the historical learning task.

[0058] Furthermore, in a possible embodiment, S100 includes collecting a pre-training data set, performing pre-training using the original model using the pre-training data set, removing the classification layer and other customized forms after the training is completed, and using the trained model as a data pre-training model. The original model may be a CNN model or a transformer model. Preferably, the present implementation uses a large transformer model as the original model, which has a higher training speed and performance than a CNN model. The pre-training data set may be the same as or different from the target data set.

[0059] In another possible embodiment, in S100, the data pre-training model uses a pre-training model provided by a third party. Since the cost of training a large transformer model is getting higher and higher, the pre-training model provided by a third party is directly used for the training of the decoder in S200, thereby saving training costs.

[0060] Furthermore, S200 includes:

[0061] S201, initialize the decoder parameters and freeze the encoder parameters, wherein the decoder parameters are y=d(r) parameters and the encoder parameters are r=e(x) parameters. The encoder is the data pre-training model obtained in S100.

[0062] S202, initialize the memory space of task b, receive the data stream of task b, and prepare to execute the class incremental learning of the bth stage.

[0063] Specifically, in the incremental learning of this embodiment, there is a continuous data stream consisting of B data sets {D 1 ,D 2 ,…,D B}, and the categories of these B data sets do not overlap. b ={(x i b ,y i b )} Nb It represents the incremental learning stage of the b-th class, with N b training examples. i b Belongs to category Y b The sample of Y b is the label space of task b. For any b≠b ’ , there is Y b ∩Y b’ When training task b, the model is only allowed to fully access D b After each task is trained, the model will be able to perform the task Y 1 ∪…Yb Evaluate performance on.

[0064] S203, perform main training, forward inference of data stream sample x through encoder e, obtain encoding feature r, input encoding feature r into decoder d to obtain classification prediction y, perform cross entropy loss calculation Cross Entropy(y, label), and complete back propagation to update decoder parameters. If the main training gradient constraint α is set, set the main training gradient clipping coefficient to α, where α is preferably 0.5.

[0065] S204, perform auxiliary training, retrieve the memory space information of task 1 to task b-1 {m 1 ,…,m b-1}, input the memory space information into the decoder d to obtain the classification prediction y, perform cross entropy loss calculation Cross Entropy(y,[1,2,…,b-1]), and complete back propagation to update the decoder parameters. If the auxiliary training gradient constraint β is set, set the auxiliary training gradient clipping coefficient to β. The memory space information belongs to the historical task memory.

[0066] By setting the main training gradient constraint α and the auxiliary training gradient constraint β, the steady state of the new task and the old task in the parameter space can be maintained. Appropriate settings of α and β can produce better incremental learning training effects and can be adjusted according to the training situation. In this embodiment, a main training gradient constraint experience value of 0.5 is provided. In other possible embodiments, the auxiliary training constraint β may not be set.

[0067] Combination Figure 2 , the auxiliary training data is approximately compressed, corresponding to Figure 2 The approximate replay step in the old task, while the main training is mainly for Figure 2 In the step of training the decoder on the new task, in this embodiment, it is equivalent to performing two class standard trainings on a decoder, and the two trainings are affected by α and β respectively.

[0068] In this embodiment, the main training adopts the standard training process, and the auxiliary training adopts the approximate replay. The approximate replay is further described below.

[0069] According to the back propagation formula, the training process is reflected in the parameters as the adjustment of the parameters by gradient update. Specifically, the weight of each layer of the network is updated during the training process as follows:

[0070]

[0071] In the formula, δ is the loss feedback of the next layer of the network, a is the activation input of the previous layer of the network, w is the weight parameter, and the weight plays a major role in class incremental learning. C is the final loss, and K is the cumulative count upper bound of the back propagation update process of the l-layer weight. is the value of the jth component of the return loss vector received by layer l, is the j-th row and k-th column of the weight parameter matrix of layer l.

[0072] Therefore, for the incremental learning of class b, we have:

[0073]

[0074] In the formula, f is the training model, w o represents the weight parameter after standard training on the datasets from task 1 to task b, w b represents the weight of the model after completing the b-1 stage training on the dataset of task b, Δw b In normal cases, other task data participate in the gradient update generated by standard training, which is expanded as follows:

[0075]

[0076] represents the back propagation update of all samples from task 1 to task b-1 on w. Considering the additivity of gradients in the back propagation update process, training task b first and then replaying samples of other tasks is consistent with the dataset that trains tasks 1 to b at the same time at the beginning. Therefore, considering Δw b Take approximate measures:

[0077]

[0078] All sample activation inputs under the same category are weighted sampled, where Ω is the importance function used to evaluate the importance of the current sample.

[0079] a * Refers to the approximate expression of all sample information under this type of task. Assuming that there are 10,000 samples in this category, normal replay should select at least 500 samples, that is, 5%. The present invention believes that in the case of pre-training, replay a * To replace this 5%, if we take the above example, according to conservative estimates, training can be up to 10 times faster than traditional replay methods and 2-3 times faster than existing methods based on pre-trained models.

[0080] Furthermore, in a possible embodiment, the replay of all samples from task 1 to task b-1 can be approximated as:

[0081]

[0082] Based on this, for any number of samples in any category, it is expressed as a unit of information participating in the replay training process. Therefore, the replay training of class incremental learning of weight w can be expressed as follows:

[0083]

[0084] The incremental learning process in phase b is divided into two parts: learning a new task b and consolidating old tasks 1 to b-1. Both parts can use a hyperparameter to balance the relationship between them. In this embodiment, the hyperparameters are α and β respectively.

[0085] S205, update the memory space information of task b, by {m 1 ,…,m b-1} becomes {m 1 ,…,m b}

[0086] in,

[0087]

[0088] In the formula, m b is the memory space information of task b, N b is the number of training instances, r i is the encoding feature and Ω is the importance function.

[0089] The symbols are replaced based on the scene description, r is the output of the encoder and the input of the decoder, that is, a of the decoder.

[0090] Pass S205, complete Figure 2 Update the historical task memory step in .

[0091] S206, return to S202 to perform incremental learning of class b+1 stage

[0092] Further, S300 includes performing prediction on tasks 1 to b to obtain prediction results about tasks 1 to b.

[0093] In a possible embodiment, the auxiliary training gradient constraint β of task b is adjusted to change the performance of task b. Specifically, if the performance of task b is enhanced, the β value is increased for the auxiliary training of task b. Conversely, if the performance of other tasks is improved by forgetting task b, the β value can be reduced or the memory space information of task b can be removed.

[0094] It should be noted that for this embodiment, there is no difference in content between the target data set and the target future data stream. The former is generally regarded as an abstract expression of the target task, and the latter is regarded as a description of the actual scene, with a steady stream of new information generated. In training, a pre-set data set is usually used, and in order to simulate the real-world incremental learning scenario, it is divided into many sub-data sets; the difference is that the data itself is continuously generated in the real world, so there is no need to simulate it.

[0095] Table 1 shows the performance comparison between the efficient replay method of incremental learning based on the pre-trained model of this embodiment and different traditional replay methods. The traditional replay methods used include RMM (Reinforced Memory Management), GEM (Gradient Episodic Memory), iCaRL (Incremental Classifier and Representation Learning), and WA (Weight Alignment).

[0096] Table 1 Performance comparison between different methods

[0097]

[0098] Among them, Base-n, Inc-m means that for a data set, standard training starts from task n, and m tasks are incrementally learned in each stage. It indicates the top 1 accuracy of the model on all historical tasks after the last stage of learning. Represents each incremental learning stage The training cycle is represented by how many rounds of full training of the current dataset are required for each incremental learning stage, and the sample set is represented by the size of the historical sample information counted in the original sample unit.

[0099] Table 1 shows the top1 accuracy of the last stage. Usually, in the process of incremental learning, this value continues to decrease. Therefore, the top1 accuracy of the last stage must be lower than that of each incremental learning stage. The average accuracy of .

[0100] As can be seen from Table 1, the top 1 accuracy and average accuracy of this embodiment are better than those of the traditional replay method. The performance of the method of this embodiment is significantly higher than that of the traditional replay method, while the training speed is faster and less historical sample information is required.

[0101] Compared with the standard training process, the pre-trained model-based incremental learning efficient replay method of the present invention has only one more gradient constraint adjustment, which is almost equivalent to the standard training process and is hardware-friendly. Under the condition of the pre-trained model, the training can be greatly accelerated without losing performance, and only a minimum amount of historical sample information is required.

[0102] The efficient replay method for incremental learning based on a pre-trained model of the present invention combines the high performance of a large model with the industrial friendliness of the replay idea, and realizes incremental learning training with low computational cost and low historical sample retention, which has the advantages of both while avoiding their respective disadvantages.

[0103] Embodiment 2

[0104] As another specific embodiment of the present invention, the present invention provides a class incremental learning efficient replay device based on a pre-training model, referring to Figure 3 , including at least one memory, at least one processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the efficient replay method of class incremental learning based on a pre-trained model in the embodiment.

[0105] The device implemented in this embodiment achieves an efficient replay method for incremental learning based on a pre-trained model, and uses replay on the basis of the pre-trained model, thereby achieving the advantages of high algorithm performance brought by the pre-trained large model and simplicity brought by the replay method, and solves the problems that the traditional replay algorithm is not effective enough when used alone and the pre-trained model is not suitable for standard deep learning training.

[0106] Embodiment 3

[0107] As another specific embodiment of the present invention, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the efficient replay method of incremental learning based on a pre-trained model in the embodiment are implemented.

[0108] The computer-readable storage medium of this embodiment stores a program that can implement the steps of an efficient replay method for class incremental learning based on a pre-trained model, thereby combining the advantages of a pre-trained model and replay to complete class incremental learning.

[0109] Embodiment 4

[0110] As another specific embodiment of the present invention, the present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the efficient replay method of class incremental learning based on a pre-trained model in the embodiment.

[0111] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. An efficient replay method for incremental learning based on a pre-trained model, characterized in that: The method comprises: Get data pre-training model; Based on the acquired data pre-training model, performing class incremental learning on the target data set; According to the data pre-training model and the class incremental learning results, historical learning tasks are predicted.

2. The efficient replay method for incremental learning based on a pre-trained model according to claim 1, characterized in that: The performing class incremental learning on the target data set based on the acquired data pre-training model includes using the acquired data pre-training model as an encoder, performing class incremental learning on the target data set on the original decoder, and obtaining a classification prediction model without catastrophic forgetting of historical tasks.

3. The efficient replay method for class incremental learning based on a pre-training model according to claim 2 is characterized in that: The predicting of historical learning tasks based on the data pre-training model and the class incremental learning results includes using the classification prediction model as a decoder and jointly predicting historical learning tasks with the encoder and the decoder obtained through training.

4. The efficient replay method for class incremental learning based on a pre-training model according to claim 2 is characterized in that: The data pre-training model is a model pre-trained using a pre-training data set or a third-party pre-training model.

5. The efficient replay method for incremental learning based on a pre-trained model according to claim 2, characterized in that: The performing class incremental learning on the target data set includes performing class incremental learning on the target data set using an approximate replay algorithm on the original decoder.

6. The efficient replay method for incremental learning based on a pre-trained model according to claim 5, characterized in that: The step of using an approximate replay algorithm on the original decoder to perform incremental learning on the target data set includes: Initialize the decoder y=d(r) parameters and freeze the encoder r=e(x) parameters; Initialize the memory space of task b, receive the data stream of task b, and prepare to execute the incremental learning of the b-th phase; For main training, the data stream sample x is forward inferred through the encoder e to obtain the encoded feature r, which is input into the decoder d to obtain the classification prediction y, and the cross entropy loss is calculated Cross Entropy(y, label), and back propagation is completed to update the decoder parameters; Perform auxiliary training and retrieve the memory space information {m1,…,m b-1 }, input the memory space information into the decoder d to obtain the classification prediction y, perform cross entropy loss calculation Cross Entropy(y,[1,2,…,b-1]), and complete back propagation to update the decoder parameters; Update the memory space information of task b, from {m1,…,m b-1 } becomes {m1,…,m b }; Return to perform incremental learning of class b+1.

7. The efficient replay method for class incremental learning based on a pre-training model according to claim 6 is characterized in that: The memory space m of task b b The calculation formula is: Where N b is the number of training instances, r i is the encoding feature and Ω is the importance function.

8. The efficient replay method for incremental learning based on a pre-trained model according to claim 6, characterized in that: In the main training, if the main training gradient constraint α is set, the gradient clipping coefficient of this round of training is set to α; In auxiliary training, if the auxiliary training gradient constraint β is set, the gradient clipping coefficient of this round of training is set to β.

9. The efficient replay method for incremental learning based on a pre-trained model according to claim 8, characterized in that: The main training gradient constraint α is 0.

5.

10. An efficient replay method for incremental learning based on a pre-trained model, characterized in that: The method comprises: Get the image data pre-training model; Based on the acquired image data pre-training model, performing class incremental learning on the target image data set; According to the image data pre-training model and the class incremental learning results, historical learning tasks are predicted.

Citation Information

Patent Citations

  • Meta-learning-based domain increment method

    CN112308211A

  • Incremental learning method for generating playback based on variational auto-encoder

    CN116797878A

  • Incremental learning method based on meta learning and playback

    CN118035863A

Cited By

  • Electrical equipment fault diagnosis method based on category increment updating strategy and memory management optimization

    CN119337237A

  • Image classification method of class incremental learning based on kernel sparsity and entropy

    CN121686094A