CN118072099A - Class increment learning method based on joint distillation playback strategy - Google Patents

Abstract

The invention discloses a class increment learning method based on a joint distillation playback strategy, which comprises the following steps: 1) Training a distillation model of an initial task; 2) Setting a buffer area to store distilled data of an initial task; 3) Establishing a combined distillation model based on distillation data of an old task and original data of a new task; 4) The co-distilled data is played back during the next task training. According to the technical scheme, the distillation data of the old task and the original data of the new task are utilized to carry out combined distillation, and category characteristic information of the old task is fully considered, so that characteristic ambiguity is avoided. On the basis, the same learning rate is used in the distillation process of new and old tasks, and an optimal buffer area size is searched for storing distilled data for playback, so that catastrophic forgetting in a class increment learning scene is effectively relieved, and the classification effect of images is improved.

Description

Class-incremental learning method based on joint distillation and replay strategy

Technical Field

The present invention relates to the technical field of image classification, and in particular to a class incremental learning method based on a joint distillation replay strategy.

Background technique

In recent years, incremental learning has gradually become one of the research hotspots in the field of deep learning. There are three basic scenarios for incremental learning: task incremental learning, domain incremental learning, and class incremental learning. In the task incremental learning scenario, the model needs to gradually learn a set of clearly distinguishable tasks; in the domain incremental learning scenario, the model needs to learn the same problem in different environments; in the class incremental learning scenario, the model needs to distinguish all classes. Since real-world scenarios often generate new classes over time, class incremental learning is closer to real-world application scenarios and more challenging than task incremental learning and domain incremental learning. However, as new categories continue to increase, the model must be continuously updated to incorporate the knowledge of new tasks, and this learning process may destroy or even completely overwrite the knowledge of old tasks, leading to catastrophic forgetting problems. Therefore, it is particularly important to solve the catastrophic forgetting problem in incremental learning.

In current research, methods to mitigate catastrophic forgetting in incremental learning can be roughly divided into three categories: regularization-based methods, bias correction-based methods, and replay-based methods. Regularization-based methods impose constraints on important parameters in old tasks by adding regularization terms in the learning of new tasks, balancing new and old tasks and limiting the forgetting of old knowledge. However, in practical applications, it is difficult to design a reasonable metric to estimate the importance of each parameter in a neural network. The starting point of the bias correction-based method is to solve the bias problem in incremental learning and prevent the incremental learning network from being biased towards the category of the new task, but new biases may be introduced during the correction process.

Replay-based methods can be divided into two categories according to the different replay contents: feature replay and data replay. Feature replay-based incremental learning methods mitigate forgetting by saving and replaying old category sample features or old category prototype vectors in the deep feature space. Data replay-based incremental learning methods mitigate forgetting by saving and replaying examples of old tasks and generating data. With the development of dataset distillation technology, dataset distillation technology can be combined with incremental learning scenarios. The core idea of dataset distillation technology is to optimize the synthetic dataset, compress the knowledge of the old task training dataset into a small synthetic distilled dataset, and replay the synthetic distilled dataset of the old task when learning the new task. However, when the dataset distillation technology is applied to each task separately, the synthetic image does not take into account the category characteristics of the old task dataset, which may cause feature ambiguity, especially when the sample category characteristics of the new task are similar to those of the old task, which will aggravate the feature ambiguity and increase the forgetfulness of these categories.

Therefore, a new technology is needed to solve the above problems.

Summary of the invention

The purpose of the present invention is to provide a class incremental learning method based on a joint distillation replay strategy to address the deficiencies of the prior art. This method first adds the distilled data of the old task to the data distillation process of the original data set of the new task. This joint distillation process allows similar category features between the new and old task samples to be considered. Secondly, the same learning rate is used in the distillation process of the new and old tasks, and an optimal buffer size is found to store the distilled data. Finally, the joint distilled data stored in the buffer is replayed during the learning process of the next task to achieve class incremental learning.

The technical solution for achieving the purpose of the present invention is:

The incremental learning method based on the joint distillation replay strategy is different from the existing technology in that it includes the following steps:

1) Train the distillation model for the initial task: Given the training dataset for the initial task The distillation model parameter θ of the initial task, l( _xi , θ) is the loss function of the data _xi on the distillation model. The goal of distillation model optimization is to obtain a θ ^* through training so that the loss of the model on the entire data set is minimized, which is defined as follows:

Assume that the distillation model is randomly initialized with parameters θ ₀ , θ ₀ satisfies the distribution of p(θ ₀ ), and use standard stochastic gradient descent to optimize and update the distillation data. and learning rate/> Thereby updating the distillation model parameters of the initial task. Assuming that the k-th distillation model parameter update is now performed, then:

At this point, the optimization objective of the distillation model of the initial task becomes to minimize the loss function Thus, the distilled data of the initial task is obtained:

2) Set up a buffer to store the distillation data of the initial task: Initialize the buffer The distillation data of the initial task in step 1) Stored in a buffer, then:

3) Establish a joint distillation model based on the distilled data of the old task and the original data of the new task: Assuming that in a class incremental learning scenario, the neural network needs to continuously learn T tasks. In addition to a simple data set distillation for the initial task, each of the remaining tasks is jointly distilled, and the obtained joint distillation data is stored in the buffer in step 2). In order to enable the model to continue learning in the class incremental learning scenario, a fixed learning rate η needs to be used in the joint distillation process. The process is shown in the following formula:

in, and/> Respectively represent the distilled data of the previous task and the distilled data of the current task, /> Represents the original training dataset of the current task, /> Represents the loss of the distilled data of the previous task and the original training data set of the current task under the parameters of the distilled model;

4) Replay the joint distillation data during the next task training process: With the arrival of the next task data stream, in order not to forget the knowledge of the old task, the joint distillation data of the old task in step 3) is taken out from the buffer, and the joint distillation data and the original training data of the new task are input into the classification network together. The training process can be expressed as:

in, Indicates that each step s is from the training data set/> A batch of data selected from / > Represents the parameters of the classification network,/> It is a joint distilled dataset of old tasks taken from the buffer, including the distilled data of all categories that the classification network has learned.

The buffer storage process in step 2) is as follows: after the joint distillation process is completed, the joint distillation data of the old task is stored in the set buffer, and in order to enable the model to learn the feature information of the old category more fully, the buffer size is set to 100, that is, after the joint distillation, 100 distilled images are saved for each old category to reduce the forgetting of the old category. The distillation data stored in the buffer can be expressed as:

in, Represents the distilled data obtained after a simple data set distillation of the initial task. Represents the joint distilled data obtained by jointly distilling the distilled data of the old task and the original data of the new task.

The joint distillation process in step 3) is as follows: Initialize the distillation number of the current task and distillation model parameters, selecting a batch of data at each iteration of the optimizer/> Then use the distillation data/> Perform K=1, 2, …, k gradient descent updates on the joint distillation model parameters, then calculate the loss of the batch data _xq under the distillation model parameters and update the distillation data:

Where p(θ ₀ ) is the distribution of the initialization weights, the batch size of the batch data x _q is n, and the optimizer iterates Q times.

The joint distillation playback process in step 4) is as follows: when the classification network learns a new task, the joint distillation data of the old task is played back, that is, the training data set currently input to the classification network is Then the class incremental learning process based on the joint distillation replay strategy can be expressed as:

Therefore, after all tasks are learned, the classification network parameters are updated to The buffer is updated as />

This technical solution uses a joint distillation strategy to update the distillation data of the old task, which can avoid feature ambiguity in the distillation process; in the process of joint distillation, the new and old tasks are kept using the same learning rate, thereby ensuring the effectiveness of incremental learning.

This method no longer processes the distilled data of each learning task separately, but adds the distilled data of the old task to the data distillation process of the original data set of the new task. On this basis, an optimal buffer size is found to store the joint distilled data to maintain the memory of historical tasks. Finally, when a new task arrives, the joint distilled data of the old task is replayed. The proposed joint distillation replay strategy can effectively alleviate catastrophic forgetting and improve the classification performance in the incremental learning scenario.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a class incremental learning method based on a joint distillation replay strategy provided in an embodiment.

FIG2 is a schematic flow diagram of a combined distillation method provided in an embodiment.

Detailed ways

The content of the present invention is further described below in conjunction with the drawings and embodiments, but the present invention is not limited thereto.

Example:

Referring to Figure 1, the incremental learning method based on the joint distillation replay strategy is different from the existing technology in that it includes the following steps:

1) Train the distillation model for the initial task: Given the training dataset for the initial task The distillation model parameter θ of the initial task, l( _xi , θ) is the loss function of the data _xi on the distillation model. The goal of distillation model optimization is to obtain a θ ^* through training so that the loss of the model on the entire data set is minimized, which is defined as follows:

Assume that the distillation model is randomly initialized with parameters θ ₀ , θ ₀ satisfies the distribution of p(θ ₀ ), and use standard stochastic gradient descent to optimize and update the distillation data. and learning rate/> Thus, the distillation model parameters of the initial task are updated. Assuming that the k-th distillation model parameter update is now performed, then:

At this point, the optimization objective of the distillation model of the initial task becomes to minimize the loss function Thus, the distilled data of the initial task is obtained:

2) Set up a buffer to store the distillation data of the initial task: Initialize the buffer The distillation data of the initial task in step 1) Stored in a buffer, then:

3) Establish a joint distillation model based on the distilled data of the old task and the original data of the new task: Referring to Figure 2, assuming that in a class incremental learning scenario, the neural network needs to continuously learn T tasks. In addition to a simple data set distillation for the initial task, each of the remaining tasks t∈{2,...,T} is jointly distilled, and the obtained joint distillation data is stored in the buffer in step 2). In order to enable the model to continue learning in a class incremental learning scenario, a fixed learning rate η needs to be used in the joint distillation process. The process is shown in the following formula:

in, and/> Respectively represent the distilled data of the previous task and the distilled data of the current task, /> Represents the original training dataset of the current task, /> Represents the loss of the distilled data of the previous task and the original training data set of the current task under the parameters of the distilled model;

4) Replay the joint distillation data during the next task training process: With the arrival of the next task data stream, in order not to forget the knowledge of the old task, the joint distillation data of the old task in step 3) is taken out from the buffer, and the joint distillation data and the original training data of the new task are input into the classification network together. The training process can be expressed as:

in, Indicates that each step s is from the training data set/> A batch of data selected from / > Represents the parameters of the classification network,/> It is a joint distilled dataset of old tasks taken from the buffer, including the distilled data of all categories that the classification network has learned.

The buffer storage process in step 2) is as follows: after the joint distillation process is completed, the joint distillation data of the old task is stored in the set buffer, and in order to enable the model to learn the feature information of the old category more fully, the buffer size is set to 100, that is, after the joint distillation, 100 distilled images are saved for each old category to reduce the forgetting of the old category. The distillation data stored in the buffer can be expressed as:

in, Represents the distilled data obtained after a simple data set distillation of the initial task. Represents the joint distilled data obtained by jointly distilling the distilled data of the old task and the original data of the new task.

Referring to FIG. 2 , the joint distillation process in step 3) is specifically as follows: Initialize the distillation number of the current task and distillation model parameters, selecting a batch of data at each iteration of the optimizer/> Then use the distillation data/> Perform K=1, 2, …, k gradient descent updates on the joint distillation model parameters, then calculate the loss of the batch data _xq under the distillation model parameters and update the distillation data:

Where p(θ ₀ ) is the distribution of the initialization weights, the batch size of the batch data x _q is n, and the optimizer iterates Q times.

The joint distillation playback process in step 4) is as follows: when the classification network learns a new task, the joint distillation data of the old task is played back, that is, the training data set currently input to the classification network is Then the class incremental learning process based on the joint distillation replay strategy can be expressed as:

Therefore, after all tasks are learned, the classification network parameters are updated to The buffer is updated as />

Claims (4)

1. A class incremental learning method based on a joint distillation replay strategy, characterized by comprising the following steps: 1) Train the distillation model for the initial task: Given the training dataset for the initial task The distillation model parameter θ of the initial task, l( _xi , θ) is the loss function of the data _xi on the distillation model. The goal of distillation model optimization is to obtain a θ ^* through training so that the loss of the model on the entire data set is minimized, which is defined as follows: Assume that the distillation model is randomly initialized with parameters θ ₀ , θ ₀ satisfies the distribution of p(θ ₀ ), and use standard stochastic gradient descent to optimize and update the distillation data. and learning rate/> Thereby updating the distillation model parameters of the initial task. Assuming that the k-th distillation model parameter update is now performed, then: At this point, the optimization objective of the distillation model of the initial task becomes to minimize the loss function Thus, the distilled data of the initial task is obtained: 2) Set up a buffer to store the distillation data of the initial task: Initialize the buffer The distillation data of the initial task in step 1) Stored in a buffer, then: 3) Establish a joint distillation model based on the distilled data of the old task and the original data of the new task: Assuming that in a class incremental learning scenario, the neural network needs to continuously learn T tasks. In addition to a simple data set distillation for the initial task, each of the remaining tasks t∈{2,...,T} is jointly distilled, and the obtained joint distillation data is stored in the buffer in step 2). In order to enable the model to continue learning in a class incremental learning scenario, a fixed learning rate η needs to be used in the joint distillation process. The process is shown in the following formula: in, and/> Respectively represent the distilled data of the previous task and the distilled data of the current task, /> Represents the original training dataset of the current task, /> Represents the loss of the distilled data of the previous task and the original training data set of the current task under the parameters of the distilled model; 4) Replay the joint distillation data during the next task training process: With the arrival of the next task data stream, in order not to forget the knowledge of the old task, the joint distillation data of the old task in step 3) is taken out from the buffer, and the joint distillation data and the original training data of the new task are input into the classification network together. The training process can be expressed as: in, Indicates that each step s is from the training data set/> A batch of data selected from / > Represents the parameters of the classification network,/> It is a joint distilled dataset of old tasks taken from the buffer, including the distilled data of all categories that the classification network has learned. 2. The class incremental learning method based on the joint distillation playback strategy according to claim 1 is characterized in that the buffer storage process in step 2) is: after the joint distillation process is completed, the joint distillation data of the old task is stored in the set buffer, and in order to enable the model to learn the feature information of the old category more fully, the buffer size is set to 100, that is, after the joint distillation, 100 distilled images are saved for each old class to reduce the forgetting of the old class. The distillation data stored in the buffer can be expressed as: in, Represents the distilled data obtained after a simple data set distillation of the initial task, /> Represents the joint distilled data obtained by jointly distilling the distilled data of the old task and the original data of the new task. 3. The incremental learning method based on the joint distillation playback strategy according to claim 1 is characterized in that the joint distillation process in step 3) is specifically: initializing the distillation data of the current task and distillation model parameters, selecting a batch of data at each iteration of the optimizer/> Then use the distillation data/> Perform K=1, 2, …, k gradient descent updates on the joint distillation model parameters, then calculate the loss of the batch data _xq under the distillation model parameters and update the distillation data: Where p(θ ₀ ) is the distribution of the initialization weights, the batch size of the batch data x _q is n, and the optimizer iterates Q times. 4. The incremental learning method based on the joint distillation replay strategy according to claim 1 is characterized in that the joint distillation replay process in step 4) is: when the classification network learns a new task, the joint distillation data of the old task is replayed, that is, the training data set currently input to the classification network is Then the class incremental learning process based on the joint distillation replay strategy can be expressed as: Therefore, after all tasks are learned, the classification network parameters are updated to The buffer is updated as />