A data forgetting method based on error maximization noise and neuron pruning

Abstract

The present invention relates to the field of data forgetting, and in particular to a data forgetting method based on error maximization noise and neuron pruning, comprising: generating an error maximization noise matrix; initializing model parameters according to the connection sensitivity of each model parameter to induce model forgetting; updating a target forgotten data set and a retained data set; periodically detecting the performance change amplitude, and determining whether to perform adaptive adjustment for a data forgetting process according to the performance change amplitude; if adaptive adjustment is performed for data forgetting process parameters, periodically iteratively updating the noise intensity and pruning scale according to the performance change amplitude; and stopping iterative updating for the forgetting process parameters under the condition that the forgetting effect is met or a preset number of iterations is reached. The present invention retains the overall performance of the model to the greatest extent while ensuring the data forgetting effect.

Description

Technical Field

[0001] The present invention relates to the field of data forgetting, and in particular to a data forgetting method based on error maximization noise and neuron pruning. Background Art

[0002] In the field of personal sensitive information, in order to solve the data privacy issues of large-scale data sets used in artificial intelligence and machine learning, data forgetting methods are used to ensure that sensitive data is not used in subsequent processes to ensure the security of sensitive data. The existing data forgetting methods have the following problems: the error maximization noise method may cause unnecessary errors in the model when processing other data, affecting the overall performance, and the neuron pruning method often cannot completely eliminate the data impact due to the locality of pruning, and may cause a significant decrease in model performance. Therefore, how to achieve complete data forgetting without significantly affecting the model performance is an urgent problem to be solved by technicians in this field. Summary of the invention

[0003] To this end, the present invention provides a data forgetting method based on error maximization noise and neuron pruning to overcome the problem in the prior art that it is unable to minimize the impact on the performance of the updated model while ensuring complete data forgetting, resulting in low efficiency in subsequent data processing of the model.

[0004] To achieve the above object, the present invention provides a data forgetting method based on error maximization noise and neuron pruning, comprising:

[0005] Generate an error-maximizing noise matrix;

[0006] The connection sensitivity of each model parameter is detected, and the model parameters are initialized according to the connection sensitivity of each model parameter to induce model forgetting;

[0007] Update the target forgotten dataset and the retained dataset;

[0008] Periodically detect the performance change range, and determine whether to perform adaptive adjustments to the data forgetting process based on the performance change range;

[0009] If the forgetting process parameters are adaptively adjusted, the noise intensity and pruning scale are iteratively updated periodically according to the performance change amplitude;

[0010] When the forgetting effect is met or the preset number of iterations is reached, the iterative update of the forgetting process parameters is stopped.

[0011] Furthermore, the process of generating the error maximization noise matrix includes:

[0012] Freeze the weights of the pre-trained model and randomly initialize a noise matrix from a standard normal distribution;

[0013] Obtain the error maximization noise matrix according to the optimization function;

[0014] The optimization function is formula (1),

[0015] (1),

[0016] in, To maximize the noise for the error, is the classification loss corresponding to the target forgotten category, is the currently trained model, is a noise matrix of the same size as the input to the model, For The strength of regularization, are the pre-trained model parameters, Forget the labels of each category of data in the target dataset.

[0017] Further, the connection sensitivity of each model parameter is detected, and the model parameters that need to be initialized are determined according to the connection sensitivity of each model parameter;

[0018] If the parameters are initialized, the target forgetting model is induced to forget the target forgetting data according to the initialized model parameters.

[0019] Furthermore, the connection sensitivity of each model parameter is determined according to formula (2):

[0020] (2),

[0021] in, For the The connection sensitivity of the model parameters, For a given data set, the function Used to determine the loss of the model, For the model parameters, is the input of the model.

[0022] Furthermore, the target forgotten dataset and the retained dataset are updated periodically, and the updating process is performed according to formula (3):

[0023] (3),

[0024] in, To retain the dataset, For the target forget dataset, To obtain the best model parameters, For the The model parameters obtained by the cycle update are derived. For the The model parameters are updated in the next cycle.

[0025] Furthermore, before executing the data forgetting operation, the initial performance index of the model is recorded, and during the execution of the data forgetting operation, the performance change amplitude of the performance index at the current moment is periodically detected, and whether to adaptively adjust the forgetting process parameters is determined according to the absolute value of the performance change amplitude;

[0026] If the absolute value of the performance change is greater than the preset performance change, the forgetting process parameters are adaptively adjusted;

[0027] If the absolute value of the performance change amplitude is less than or equal to the preset performance change amplitude, no adaptive adjustment is performed on the forgetting process parameters;

[0028] For a single moment, the performance change range is the difference between the initial performance indicator reference value and the real-time performance indicator reference value at that moment;

[0029] The forgetting process parameters include noise intensity and pruning scale.

[0030] Furthermore, the process of adaptively adjusting the noise intensity is periodically setting the noise intensity according to the performance variation amplitude;

[0031] The setting method of each noise intensity is determined according to the performance change range, where:

[0032] If the performance change amplitude is less than or equal to 0 and the absolute value of the performance change amplitude is greater than the preset performance change amplitude, the noise intensity at the current moment is reduced;

[0033] If the performance change amplitude is greater than 0 or the absolute value of the performance change amplitude is less than or equal to the preset performance change amplitude, the noise intensity at the current moment is maintained to perform subsequent data forgetting operations.

[0034] Furthermore, if the noise intensity at the current moment is reduced, the adjusted noise intensity is determined according to formula (4):

[0035] (4),

[0036] in, is the noise intensity at the current moment, is the noise intensity in the next performance adjustment cycle starting from the current time, is the preset noise learning rate, is the performance change at the current moment, The preset performance variation range.

[0037] Furthermore, the process of adaptively adjusting the pruning scale is to periodically set the pruning scale according to the performance variation range;

[0038] The setting method of each pruning scale is determined according to the performance change range. The setting methods of pruning scale include:

[0039] If the performance change range is greater than the preset performance change range, the pruning scale at the current moment is reduced;

[0040] If the performance change range is less than or equal to the preset performance change range, the current pruning scale is maintained to perform subsequent data forgetting operations.

[0041] Furthermore, if the current pruning scale is reduced, the adjusted pruning scale is determined according to formula (5):

[0042] (5),

[0043] in, is the trimming scale at the current moment, is the pruning scale in the next performance adjustment cycle starting from the current time, The default pruning learning rate.

[0044] Compared with the prior art, the beneficial effect of the present invention lies in that, in the technical scheme of the present invention, by combining the error maximization noise method, the neuron pruning method and the adaptive adjustment method, complete data forgetting is achieved without significantly affecting the data forgetting part of the model, and in the data forgetting process, there is no need to retrain the model, so the forgetting effect of the model on specific data can be guaranteed, thereby improving the processing efficiency of the data forgetting process.

[0045] Furthermore, the optimization function used in the process of generating the error maximization noise matrix in the present invention can maximize the error of specific data in the model, thereby weakening the model's dependence on the data, and by regularizing the noise parameters, it can accurately control the intensity and direction of the noise, ensuring that while achieving data forgetting, the impact on the overall performance of the model is minimized as much as possible.

[0046] Furthermore, the present invention detects the connection sensitivity of model parameters based on the obtained error maximization noise matrix, and selectively prunes key neurons according to the connection sensitivity of each model parameter, which can further reduce the model's memory for specific data, ensure the effectiveness of neuron pruning, and avoid unnecessary impact on model performance.

[0047] Furthermore, the present invention dynamically adjusts the noise intensity and the neuron pruning scale according to the performance change amplitude in the process of actually executing the data forgetting operation. By real-time monitoring of the model performance, an appropriate adjustment strategy is selected to avoid excessive negative impact of noise and neuron pruning on the model, thereby ensuring the data forgetting effect while retaining the overall performance of the model to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a data forgetting method based on error maximization noise and neuron pruning according to the present invention;

[0049] Figure 2 A flow chart of the present invention for determining whether to perform adaptive adjustment on forgetting process parameters according to the absolute value of the performance change amplitude;

[0050] Figure 3 A flow chart of iteratively updating the noise intensity according to the performance change amplitude of the present invention;

[0051] Figure 4 This is a flow chart of iteratively updating the pruning scale according to the performance change range of the present invention. DETAILED DESCRIPTION

[0052] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0053] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0054] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0055] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0056] See also Figures 1 to 4 As shown, the present invention provides a data forgetting method based on error maximization noise and neuron pruning, comprising:

[0057] Generate an error-maximizing noise matrix;

[0058] The connection sensitivity of each model parameter is detected, and the model parameters are initialized according to the connection sensitivity of each model parameter to induce model forgetting;

[0059] Update the target forgotten dataset and the retained dataset;

[0060] Periodically detect the performance change range, and determine whether to perform adaptive adjustments to the data forgetting process based on the performance change range;

[0061] If the forgetting process parameters are adaptively adjusted, the noise intensity and pruning scale are iteratively updated periodically according to the performance change amplitude;

[0062] When the forgetting effect is met or the preset number of iterations is reached, the iterative update of the forgetting process parameters is stopped.

[0063] Wherein, the model parameters are basic values ​​constituting the model. The model parameters in the present invention include but are not limited to weights and biases. The user defines a given data set, a target forgotten data set and a retained data set according to actual needs. The given data set is a set of data that needs to be processed by the model in the present invention, the target forgotten data set is a set of data that needs to be forgotten, and the retained data set is a set of data other than the given data set in the given data set. The forgetting effect is satisfied when the prediction accuracy of the model in the present invention for the data in the target forgotten data set is less than or equal to the preset prediction accuracy. How to determine the prediction accuracy of the model for the data in the target forgotten data set is easy to understand for those skilled in the art and will not be described in detail here. The values ​​of the preset prediction accuracy and the preset number of iterations can be determined by the user according to the actual working scenario. For example, the user can set it according to the running history record. The higher the user's requirement for the forgetting effect, the smaller the value of the preset prediction accuracy. A value of the preset prediction accuracy is provided, and the value of the preset prediction accuracy is 0. The higher the user's requirement for the model performance to be free from influence, the larger the value of the preset number of iterations. A method for determining the value of the preset number of iterations is provided, and the maximum value of the number of iterative updates when the forgetting effect is satisfied in the running history record that satisfies the user's requirement for the model performance to be free from influence is recorded as the preset number of iterations.

[0064] In the present invention, several operation history records are correspondingly set up, and any history record records the performance change range and the number of iterative updates during at least one execution of the data forgetting operation, and each history record corresponds to a qualified mark, which records whether the degree to which the model performance is affected during the execution of the data forgetting operation meets the user requirements.

[0065] Specifically, the generation process of the error maximization noise matrix includes:

[0066] Freeze the weights of the pre-trained model and randomly initialize a noise matrix from a standard normal distribution;

[0067] Obtain the error maximization noise matrix according to the optimization function;

[0068] The optimization function is formula (1),

[0069] (1),

[0070] in, To maximize the noise for the error, is the classification loss corresponding to the target forgotten category, is the currently trained model, is a noise matrix of the same size as the input to the model, For The strength of regularization, are model parameters, Forget the labels of each category of data in the target dataset.

[0071] in, express In the parameters Under the expectation, the error maximization noise matrix corresponding to the target forgotten category is obtained according to formula (1), and the weight of the pre-trained model is frozen, that is, the weight of the model is guaranteed not to be updated. How to freeze the weight of the pre-trained model is easy to understand for those skilled in the art and will not be described in detail here. The target forgotten data category is the category of data in the target forgotten data set. The function used to determine the classification loss corresponding to the target forgotten category in the present invention is the cross entropy loss function. The value of the regularization strength can be set by the user according to the actual model used. The value range of the regularization strength is (0,1). How to select the appropriate regularization strength according to the actual situation is easy to understand for those skilled in the art and will not be described in detail here.

[0072] An example of randomly initializing a noise matrix with the same size as the model input is provided. If the model input is a 224x224x3 image, i.e., a width of 224 pixels, a height of 224 pixels, and 3 color channels, then the added noise should also be a 224x224x3 matrix, in which each element is randomly drawn from a normal distribution (0,1). In the present invention, each data sample has a corresponding label, and the labels of each category of data in the target forgotten data set represent the categories that need to be forgotten. For example, if the model has been trained with data containing "cats" and "dogs", but now the category of "dogs" needs to be forgotten, the label of "dogs" is recorded as the label of the category of data "dogs" in the target forgotten data set.

[0073] Specifically, the connection sensitivity of each model parameter is detected, and the model parameters that need to be initialized are determined according to the connection sensitivity of each model parameter;

[0074] If the parameters are initialized, the target forgetting model is induced to forget the target forgetting data set according to the initialized model parameters.

[0075] Among them, the parameter initialization condition is that all model parameters that need to be initialized are initialized, and a preset number of model parameters are selected from large to small according to the connection sensitivity. The user can determine the value of the preset number according to the actual working scenario. The higher the user's requirement for the forgetting effect, the larger the value of the preset number. A preset number of values ​​is provided, and the preset number of values ​​is 30% of the number of model parameters. The selected model parameters are initialized, and the initialization process is to randomly assign values ​​to each model parameter. How to determine the preset number of model parameters that need to be initialized according to the actual working scenario and how to initialize the selected model parameters are easy for technical personnel in this field to understand and will not be elaborated here.

[0076] Specifically, the connection sensitivity of each model parameter is determined according to formula (2):

[0077] (2),

[0078] in, For the The connection sensitivity of the model parameters, For a given data set, the function Used to determine the loss of the model, For the model parameters, is the input of the model.

[0079] Among them, the function that determines the model loss in formula (2) is the cross entropy loss function, express exist The expectations below.

[0080] Specifically, the target forgotten data set and the retained data set are updated, and the updating process is performed according to formula (3):

[0081] (3),

[0082] in, To retain the dataset, For the target forget dataset, To obtain the best model parameters, For the The model parameters obtained by the cycle update are derived. For the The model parameters are updated in the next cycle.

[0083] Among them, according to as well as Formula (3) is solved, and the prediction accuracy of the model for the data in the target forgotten dataset and the retained dataset is minimized and the prediction accuracy of the model for the data in the retained dataset is maximized by performing cyclic updates on the target forgotten dataset and the retained dataset.

[0084] Specifically, before executing the data forgetting operation, the initial performance indicators of the model are recorded, and during the data forgetting operation, the performance change amplitude of the performance indicators at the current moment is periodically detected, and whether to adaptively adjust the forgetting process parameters is determined according to the absolute value of the performance change amplitude;

[0085] If the absolute value of the performance change is greater than the preset performance change, the forgetting process parameters are adaptively adjusted;

[0086] If the absolute value of the performance change amplitude is less than or equal to the preset performance change amplitude, no adaptive adjustment is performed on the forgetting process parameters;

[0087] For a single moment, the performance change range is the difference between the initial performance indicator reference value and the real-time performance indicator reference value at that moment;

[0088] The forgetting process parameters include noise intensity and pruning scale.

[0089] Among them, the present invention applies a cyclic performance adjustment cycle, and the duration of the performance adjustment cycle can be determined by the user. The higher the user's requirement for the model performance to be free from influence, the shorter the duration of the performance adjustment cycle. A performance adjustment cycle duration is provided, and the performance adjustment cycle is 5 minutes. At the end of each performance adjustment cycle, the performance change amplitude of the performance index at the current moment is detected, and it is determined whether to adaptively adjust the forgetting process parameters according to the performance change amplitude. Performing data forgetting operation means using error maximization noise and neuron pruning to process the model;

[0090] The user can select the performance indicator according to the actual working scenario. The present invention does not make specific limitations on the performance indicator used. A performance indicator used in the present invention is provided. The performance indicator is a loss function value. The loss function value is the absolute value of the difference between the output predicted by the loss function calculation model and the actual output. Different loss functions need to be selected for different tasks. How to select the loss function according to the actual situation is easy to understand for those skilled in the art and will not be elaborated here. The initial performance indicator reference value is the value of the performance indicator detected before executing data forgetting, and the real-time performance indicator reference value is the value of the performance indicator detected at the end of each performance adjustment cycle.

[0091] The value of the preset performance variation range can be determined by the user according to the actual working scenario. For example, the user can set it according to the operation history record. The higher the user's requirement for the model performance to be free from influence, the smaller the value of the preset performance variation range is. A method for determining the value of the preset performance variation range is provided, and the average of the absolute values ​​of the performance variation ranges at each moment in the operation history record that meets the user's requirement for the model performance to be free from influence and is not adjusted for the forgetting process parameters is recorded as the preset performance variation range; the noise intensity is the noise intensity when noise is injected into the model, and the injected noise is the generated error maximization noise. The pruning scale is used to determine the number of neurons retained during the neuron pruning process.

[0092] Specifically, the process of adaptively adjusting the noise intensity is to periodically set the noise intensity according to the performance change amplitude;

[0093] The setting method of each noise intensity is determined according to the performance change range, where:

[0094] If the performance change amplitude is less than or equal to 0 and the absolute value of the performance change amplitude is greater than the preset performance change amplitude, the noise intensity at the current moment is reduced;

[0095] If the performance change amplitude is greater than 0 or the absolute value of the performance change amplitude is less than or equal to the preset performance change amplitude, the noise intensity at the current moment is maintained to perform subsequent data forgetting operations.

[0096] Among them, if the absolute value of the performance change amplitude is greater than the preset performance change amplitude, from the start time of the adaptive adjustment, at the end of each performance adjustment cycle, the noise intensity at the current moment is set according to the performance change amplitude at the current moment. At the end time of a single performance adjustment cycle, if the performance change amplitude at this moment is greater than zero and the absolute value of the performance change amplitude at this moment is greater than the preset performance change amplitude, the noise intensity in the next performance adjustment cycle with this moment as the starting moment is set according to formula (4).

[0097] Specifically, if the noise intensity at the current moment is reduced, the adjusted noise intensity is determined according to formula (4):

[0098] (4),

[0099] in, is the noise intensity at the current moment, is the noise intensity in the next performance adjustment cycle starting from the current time, is the preset noise learning rate, is the absolute value of the performance change at the current moment, The preset performance variation range.

[0100] The value of the preset noise learning rate can be set by the user according to actual needs. A value of the preset noise learning rate is provided, and the value of the preset noise learning rate is 0.001.

[0101] Specifically, the process of iteratively updating the pruning scale is to periodically set the pruning scale according to the magnitude of performance change;

[0102] The setting method of each pruning scale is determined according to the performance change range. The setting methods of pruning scale include:

[0103] If the absolute value of the performance change range is greater than the preset performance change range, the pruning scale at the current moment is reduced;

[0104] If the absolute value of the performance change range is less than or equal to the preset performance change range, the pruning scale at the current moment is maintained to perform subsequent data forgetting operations.

[0105] Among them, if the absolute value of the performance change amplitude is greater than the preset performance change amplitude, starting from the start time of the adaptive adjustment, at the end of each performance adjustment cycle, the pruning scale at the current moment is set according to the performance change amplitude at the current moment. At the end time of a single performance adjustment cycle, if the absolute value of the performance change amplitude at this moment is greater than the preset performance change amplitude, the pruning scale within the next performance adjustment cycle with this moment as the starting moment is set according to formula (5).

[0106] Specifically, if the current pruning scale is reduced, the adjusted pruning scale is determined according to formula (5):

[0107] (5),

[0108] in, is the trimming scale at the current moment, is the pruning scale in the next performance adjustment cycle starting from the current time, The default pruning learning rate.

[0109] The value of the preset noise learning rate can be set by the user according to actual needs, and a value of a preset pruning learning rate is provided, and the value of the preset pruning learning rate is 0.001.

[0110] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A data forgetting method based on error maximization noise and neuron pruning, characterized in that: include: Generate an error-maximizing noise matrix; The connection sensitivity of each model parameter is detected, and the model parameters are initialized according to the connection sensitivity of each model parameter to induce model forgetting; Update the target forgotten dataset and the retained dataset; Periodically detect the performance change range, and determine whether to adaptively adjust the forgetting process parameters based on the performance change range; If the parameters of the data forgetting process are adjusted adaptively, the noise intensity and pruning scale are iteratively updated periodically according to the magnitude of performance changes; When the forgetting effect is satisfied or the preset number of iterations is reached, the iterative update of the forgetting process parameters is stopped; The generation process of the error maximization noise matrix includes: Freeze the weights of the pre-trained model and randomly initialize a noise matrix from a standard normal distribution; Obtain the error maximization noise matrix according to the optimization function; The optimization function is formula (1), Among them, N is the error maximization noise, L(f, y) is the classification loss corresponding to the target forgotten category, f is the currently trained model, ω noise is a noise matrix of the same size as the model input, and λ is the noise The strength of regularization, θ is the pre-training model parameter, y is the label of each category of data in the target forgetting dataset, E (θ) [-L(f,y)+λ||ω noise ||] means -L(f,y)+λ||ω noise ||Expectation under parameter θ; Detect the connection sensitivity of each model parameter, and determine the model parameters that need to be initialized according to the connection sensitivity of each model parameter; Wherein, if under the parameter initialization condition, the target forgetting model is induced to forget the target forgetting data set according to the initialized model parameters; The connection sensitivity of each model parameter is determined according to formula (2): Among them, s j (D) is the connection sensitivity of the jth model parameter, D is a given data set, and the function l(θ j ; x, y) is used to determine the model loss, θ j is the jth model parameter, x is the input of the model, express The expectation under (x, y)~D.

2. The data forgetting method based on error maximization noise and neuron pruning according to claim 1, characterized in that: The target forgotten data set and the retained data set are updated according to formula (3). i * =minL(θ^;D r )stθ^=argmaxL(θ t ;D w ) (3), Among them, D r To retain the data set, D w is the target forgetting dataset, θ * is the optimal model parameter finally obtained, θ^ is the derivative of the model parameter updated for the tth cycle, θ t The model parameters updated in the tth cycle.

3. The data forgetting method based on error maximization noise and neuron pruning according to claim 2, characterized in that: Before executing the data forgetting operation, the initial performance indicators of the model are recorded, and during the data forgetting operation, the performance change amplitude of the performance indicators at the current moment is periodically detected, and whether to adaptively adjust the forgetting process parameters is determined according to the absolute value of the performance change amplitude; If the absolute value of the performance change is greater than the preset performance change, the forgetting process parameters are adaptively adjusted; If the absolute value of the performance change amplitude is less than or equal to the preset performance change amplitude, no adaptive adjustment is performed on the forgetting process parameters; For a single moment, the performance change range is the difference between the initial performance indicator reference value and the real-time performance indicator reference value at that moment; The forgetting process parameters include noise intensity and pruning scale.

4. The data forgetting method based on error maximization noise and neuron pruning according to claim 3, characterized in that: The process of adaptively adjusting the noise intensity is to periodically set the noise intensity according to the performance change amplitude; The setting method of each noise intensity is determined according to the performance change range, where: If the performance change amplitude is less than or equal to 0 and the absolute value of the performance change amplitude is greater than the preset performance change amplitude, the noise intensity at the current moment is reduced; If the performance change amplitude is greater than 0 or the absolute value of the performance change amplitude is less than or equal to the preset performance change amplitude, the noise intensity at the current moment is maintained to perform subsequent data forgetting operations.

5. The data forgetting method based on error maximization noise and neuron pruning according to claim 4, characterized in that: If the noise intensity at the current moment is reduced, the adjusted noise intensity is determined according to formula (4): Among them, η t is the noise intensity at the current moment, η t+1 is the noise intensity in the next performance adjustment cycle starting from the current moment, α is the preset noise learning rate, ΔP is the performance change amplitude at the current moment, and ∈ is the preset performance change amplitude.

6. The data forgetting method based on error maximization noise and neuron pruning according to claim 5, characterized in that: The process of adaptively adjusting the pruning scale is to periodically set the pruning scale according to the performance change amplitude; The setting method of each pruning scale is determined according to the performance change range. The setting methods of pruning scale include: If the absolute value of the performance change range is greater than the preset performance change range, the pruning scale at the current moment is reduced; If the absolute value of the performance change range is less than or equal to the preset performance change range, the pruning scale at the current moment is maintained to perform subsequent data forgetting operations.

7. The data forgetting method based on error maximization noise and neuron pruning according to claim 6, characterized in that: If the current pruning scale is reduced, the adjusted pruning scale is determined according to formula (5): Among them, λ t is the pruning scale at the current moment, λ t+1 is the pruning scale in the next performance adjustment cycle starting from the current moment, and β is the preset pruning learning rate.

Images