Noise simulation method, noise evaluation method, system and storage medium
By building a noise feature library and training learning to generate a noise model that is adapted to multiple scenarios, the complex and incompatible noise distribution in optical simulation calculations are solved, and the simple generation and high accuracy of the noise model are achieved.
Patent Information
- Application Number
- CN202311863716.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-29
- Publication Date
- 2025-07-01
AI Technical Summary
In the field of optical simulation computing, it is difficult for the prior art to generate noise distributions that meet real scenes, and the noise model generation method is complex and incompatible with multi-scene requirements.
By constructing a noise feature library, using device sampling to obtain the noise set, building a parameterized initial noise model based on expert experience, and training the target noise model through reinforcement learning and heuristic learning to generate a noise distribution adapted to multiple scenarios.
It achieves stronger compatibility and simple operation of noise models, and can generate noise distributions that are more suitable for real scenes, improving the accuracy of artificial intelligence simulation learning.
Smart Images
Figure CN120234927A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of optical computing and machine learning, and particularly to a noise simulation method, a noise evaluation method, a system, and a storage medium. Background Art
[0002] In machine learning, noise refers to the unwanted signals introduced into the data, and noise is inevitable in machine learning.
[0003] In the process of simulated learning of artificial intelligence, if there is no noise, the simulated learning of artificial intelligence is an ideal state. When applying artificial intelligence to real life, the data received by artificial intelligence will inevitably contain noise. Therefore, the difference between the output result of artificial intelligence learned in the ideal state and the real result is relatively large.
[0004] For this reason, in the process of simulated learning of some artificial intelligence, the noise will be extracted first and then added in the subsequent simulated learning. However, in the field of optical simulation computing, due to the extremely complex pattern of noise, and when it is necessary to generate noise with a correct distribution, it is first necessary to determine the noise pattern, and then perform parameter search or tuning based on the determined noise pattern to find the noise generation model corresponding to the determined noise pattern, so as to generate a more appropriate noise distribution. Summary of the Invention
[0005] The present invention provides a noise simulation method, a noise evaluation method, a system, and a storage medium, aiming to simplify the generation method of the noise model and generate a more appropriate noise distribution.
[0006] According to the first aspect of the present invention, a noise simulation method is provided, including: obtaining a conditional noise set by using the sample data signal sampled by a device, where the condition is the generation condition of the noise signal; constructing a noise feature library by using the noise set, and the noise feature library includes at least two types of noise features; constructing an initial noise model with parameterization based on preset expert experience, where the noise generator outputs noise feature information corresponding to the at least two types of noise features according to the initial noise model; and training and learning the initial noise model by using the noise feature library to obtain a trained target noise model.
[0007] Further, the construction method of the noise feature library includes: constructing a noise feature method of interest based on preset expert experience, and incorporating the noise feature method into a pre-constructed noise feature method database; and calculating the noise set by using at least two noise feature methods in the noise feature method database to obtain the at least two types of noise features.
[0008] Further, the training and learning includes any one or a combination of reinforcement learning and heuristic learning.
[0009] Further, the step of training and learning the initial noise model by using the noise feature library includes: selecting two types of parameters from the noise set and the noise feature library, where one type of parameter is used for the reinforcement learning of the initial noise model, and the other type of parameter is used for the heuristic learning of the initial noise model.
[0010] Further, in the case where the training and learning is reinforcement learning, the step of training and learning the initial noise model includes: 1) using the noise set to train and update the initial noise model, and calculating a first state quantity for the initial noise model after training and updating; 2) calculating the first sum of squares of the first state quantity; 3) determining whether the first sum of squares is less than a first preset threshold, or whether the number of times of training and updating the initial noise model is greater than or equal to a first preset number of times; if the first sum of squares is less than the first preset threshold, or the number of times of training and updating is greater than or equal to the first preset number of times, then the reinforcement learning of the noise model is completed to obtain a trained target noise model; if the first sum of squares is greater than or equal to the first preset threshold, or the number of times of training and updating is less than the first preset number of times, then repeat steps 1)-3).
[0011] Further, the calculation method of the first state quantity includes: obtaining the training noise features output by the initial noise model when using the noise set to train and update the initial noise model; calculating the difference between the training noise features and the data features to obtain the first state quantity.
[0012] Further, the step of training and learning the initial noise model further includes: when the first sum of squares is greater than or equal to the first preset threshold, or the number of times of training and updating is less than the first preset number of times, then when continuing to train and update the noise model, modify the parameters of the initial noise model.
[0013] Further, in the case where the training and learning is heuristic learning, the step of training and learning the initial noise model includes: A) constructing an initialization condition; B) under each of the initialization conditions, using the noise set to train and update the initial noise model, and calculating a second state quantity after the training and update of the initial noise model; C) calculating a second sum of squares of the second state quantity; D) determining whether the second sum of squares is less than a second preset threshold, or whether the number of times of training and updating the initial noise model is greater than or equal to a second preset number of times; if the second sum of squares is less than the second preset threshold, or the number of times of training and updating is greater than or equal to the second preset number of times, then the reinforcement learning of the initial noise model is completed to obtain a trained target noise model; if the second sum of squares is greater than or equal to the second preset threshold, or the number of times of training and updating is less than the second preset number of times, then repeat steps A)-D).
[0014] According to a second aspect of the present invention, the present invention also provides a method for noise evaluation of optical analog computing, including: obtaining a sample signal in machine learning, and calculating an effective signal in the sample signal according to an input and a weight; removing the effective signal from the sample signal to obtain a sample noise signal; establishing a noise model matching the sample noise signal, and generating a simulated noise signal by using the noise model; comparing the simulated noise signal with the sample noise signal to evaluate the noise model; wherein, the noise model is obtained by using the noise simulation method described in any one of the above.
[0015] Further, when evaluating the noise model, the evaluation method of the target noise model is reinforcement learning evaluation or / and heuristic learning evaluation.
[0016] Further, when the target noise model is evaluated using reinforcement learning, the evaluation steps include: A1) judging the data distribution of the noise with the additional specific parameter conditions, and selecting a target noise model according to the data distribution; B1) updating the target noise model with the noise that conforms to the data distribution and has the additional specific parameter conditions, and obtaining the state quantity after the target noise model is updated; C1) calculating the difference value between the true feature and the sample noise by using the state quantity after the target noise model is updated; D1) calculating a first conditional value for exiting the evaluation by using the difference value, obtaining a preset first maximum number of iterations for the evaluation, and obtaining a first target conditional threshold for exiting the evaluation loop; F1) iterating on the noise that conforms to the data distribution and has the additional specific parameter conditions, and judging whether the first conditional value is less than the first target conditional threshold, or whether the number of iterations is equal to the first maximum number of iterations; if the first conditional value is less than the first target conditional threshold or the number of iterations is equal to the first maximum number of iterations, then exit the iteration and complete the evaluation of the reinforcement learning; if the first conditional value is greater than or equal to the first target conditional threshold and the number of iterations is not equal to the first maximum number of iterations, then modify the parameter conditions of the noise with the additional specific parameter conditions, and repeat steps A1)-F1).
[0017] Further, the data distribution includes at least one of a standard Gaussian distribution, a Poisson distribution, and a chi-square distribution.
[0018] Further, when the target noise model is evaluated using heuristic learning, the evaluation steps include: obtaining the parameters encoded by the heuristic learning method; generating an initialization condition for the encoding method based on the encoding method of the heuristic learning method; optimizing the noise model under each of the initialization conditions; obtaining a second target conditional threshold and a second maximum number of iterations for exiting the evaluation; evaluating each set of encoded parameters after optimization, and calculating a second conditional value for exiting the evaluation; judging whether the second conditional value for exiting is less than the second target conditional threshold or the number of times of evaluation is equal to the second maximum number of iterations, and if so, then exit the iteration and complete the evaluation of the heuristic learning; if not, then perform a new round of heuristic learning evaluation according to the parameters encoded by different heuristic learning methods.
[0019] Further, in the step of optimizing the noise model under each of the initialization conditions, the noise model includes a noise model generated according to expert experience and a model generated according to the reinforcement learning evaluation.
[0020] Further, the parameter information includes any one or a combination of at least any two of temperature, weight, and input value.
[0021] According to the third aspect of the present invention, the present invention further provides a noise simulation system, including: a conditional noise set acquisition module, configured to sample sample data signals from a device, and obtain a training noise set from the sample data signals, where the training noise set is obtained by extracting a sample background signal from the sample data signals, and the condition is the sample generation condition of the sample data signals; a feature library construction module, configured to construct a noise feature library by using the training noise set, where the noise feature library includes at least two types of noise features; a noise model construction module, configured to construct an initial noise model with parameterization based on preset expert experience, so that a noise generator outputs noise feature information corresponding one-to-one to the at least two types of noise features according to the initial noise model; a noise model learning module, configured to perform training and learning on the initial noise model by using the noise feature library to obtain a trained target noise model.
[0022] According to the fourth aspect of the present invention, the present invention further provides a noise evaluation system, including: a signal acquisition module, configured to acquire sample signals in machine learning, and calculate effective signals in the sample signals according to inputs and weights; a signal processing module, configured to remove the effective signals from the sample signals to obtain sample noise signals; a simulated noise signal generation module, configured to establish a noise model matching the sample noise signals, and generate simulated noise signals by using the noise model; a noise model evaluation module, configured to compare the simulated noise signals with the sample noise signals to evaluate the noise model; where the noise model is obtained by using the above-mentioned noise simulation system.
[0023] According to the fifth aspect of the present invention, the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the noise simulation method described in any one of the above is implemented.
[0024] According to the sixth aspect of the present invention, the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the noise evaluation method for optical analog computing described in any one of the above is implemented.
[0025] Through one or more of the above-mentioned embodiments in the present invention, at least the following technical effects can be achieved:
[0026] In the technical solution disclosed by the present invention, when generating a noise model, by using the generation conditions of noise and using different types of noise features to train a parameterized noise model, the noise model can be made to adapt to the noise requirements in various scenarios, thereby ensuring the generalization of the noise model. Moreover, during the training process of the noise model, it is not necessary to first determine the pattern of the noise, nor is it necessary to perform parameter search or tuning based on the determined noise pattern. Therefore, the generation method of the noise model is not only more compatible, but also simpler to operate, and can generate a noise distribution that better fits the real scenario. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0028] Figure 1 It is a flowchart of the noise simulation method provided by an embodiment of the present invention;
[0029] Figure 2 It is a flowchart of the noise evaluation method for optical simulation calculation provided by an embodiment of the present invention;
[0030] Figure 3 It is a framework diagram of the noise simulation system provided by an embodiment of the present invention;
[0031] Figure 4 It is a framework diagram of the noise evaluation system for optical simulation calculation provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0033] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after without special explanation.
[0034] In machine learning, noise refers to the unwanted signals introduced into the data, and noise is inevitable in machine learning. In AI simulation learning, if no noise is added, it is an ideal state and no noise will be automatically introduced. Therefore, adding noise that conforms to the characteristics of machine learning during the AI simulation learning process can help the model better learn the authenticity of the data, make it closer to the state in machine learning, and improve its accuracy.
[0035] The existing methods for generating noise are mainly in the non-simulation computing field, and basically all are known noise models constructed based on expert experience. Then, noise is generated on the determined noise model. There are also some AI-based ways to construct noise. The disadvantages are that large-scale noise generation cannot be done and it can only be used in specific fields.
[0036] Therefore, in the simulation learning process of some artificial intelligence, noise will be extracted first and then added in the subsequent simulation learning. However, in the field of optical simulation computing, due to the extremely complex pattern of noise, and when noise with the correct distribution needs to be generated, first, the noise pattern needs to be determined, and then parameter search or tuning is performed based on the determined noise pattern to find the correct noise generation model, so as to generate noise with the correct distribution.
[0037] To solve the above problems, this application provides a noise simulation method, a noise evaluation method, a system, and a storage medium, specifically as follows:
[0038] Figure 1 The noise simulation method provided by the embodiments of the present invention is shown as follows and includes:
[0039] S101. Obtain a conditional noise set using the sample data signal sampled by the device, where the condition is the generation condition of the noise signal;
[0040] S102. Construct a noise feature library using the noise set, and the noise feature library includes at least two types of noise features;
[0041] S103. Based on the preset expert experience, construct an initial noise model with parameterization processing, where the noise generator outputs noise feature information corresponding one by one to at least two types of noise features according to the initial noise model;
[0042] S104. Use the noise feature library to train and learn the initial noise model to obtain a trained target noise model.
[0043] In this embodiment, the data features in the data set are the true noises of the sample data. When constructing a parameterized initial noise model, multiple models will be constructed according to different parameters, and the set of finally learned true noise models is a subset of the parameterized noise model, where the parameters can be temperature, input, weight, etc.; in addition, the construction of the noise feature library can store more noise features, and sufficient data support can be provided when constructing / evaluating complex noise generators.
[0044] The noise simulation method provided in this embodiment, when generating a noise model, by using the generation conditions of the noise and using different types of noise features to train the parameterized noise model, can enable the noise model to adapt to the noise requirements in various scenarios, thereby ensuring the generalization of the noise model. Moreover, during the training process of the noise model, it is not necessary to first determine the noise pattern, nor is it necessary to perform parameter search or tuning based on the determined noise pattern. Therefore, the generation method of the noise model is not only more compatible, but also more convenient to operate, and can generate a noise distribution that is more in line with the real scenario.
[0045] In one embodiment, the method for constructing a noise feature library includes: constructing a noise feature method of interest based on preset expert experience, and incorporating the noise feature method into a pre-constructed noise feature method database; calculating a noise set using at least two noise feature methods in the noise feature method database to obtain at least two types of noise features.
[0046] In this embodiment, when constructing the noise feature method and the noise feature method database, based on the target generation conditions, when the noise conforms to a preset distribution, a noise feature method of interest is constructed, that is, a noise feature method that conforms to the target generation conditions is constructed, and the noise feature method is stored in the noise feature method database. For example, under the conditions of a specific temperature / input / weight, the noise conforms to a Gaussian distribution, or a Poisson distribution, or a chi-square distribution, etc., or a combination of the above distributions, and a noise feature method of interest is constructed and all the constructed noise feature methods are stored in the noise feature method database.
[0047] And after obtaining at least two types of noise features, one of the at least two types of noise is used for evaluation during the construction of the noise model, and the other is used for iterative evaluation during the learning of the noise model.
[0048] In one embodiment, the noise set with noise set conditions obtained by using the sample data signal sampled by the device in step S101 includes: under specific conditions, such as input, weight, temperature, humidity, etc., the sample data signal sampled from the device, and calculating the effective signal in the sample signal according to the input and weight; removing the effective signal from the sample signal to obtain a sample noise signal. In one embodiment, the training learning includes any one or a combination of reinforcement learning and heuristic learning. Among them, the training and learning of the initial noise model using the noise feature library in step S104 includes: selecting two types of parameters from the noise set and the noise feature library, where one type of parameter is used for the reinforcement learning of the initial noise model, and the other type of parameter is used for the heuristic learning of the initial noise model. Specifically:
[0049] In one embodiment, in the case where the training learning is reinforcement learning, the steps of training and learning the initial noise model include:
[0050] 1) Use the noise set to train and update the initial noise model, and calculate the first state quantity for training and updating the initial noise model;
[0051] 2) Calculate the first sum of squares of the first state quantity;
[0052] 3) Determine whether the first sum of squares is less than the first preset threshold, or whether the number of times of training and updating the initial noise model is greater than or equal to the first preset number of times;
[0053] If the first sum of squares is less than the first preset threshold, or the number of times of training and updating is greater than or equal to the first preset number of times, then complete the reinforcement learning of the noise model to obtain the trained target noise model;
[0054] If the first sum of squares is greater than or equal to the first preset threshold, or the number of times of training and updating is less than the first preset number of times, then repeat steps 1)-3).
[0055] In this embodiment, the first sum of squares represents the difference in noise feature comparison, and the noise model is trained by using the method of reinforcement learning, so that the noise model can quickly search the received data signal to obtain the local optimal solution under the noise model.
[0056] In one embodiment, the calculation method of the first state quantity includes:
[0057] Obtain the training noise features output by the initial noise model when using the noise set to train and update the initial noise model; calculate the difference between the training noise features and the data features to obtain the first state quantity.
[0058] In this embodiment, the difference between the training noise features and the true noise features is represented by the first state quantity.
[0059] In one embodiment, the step of training and learning the initial noise model further includes: if the first sum of squares is greater than or equal to the first preset threshold, or the number of training updates is less than the first preset number, then when continuing to train and update the noise model, modify the parameters of the initial noise model.
[0060] In this embodiment, the parameters of the noise model are the above-mentioned target generation conditions. If the first sum of squares is greater than or equal to the first preset threshold, or the number of training updates is less than the first preset number, it may indicate that under this target condition, a suitable model cannot be trained. At this time, by modifying the parameters of the noise model, the noise model can be trained under another parameter.
[0061] In one embodiment, in the case where the training and learning is heuristic learning, the steps of training and learning the initial noise model include:
[0062] A) Construct an initialization condition;
[0063] B) Under each initialization condition, use the noise set to train and update the initial noise model, and calculate the second state quantity after the training and update of the initial noise model;
[0064] C) Calculate the second sum of squares of the second state quantity;
[0065] D) Determine whether the second sum of squares is less than the second preset threshold, or whether the number of training updates of the initial noise model is greater than or equal to the second preset number;
[0066] If the second sum of squares is less than the second preset threshold, or the number of training updates is greater than or equal to the second preset number, then complete the reinforcement learning of the initial noise model to obtain the trained target noise model;
[0067] If the second sum of squares is greater than or equal to the second preset threshold, or the number of training updates is less than the second preset number, then repeat steps A)-D).
[0068] In this embodiment, the calculation method of the second state quantity is the same as that of the first state quantity in the above embodiment, which will not be elaborated in this embodiment. This embodiment uses the heuristic learning method to train the initial noise model, so that the target noise model can fully search the received data signal, and the search result can be close to the global optimal solution.
[0069] In one embodiment, when training and learning the initial noise model, the trained target noise model is obtained by means of reinforcement learning and heuristic learning of the noise model.
[0070] In this embodiment, the noise model is trained by means of reinforcement learning and heuristic learning, so that the noise model can quickly search for the received data signal, fully search for the data signal, and obtain the local optimal solution and the global optimal solution under the noise model.
[0071] Please refer to Figure 2 , this embodiment of the present application also provides a method for evaluating the noise of optical analog computing. The noise evaluation method includes:
[0072] S201. Obtain a sample signal in machine learning, and calculate the effective signal in the sample signal according to the input and weight;
[0073] S202. Remove the effective signal from the sample signal to obtain a sample noise signal;
[0074] S203. Establish a noise model matching the sample noise signal, and generate a simulated noise signal by using the noise model;
[0075] S204. Compare the simulated noise signal with the sample noise signal to evaluate the noise model;
[0076] In step S201, the sample signal can also be an electrical signal, an optoelectronic hybrid signal, etc. The sample signal is mixed with an effective signal and a noise signal.
[0077] In step S204, the noise model is trained by using the noise simulation method in any of the above embodiments.
[0078] The method for evaluating the noise of optical analog computing proposed in this embodiment can evaluate the noise, so that the qualified noise can be applied to the simulation learning of artificial intelligence, which can help the simulation learning model of artificial intelligence better learn the authenticity of data and be closer to the state of machine learning. When applying the simulation learning model of artificial intelligence to machine learning, since the simulation learning model of artificial intelligence has considered the influence of noise, the accuracy of the output result of artificial intelligence is improved.
[0079] In one embodiment, when evaluating the noise model, the evaluation method of the target noise model is reinforcement learning evaluation or / and heuristic learning evaluation.
[0080] The reinforcement learning evaluation described in this embodiment can quickly search for the local optimal solution under the current noise generation model, while the heuristic learning evaluation can ensure that the result is close to the global optimal solution after full search.
[0081] In one embodiment, when the target noise model is evaluated using reinforcement learning, the evaluation steps include: A1) judging the data distribution of the noise with additional specific parameter conditions, and selecting the target noise model according to the data distribution; B1) updating the target noise model with the noise that meets the additional specific parameter conditions of the data distribution, and obtaining the state quantity after the target noise model is updated; C1) calculating the difference value between the true feature and the sample noise using the state quantity after the target noise model is updated; D1) calculating the first condition value for exiting the evaluation using the difference value, obtaining the first maximum number of iterations preset for the evaluation, and obtaining the first target condition threshold for exiting the evaluation loop; F1) iterating on the noise that meets the additional specific parameter conditions of the data distribution, and judging whether the first condition value is less than the first target condition threshold or whether the number of iterations is equal to the first maximum number of iterations; if the first condition value is less than the first target condition threshold or the number of iterations is equal to the first maximum number of iterations, exit the iteration and complete the evaluation of the reinforcement learning; if the first condition value is greater than or equal to the first target condition threshold and the number of iterations is not equal to the first maximum number of iterations, modify the parameter conditions of the noise with the additional specific parameter conditions, and repeat steps A1)-F1).
[0082] In this embodiment, the data distribution includes at least one of a standard Gaussian distribution, a Poisson distribution, and a chi-square distribution. Exemplarily, if the data distribution in this embodiment is a Gaussian distribution, the noise that meets the specific parameter conditions of the Gaussian distribution can be used as the training sample for training the target noise model, so that the target noise model can learn the noise, and in the process of application, the target noise model can reduce the influence of the output result by the noise, thereby improving the accuracy of the output of the target noise model.
[0083] In this embodiment, the value of the state quantity is the difference between the noise feature generated by the target noise model and the true noise feature, which can be represented by {Fea1 - sFea1}. The noise feature can be extracted using the target noise model, where Fea1 is the noise feature generated by the target noise model and sFea1 is the true noise feature.
[0084] The calculation method of the first condition value is the sum of squares of the state quantity, which can be represented by {sum(Fea1 - sFea1) 2}.
[0085] In one embodiment, when the target noise model is evaluated using heuristic learning, the evaluation steps include: obtaining parameters encoded by multiple heuristic learning methods; generating initialization conditions for the encoding method based on the encoding method of the heuristic learning method; optimizing the noise model under each initialization condition; obtaining a second target condition threshold for exiting the evaluation and a second maximum number of iterations; evaluating each set of encoded parameters after optimization and calculating a second condition value for exiting the evaluation; determining whether the second condition value for exiting is less than the second target condition threshold or the number of evaluations is equal to the second maximum number of iterations. If so, exit the iteration and complete the heuristic learning evaluation; if not, perform a new round of heuristic learning evaluation based on the parameters encoded by different heuristic learning methods.
[0086] In this embodiment, in the step of optimizing the noise model under each initialization condition, the noise model includes a noise model generated according to expert experience and a model generated according to reinforcement learning evaluation.
[0087] Specifically, in this embodiment, in the step of optimizing the noise model under each initialization condition, if the noise model is generated according to expert experience, the criteria for heuristic evaluation can be achieved, thereby evaluating the noise data globally; while if the noise model is the model generated by the reinforcement learning evaluation in the above embodiment, during the evaluation learning process, reinforcement learning is also performed. This ensures that when the heuristic learning evaluation searches for the global optimal solution, local optimal solutions are also evaluated using reinforcement learning. Since the local optimal solutions have already been given during the reinforcement learning evaluation, during the heuristic learning evaluation, only a few local optimal solutions need to be searched for the global optimal solution. This reduces the amount of data and time for the global search and improves the efficiency of evaluating the noise.
[0088] Please refer to Figure 3 , the embodiment of the present application also provides a noise simulation system. The noise simulation system includes: a conditional noise set acquisition module 1, a feature library construction module 2, a noise model construction module 3, and a noise model learning module 4; wherein, the conditional noise set acquisition module is used to sample the sample data signal from the device and obtain a training noise set from the sample data signal. The training noise set is obtained by extracting the sample background signal from the sample data signal, and the condition is the sample generation condition of the sample data signal; the feature library construction module 2 is used to construct a noise feature library using the noise set, and the noise feature library includes at least two types of noise features; the noise model construction module 3 is used to construct an initial noise model with parameterization based on preset expert experience, so that the noise generator outputs noise feature information corresponding to at least two types of noise features according to the initial noise model; the noise model learning module 4 is used to train and learn the initial noise model using the noise feature library to obtain a trained target noise model.
[0089] The noise simulation system disclosed in this embodiment, when generating a noise model, can make the noise model adapt to the noise requirements in various scenarios by using the generation conditions of noise and training the parameterized noise model with different types of noise characteristics, thereby ensuring the generalization of the noise model. Moreover, during the training process of the noise model, it is not necessary to first determine the noise pattern, nor to perform parameter search or tuning based on the determined noise pattern. Therefore, the generation method of the noise model is not only more compatible, but also easier to operate, and can generate a noise distribution that is more consistent with the real scenario.
[0090] In one embodiment, the feature library construction module 2 includes: a noise feature method construction unit and a noise feature calculation unit; the noise feature method construction unit is used to construct a noise feature method of interest based on preset expert experience and incorporate the noise feature method into a pre-constructed noise feature method database; the noise feature calculation unit is used to calculate at least two types of noise features for a noise set by using at least two noise feature methods in the noise feature method database.
[0091] In one embodiment, the noise model learning module 4 is specifically configured to select two types of parameters from the noise set and the noise feature library, where one type of parameter is used for the reinforcement learning of the initial noise model, and the other type of parameter is used for the heuristic learning of the initial noise model.
[0092] In one embodiment, the noise model learning module 4 includes: a first training update unit, a first calculation unit, and a first execution unit; wherein, the first training update unit is used to train and update the initial noise model using the noise set and calculate a first state quantity for the trained and updated initial noise model; the first calculation unit is used to calculate the first sum of squares of the first state quantity; the first execution unit is used to determine whether the first sum of squares is less than a first preset threshold, or whether the number of times of training and updating the initial noise model is greater than or equal to a first preset number; if the first sum of squares is less than the first preset threshold, or the number of times of training and updating is greater than or equal to the first preset number, then the reinforcement learning of the noise model is completed to obtain a trained target noise model; if the first sum of squares is greater than or equal to the first preset threshold, or the number of times of training and updating is less than the first preset number, then the first training update unit, the first calculation unit, and the first execution unit are called again.
[0093] In one embodiment, the first training update unit includes: a training noise feature acquisition subunit and a first state quantity calculation subunit; wherein, the training noise feature acquisition subunit is used to acquire the training noise features output by the initial noise model when the noise set is used to train and update the initial noise model; the first state quantity calculation subunit is used to calculate the difference between the training noise features and the data features to obtain a first state quantity.
[0094] In one embodiment, the noise model learning module 4 further includes: a parameter modification unit, configured to modify the parameters of the initial noise model when the first sum of squares is greater than or equal to a first preset threshold, or the number of training updates is less than a first preset number, and then continue to train and update the noise model.
[0095] In one embodiment, the noise model learning module includes: an initialization unit, a second training and update unit, a second calculation unit, and a second execution unit; wherein, the initialization unit is configured to construct an initialization condition; the second training and update unit is configured to use a noise set to train and update the initial noise model under each initialization condition, and calculate a second state quantity after the initial noise model is trained and updated; the second calculation unit is configured to calculate a second sum of squares of the second state quantity; the second execution unit is configured to determine whether the second sum of squares is less than a second preset threshold, or whether the number of times of training and updating the initial noise model is greater than or equal to a second preset number; if the second sum of squares is less than the second preset threshold, or the number of times of training and updating is greater than or equal to the second preset number, then the reinforcement learning of the initial noise model is completed to obtain a trained target noise model; if the second sum of squares is greater than or equal to the second preset threshold, or the number of times of training and updating is less than the second preset number, then the initialization unit, the second training and update unit, the second calculation unit, and the second execution unit are called again.
[0096] Please refer to Figure 4 , an embodiment of the present invention further provides a noise evaluation system, including:
[0097] A signal acquisition module 11, a signal processing module 12, a simulated noise signal generation module 13, and a noise model evaluation module 14; the signal acquisition module 11 is configured to acquire a sample signal in machine learning, and calculate an effective signal in the sample signal according to an input and a weight; the signal processing module 12 is configured to remove the effective signal from the sample signal to obtain a sample noise signal; the simulated noise signal generation module 13 is configured to establish a noise model matching the sample noise signal, and generate a simulated noise signal by using the noise model; the noise model evaluation module 14 is configured to compare the simulated noise signal with the sample noise signal to evaluate the noise model.
[0098] In this embodiment, the noise model is obtained by using the noise simulation system in the above embodiment.
[0099] The noise evaluation system provided in this embodiment enables qualified noise to be applied to the simulation learning of optical analog computing. Since the simulation learning of optical analog computing has considered the influence of noise, the difference between the simulation learning result of optical analog computing and the true result of machine learning is reduced, and the accuracy of the simulation result of optical analog computing is improved.
[0100] In one embodiment, when the noise model evaluation module 14 evaluates a noise model, it selects two types of parameters from a noise set and a noise feature library. One type of parameter is used for the reinforcement learning of the initial noise model, and the other type of parameter is used for the heuristic learning of the initial noise model.
[0101] In one embodiment, when the noise model evaluation module 14 uses reinforcement learning to evaluate an initial noise model, the evaluation steps include: A1) judging the data distribution of the noise with additional specific parameter conditions; B1) updating the target noise model with the noise that meets the data distribution and has additional specific parameter conditions, and obtaining the state quantity of the target noise model after the update; C1) calculating the difference value between the true feature and the sample noise by using the state quantity of the target noise model after the update; D1) calculating the first condition value for exiting the evaluation by using the difference value, obtaining the first maximum number of iterations preset for the evaluation, and obtaining the first target condition threshold for exiting the evaluation loop; F1) iterating the noise that meets the data distribution and has additional specific parameter conditions, and judging whether the first condition value is less than the first target condition threshold or the number of iterations is equal to the first maximum number of iterations; if the first condition value is less than the first target condition threshold or the number of iterations is equal to the first maximum number of iterations, then exit the iteration and complete the evaluation of the reinforcement learning; if the first condition value is greater than or equal to the first target condition threshold and the number of iterations is not equal to the first maximum number of iterations, then modify the parameter conditions of the noise with additional specific parameter conditions, and repeat steps A1)-F1).
[0102] In one embodiment, the data distribution includes at least one of a Gaussian distribution, a Poisson distribution, and a chi-square distribution.
[0103] In one embodiment, when the target noise model used by the simulated noise signal generation module 13 is in the form of heuristic learning evaluation, the evaluation steps include: obtaining the parameters encoded by the heuristic learning method; generating the initialization conditions of the encoding method based on the encoding method of the heuristic learning method; optimizing the noise model under each initialization condition; obtaining the second target condition threshold and the second maximum number of iterations for exiting the evaluation; evaluating each set of encoded parameters after optimization and calculating the second condition value for exiting the evaluation; judging whether the second condition value for exiting is less than the second target condition threshold or the number of evaluations is equal to the second maximum number of iterations. If so, then exit the iteration and complete the evaluation of the heuristic learning; if not, then perform a new round of heuristic learning evaluation according to the parameters encoded by different heuristic learning methods.
[0104] In one embodiment, in the step of optimizing the noise model under each initialization condition, the noise model includes a noise model generated according to expert experience and a model generated according to the reinforcement learning evaluation.
[0105] An embodiment of the present application also provides a computer-readable storage medium, which may be disposed in an electronic device.
[0106] The electronic device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the noise simulation method described above or the noise evaluation method for optical analog computing is implemented.
[0107] Furthermore, the electronic device further includes: at least one input device and at least one output device.
[0108] The above-mentioned memory, processor, input device, and output device are connected via a bus.
[0109] Among them, the input device may specifically be a camera, a touch panel, a physical button, or a mouse, etc. The output device may specifically be a display screen.
[0110] The memory may be a high-speed random access memory (RAM) or a non-volatile memory, such as a disk memory. The memory is used to store a set of executable program codes, and the processor is coupled to the memory.
[0111] The computer-readable storage medium may be the memory described above. A computer program is stored on the computer-readable storage medium, and when the program is executed by the processor, the noise simulation method described in the foregoing embodiments or the noise evaluation method for optical analog computing is implemented.
[0112] Furthermore, the computer-readable storage medium may also be various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a RAM, a magnetic disk, or an optical disc that can store program codes.
[0113] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or modules can be in electrical, mechanical, or other forms.
[0114] The module described as a separation component may or may not be physically separated. The component shown as a module may or may not be a physical module, that is, it may be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0115] In addition, in each embodiment of the present invention, the functional modules can be integrated into a processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules.
[0116] If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention.
[0117] It should be noted that for the foregoing method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily all essential to the present invention.
[0118] In the above embodiments, the descriptions of each embodiment have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0119] The above is the description of a noise simulation method, a noise assessment method, a system and a storage medium provided by the present invention. For those skilled in the art, according to the idea of the embodiments of the present invention, there will be changes in the specific implementation manners and application scopes.
[0120] In summary, although the present invention has been disclosed above with preferred embodiments, the above preferred embodiments are not intended to limit the present invention. Those of ordinary skill in the art can make various modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention is subject to the scope defined by the claims.
Claims
1. A noise simulation method, characterized in that, Including: Obtaining a conditional noise set from the sample data signal sampled by the device, where the condition is the generation condition of the noise signal; Constructing a noise feature library using the noise set, where the noise feature library includes at least two types of noise features; Based on preset expert experience, constructing an initial noise model with parameterization processing, where the noise generator outputs noise feature information corresponding one by one to the at least two types of noise features according to the initial noise model; Training and learning the initial noise model using the noise feature library to obtain a trained target noise model.
2. The noise simulation method according to claim 1, characterized in that The construction method of the noise feature library includes: Based on preset expert experience, constructing a noise feature method of interest and incorporating the noise feature method into a pre-constructed noise feature method database; Calculating the at least two types of noise features from the noise set using at least two noise feature methods in the noise feature method database.
3. The noise simulation method according to claim 1, characterized in that The training and learning includes any one or a combination of reinforcement learning and heuristic learning.
4. The noise simulation method according to claim 1, characterized in that The step of training and learning the initial noise model using the noise feature library includes: selecting two types of parameters from the noise set and the noise feature library, where one type of parameter is used for the reinforcement learning of the initial noise model and the other type of parameter is used for the heuristic learning of the initial noise model.
5. The noise simulation method according to claim 3 or 4, characterized in that In the case where the training and learning is reinforcement learning, the step of training and learning the noise model includes: 1) Using the noise set to train and update the initial noise model and calculating a first state quantity for the initial noise model after training and updating; 2) Calculating the first sum of squares of the first state quantity; 3) Judging whether the first sum of squares is less than a first preset threshold or whether the number of times of training and updating the initial noise model is greater than or equal to a first preset number of times; If the first sum of squares is less than the first preset threshold or the number of times of training and updating is greater than or equal to the first preset number of times, then the reinforcement learning of the noise model is completed to obtain a trained target noise model; If the first sum of squares is greater than or equal to the first preset threshold or the number of times of training and updating is less than the first preset number of times, then repeat steps 1)-3).
6. The noise simulation method according to claim 5, characterized in that The calculation method of the first state quantity includes: Obtaining the training noise features output by the initial noise model when using the noise set to train and update the initial noise model; Calculating the difference between the training noise features and the data features to obtain the first state quantity.
7. The noise simulation method according to claim 5, characterized in that The step of training and learning the initial noise model further includes: If the first sum of squares is greater than or equal to the first preset threshold, or the number of training updates is less than the first preset number, then when continuing to train and update the noise model, the parameters of the initial noise model are modified.
8. The noise simulation method according to claim 3 or 4, characterized in that In the case where the training learning is heuristic learning, the step of training and learning the initial noise model includes: A) Construct initialization conditions; B) Under each of the initialization conditions, use the noise set to train and update the initial noise model, and calculate the second state quantity after the training and update of the initial noise model; C) Calculate the second sum of squares of the second state quantity; D) Determine whether the second sum of squares is less than a second preset threshold, or whether the number of training updates of the initial noise model is greater than or equal to a second preset number; If the second sum of squares is less than the second preset threshold, or the number of training updates is greater than or equal to the second preset number, then the reinforcement learning of the initial noise model is completed to obtain a trained target noise model; If the second sum of squares is greater than or equal to the second preset threshold, or the number of training updates is less than the second preset number, then repeat steps A)-D).
9. A method for evaluating the noise of optical analog computing, characterized in that, including: Obtain a sample signal in machine learning, and calculate the effective signal in the sample signal according to the input and weights; Remove the effective signal from the sample signal to obtain a sample noise signal; Establish a noise model matching the sample noise signal, and generate a simulated noise signal using the noise model; Compare the simulated noise signal with the sample noise signal to evaluate the noise model; Wherein, the noise model is obtained by using the noise simulation method according to any one of claims 1-8.
10. The method for evaluating the noise of optical analog computing according to claim 9, characterized in that, When evaluating the noise model, the evaluation method of the target noise model is reinforcement learning evaluation or / and heuristic learning evaluation.
11. The method for evaluating the noise of optical analog computing according to claim 10, characterized in that, When the target noise model uses reinforcement learning evaluation, the evaluation steps include: A1) Determine the data distribution of the noise with the additional specific parameter conditions, and select the target noise model according to the data distribution; B1) Update the target noise model with the noise with the additional specific parameter conditions that conforms to the data distribution, and obtain the state quantity after the update of the target noise model; C1) Calculate the difference value between the true feature and the sample noise using the state quantity after the update of the target noise model; D1) Calculate the first condition value for exiting the evaluation using the difference value, obtain the first maximum number of iterations preset for the evaluation, and obtain the first target condition threshold for exiting the evaluation loop; F1) Iterate over the noise with the additional specific parameter conditions that conform to the data distribution, and determine whether the first condition value is less than the first target condition threshold, or whether the number of iterations is equal to the first maximum number of iterations; If the first condition value is less than the first target condition threshold or the number of iterations is equal to the first maximum number of iterations, exit the iteration and complete the evaluation of the reinforcement learning; if the first condition value is greater than or equal to the first target condition threshold and the number of iterations is not equal to the first maximum number of iterations, modify the parameter conditions of the noise with the additional specific parameter conditions, and repeat steps A1)-F1).
12. The method for evaluating the noise of optical analog computing according to claim 11, characterized in that, The data distribution includes at least one of a standard Gaussian distribution, a Poisson distribution, and a chi-square distribution.
13. The method for evaluating the noise of optical analog computing according to claim 10, wherein When the target noise model is evaluated using heuristic learning, the evaluation steps include: Obtain the parameters encoded by the heuristic learning method; Based on the encoding method of the heuristic learning method, generate multiple initialization conditions for the encoding method; Optimize the noise model under each of the initialization conditions; Obtain the second target condition threshold and the second maximum number of iterations for exiting the evaluation; Evaluate each set of encoded parameters after optimization and calculate the second condition value for exiting the evaluation; Determine whether the second condition value for exiting is less than the second target condition threshold or the number of evaluations is equal to the second maximum number of iterations. If so, exit the iteration and complete the evaluation of the heuristic learning; if not, perform a new round of heuristic learning evaluation based on the parameters encoded by different heuristic learning methods.
14. The method for evaluating the noise of optical analog computing according to claim 13, wherein In the step of optimizing the noise model under each of the initialization conditions, the noise model includes a noise model generated according to expert experience and a model generated according to the reinforcement learning evaluation.
15. The method for evaluating the noise of optical analog computing according to claim 9, wherein The parameter information includes any one or a combination of at least any two of temperature, weight, and input value.
16. A noise simulation system, characterized in that, It includes: A conditional noise set acquisition module, configured to sample sample data signals from a device, and obtain a training noise set from the sample data signals. The training noise set is obtained by extracting a sample background signal from the sample data signals, and the condition is the sample generation condition of the sample data signals; A feature library construction module, configured to construct a noise feature library using the training noise set, and the noise feature library includes at least two types of noise features; A noise model construction module, configured to construct an initial noise model with parameterized processing based on preset expert experience, so that a noise generator outputs noise feature information corresponding to the at least two types of noise features according to the initial noise model; A noise model learning module, configured to train and learn the initial noise model using the noise feature library to obtain a trained target noise model.
17. A noise assessment system, characterized in that, It includes: A signal acquisition module, configured to acquire a sample signal in machine learning and calculate an effective signal in the sample signal according to an input and a weight; A signal processing module, configured to remove the effective signal from the sample signal to obtain a sample noise signal; A simulated noise signal generation module, configured to establish a noise model matching the sample noise signal and generate a simulated noise signal using the noise model. A noise model evaluation module, configured to compare the simulated noise signal with the sample noise signal to evaluate the noise model; Wherein, the noise model is obtained by using the noise simulation system described in claim 16.
18. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the method described in any one of claims 1 to 8 is implemented.
19. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the method described in any one of claims 9 to 15 is implemented.