WO2022057691A1 - Gaussian distribution data adjustment method based on improved gan - Google Patents
Gaussian distribution data adjustment method based on improved gan Download PDFInfo
- Publication number
- WO2022057691A1 WO2022057691A1 PCT/CN2021/117026 CN2021117026W WO2022057691A1 WO 2022057691 A1 WO2022057691 A1 WO 2022057691A1 CN 2021117026 W CN2021117026 W CN 2021117026W WO 2022057691 A1 WO2022057691 A1 WO 2022057691A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- network
- gaussian distribution
- loss
- generated
- Prior art date
Links
- 238000009826 distribution Methods 0.000 title claims abstract description 54
- 238000000034 method Methods 0.000 title claims abstract description 16
- 238000010606 normalization Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000005315 distribution function Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Definitions
- the invention relates to a Gaussian distribution data adjustment method based on an improved GAN network, and belongs to the technical field of machine learning.
- Generative Adversarial Networks (GAN, Generative Adversarial Networks) is a deep learning model and one of the most promising methods for unsupervised learning on complex distributions in recent years. The model produces fairly good outputs through mutual game learning of (at least) two modules in the framework: Generative Model and Discriminative Model.
- the technical problem to be solved by the present invention is to provide a Gaussian distribution data adjustment method based on an improved GAN network.
- the generated data output by the GAN network conforms to the Gaussian distribution law, and also satisfies that the characteristics of the generated data and the real data are consistent.
- the technical solution adopted in the present invention is: a Gaussian distribution data adjustment method based on an improved GAN network, comprising the following steps:
- the data that natural random number, artificial setting value or other systems are produced are transmitted to the data generation network, and the data generation network generates data according to the input data output;
- step S02 the generation data of step S01, the real data measuring the approximation of the generated data are transmitted to the data discrimination network, and the data discrimination network obtains the loss Loss1 according to the approximation of the generated data and the real data;
- the n data generated in batches by the data generation network are transmitted to the Gaussian distribution deviation judgment network, and the Gaussian distribution deviation judgment network uses the central limit theorem to count the distribution of the n data generated in batches, and calculate the distance between the distribution and the normal Gaussian distribution , so as to get the loss Loss2;
- the weighted sum of the loss Loss2 and the loss Loss1 as the overall loss of the data generation network, use the overall loss to train the image generation network, and the trained data generation network adjusts the input data in line with the Gaussian distribution.
- the image generating network, the image discriminating network, and the Gaussian distribution deviation judging network are sequentially trained or counted, and the other networks remain unchanged when one of them is trained or counted.
- step S03 the mean square error of the distribution of the n pieces of data generated in batches and the normal Gaussian distribution is calculated as the loss Loss2.
- Loss Loss2+ ⁇ Loss1, where ⁇ is an adjustable weighting coefficient.
- the present invention is used to improve the Gaussian distribution data adjustment of the GAN network, and the randomly generated data is distributed in a specified manner, no matter how the original real data is distributed.
- the GAN network only changes the distribution of the data, not the data itself, so the generated data is consistent with the real data in characteristics.
- the present invention is described by taking Gaussian distribution as an example, and other statistical distribution laws of data can also be used for directional normalization, so as to meet the expected requirements of data generation.
- Figure 1 is a flow chart of the method.
- This embodiment discloses a Gaussian distribution data adjustment method based on an improved GAN network, as shown in FIG. 1 , including the following steps:
- the data AZi is transmitted to the data generation network, after the data generation network carries out the corresponding calculation to the input data AZi, the output generation data;
- the data AZi represents a natural random number, and can also be manually set or data generated by other systems, such as data or images collected by a transmitter;
- step S02 the generation data of step S01, the real data BZi measuring the approximation of the generated data are transmitted to the data discrimination network, and the data discrimination network obtains the loss Loss1 according to the approximation of the generated data and the real data;
- the n data generated in batches by the data generation network are transmitted to the Gaussian distribution deviation judgment network, and the Gaussian distribution deviation judgment network uses the central limit theorem to count the distribution of the n data generated in batches, and calculate the distance between the distribution and the normal Gaussian distribution , so as to get the loss Loss2;
- the central limit theorem shows that when n is large, the sum of independent random variables approximately obeys the normal distribution N(n ⁇ , n ⁇ 2 ), therefore, the target normal distribution function is N(n ⁇ , n ⁇ 2 ), this method
- the key problem to be solved is to make the approximation of the target normal function as high as possible when n is certain.
- the weighted sum of the loss Loss2 and the loss Loss1 as the overall loss of the data generation network, use the overall loss to train the image generation network, and the trained data generation network adjusts the input data in line with the Gaussian distribution.
- the weighted summation of the loss Loss2 and the loss Loss1 is realized by the gating unit, specifically, the weighting coefficient ⁇ and the loss Loss1 are multiplied by the AND gate, and the loss Loss2 and ⁇ Loss1 are added by the OR gate.
- the image generating network, the image discriminating network, and the Gaussian distribution deviation judging network are sequentially trained or counted, and the other networks remain unchanged during one training or statistics.
- step S03 the distribution of the n pieces of data generated in batches and the normal Gaussian distribution are calculated as the mean square error as the loss Loss2.
- Loss1 is the output of the data discrimination network, that is, the characteristic distance between the generated data and the real data (such as Euclidean distance, cosine distance, etc.);
- the training is divided into two stages Execute, when training the discriminant network, assuming that the parameters of the generation network remain unchanged, generate batch data as negative samples, and real data as positive samples to train the network; when training the generation network, the discriminant network uses the trained model. The feature distance between the generated data and the real data is judged (ie Loss1).
- GAN network other aspects of the GAN network are the same as the general GAN network principle, and the image generation network, image discrimination network, and Gaussian distribution deviation judgment network can still be common VGG, RESNET, etc. or self-defined networks.
- This embodiment is used to improve the Gaussian distribution data adjustment of the GAN network, and the randomly generated data is distributed in a specified manner, no matter how the original real data is distributed.
- the GAN network only changes the distribution of the data, not the data itself, so the generated data is consistent with the real data in characteristics.
- the present invention is described by taking Gaussian distribution as an example, and other statistical distribution laws of data can also be used for directional normalization, so as to meet the expected requirements of data generation.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims (4)
- 一种基于改进GAN网络的高斯分布数据调整方法,其特征在于:包括以下步骤:A Gaussian distribution data adjustment method based on an improved GAN network, characterized in that it comprises the following steps:S01)、将自然随机数、人工设定值或者其他系统产生的数据传输至数据生成网络,数据生成网络根据输入数据输出生成数据;S01), the data generated by natural random numbers, artificial setting values or other systems are transmitted to the data generation network, and the data generation network outputs the generated data according to the input data;S02)、步骤S01的生成数据、衡量生成数据近似性的真实数据传输至数据判别网络,数据判别网络根据生成数据与真实数据的近似性得到损失Loss1;S02), the generation data of step S01, the real data measuring the approximation of the generated data are transmitted to the data discrimination network, and the data discrimination network obtains the loss Loss1 according to the approximation of the generated data and the real data;S03)、数据生成网络批量生成的n个数据传输至高斯分布偏差判断网络,高斯分布偏差判断网络使用中心极限定理统计批量生成的n个数据的分布,并计算该分布与正态高斯分布的距离,从而得出损失Loss2;S03), the n data generated in batches by the data generation network are transmitted to the Gaussian distribution deviation judgment network, and the Gaussian distribution deviation judgment network uses the central limit theorem to count the distribution of the n data generated in batches, and calculate the distance between the distribution and the normal Gaussian distribution , so as to get the loss Loss2;S04)、损失Loss2与损失Loss1加权求和,作为数据生成网络的整体损失,利用该整体损失对图像生成网络进行训练,训练后的数据生成网络对输入数据做符合高斯分布的调整。S04), the weighted sum of the loss Loss2 and the loss Loss1, as the overall loss of the data generation network, use the overall loss to train the image generation network, and the trained data generation network adjusts the input data in line with the Gaussian distribution.
- 根据权利要求1所述的基于改进GAN网络的高斯分布数据调整方法,其特征在于:图像生成网络、图像判别网络、高斯分布偏差判断网络依次进行训练或统计,其中一个训练或统计时,其他网络保持不变。The Gaussian distribution data adjustment method based on an improved GAN network according to claim 1, characterized in that: the image generation network, the image discrimination network, and the Gaussian distribution deviation judgment network are trained or counted in sequence, and when one is trained or counted, the other networks constant.
- 根据权利要求1所述的基于改进GAN网络的高斯分布数据调整方法,其特征在于:步骤S03中,将批量生成的n个数据的分布与正态高斯分布求均方差作为损失Loss2。The Gaussian distribution data adjustment method based on the improved GAN network according to claim 1, characterized in that: in step S03, the distribution of the n pieces of data generated in batches and the normal Gaussian distribution are calculated as the mean square error as the loss Loss2.
- 根据权利要求1所述的基于改进GAN网络的高斯分布数据调 整方法,其特征在于:数据生成网络的整体损失设为Loss,则Loss=Loss2+λLoss1,其中λ为可调的加权系数。The Gaussian distribution data adjustment method based on improved GAN network according to claim 1 is characterized in that: the overall loss of the data generation network is set to Loss, then Loss=Loss2+λLoss1, wherein λ is an adjustable weighting coefficient.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010985207.4A CN112116073A (en) | 2020-09-18 | 2020-09-18 | Gaussian distribution data adjusting method based on improved GAN network |
CN202010985207.4 | 2020-09-18 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022057691A1 true WO2022057691A1 (en) | 2022-03-24 |
Family
ID=73800066
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2021/117026 WO2022057691A1 (en) | 2020-09-18 | 2021-09-07 | Gaussian distribution data adjustment method based on improved gan |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN112116073A (en) |
WO (1) | WO2022057691A1 (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112116073A (en) * | 2020-09-18 | 2020-12-22 | 神思电子技术股份有限公司 | Gaussian distribution data adjusting method based on improved GAN network |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109146988A (en) * | 2018-06-27 | 2019-01-04 | 南京邮电大学 | Non-fully projection CT image rebuilding method based on VAEGAN |
CN109446339A (en) * | 2018-10-11 | 2019-03-08 | 广东工业大学 | A kind of knowledge mapping representation method based on multicore Gaussian Profile |
US20190378263A1 (en) * | 2018-06-08 | 2019-12-12 | Industrial Technology Research Institute | Industrial image inspection method and system and computer readable recording medium |
CN112116073A (en) * | 2020-09-18 | 2020-12-22 | 神思电子技术股份有限公司 | Gaussian distribution data adjusting method based on improved GAN network |
-
2020
- 2020-09-18 CN CN202010985207.4A patent/CN112116073A/en active Pending
-
2021
- 2021-09-07 WO PCT/CN2021/117026 patent/WO2022057691A1/en active Application Filing
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190378263A1 (en) * | 2018-06-08 | 2019-12-12 | Industrial Technology Research Institute | Industrial image inspection method and system and computer readable recording medium |
CN109146988A (en) * | 2018-06-27 | 2019-01-04 | 南京邮电大学 | Non-fully projection CT image rebuilding method based on VAEGAN |
CN109446339A (en) * | 2018-10-11 | 2019-03-08 | 广东工业大学 | A kind of knowledge mapping representation method based on multicore Gaussian Profile |
CN112116073A (en) * | 2020-09-18 | 2020-12-22 | 神思电子技术股份有限公司 | Gaussian distribution data adjusting method based on improved GAN network |
Also Published As
Publication number | Publication date |
---|---|
CN112116073A (en) | 2020-12-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111563275B (en) | Data desensitization method based on generation countermeasure network | |
WO2022057691A1 (en) | Gaussian distribution data adjustment method based on improved gan | |
CN108520272A (en) | A kind of semi-supervised intrusion detection method improving blue wolf algorithm | |
WO2023019601A1 (en) | Signal modulation recognition method for complex-valued neural network based on structure optimization algorithm | |
CN108052968B (en) | QSFLA-SVM perception intrusion detection method | |
CN110321451B (en) | Image retrieval algorithm based on distribution entropy gain loss function | |
CN111126602A (en) | Cyclic neural network model compression method based on convolution kernel similarity pruning | |
CN115131347B (en) | Intelligent control method for processing zinc alloy parts | |
CN111553348A (en) | Anchor-based target detection method based on centernet | |
CN110224862A (en) | Multi-agent system network appearance based on multilayer perceptron invades capability assessment method | |
CN108334900A (en) | Generation method and system, the sorting technique and system of the disaggregated model of power battery | |
CN113408648A (en) | Unit combination calculation method combined with deep learning | |
CN111027229B (en) | Wind power curve fitting method based on sparse heteroscedastic multi-spline regression | |
CN103824285B (en) | Image segmentation method based on bat optimal fuzzy clustering | |
WO2021258961A1 (en) | Network traffic classification method and system based on improved k-means algorithm | |
CN111079229B (en) | Large-scale high-speed rotation equipment unbalance dual-target optimization method | |
CN108171271A (en) | A kind of equipment deteriorates early warning method and system | |
CN110110673B (en) | Face recognition method based on bidirectional 2DPCA and cascade forward neural network | |
CN111652264A (en) | Negative migration sample screening method based on maximum mean difference | |
Wang et al. | Research on feature selection method of intrusion detection based on deep belief network | |
CN114662624A (en) | Non-intrusive load identification method and system based on retraining twin network | |
CN114004170A (en) | Reconstruction method of cascade flow field based on limited measuring points | |
CN112418317A (en) | Method for identifying and classifying precision machining structural part based on PSO-SVM | |
CN113205214A (en) | Photovoltaic power ultra-short-term prediction method based on parallel dendritic neural network | |
CN114238952A (en) | Abnormal behavior detection method, device and system and computer readable storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21868516 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21868516 Country of ref document: EP Kind code of ref document: A1 |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 20.09.2023) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21868516 Country of ref document: EP Kind code of ref document: A1 |