CN116911354A - Encoder neural network model construction method and data processing method - Google Patents
Encoder neural network model construction method and data processing method Download PDFInfo
- Publication number
- CN116911354A CN116911354A CN202311182240.3A CN202311182240A CN116911354A CN 116911354 A CN116911354 A CN 116911354A CN 202311182240 A CN202311182240 A CN 202311182240A CN 116911354 A CN116911354 A CN 116911354A
- Authority
- CN
- China
- Prior art keywords
- neural network
- result
- network model
- module
- encoder neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000003062 neural network model Methods 0.000 title claims abstract description 111
- 238000003672 processing method Methods 0.000 title claims abstract description 16
- 238000010276 construction Methods 0.000 title abstract description 11
- 230000004927 fusion Effects 0.000 claims abstract description 90
- 238000005259 measurement Methods 0.000 claims abstract description 75
- 238000000034 method Methods 0.000 claims abstract description 63
- 238000012549 training Methods 0.000 claims abstract description 24
- 238000013528 artificial neural network Methods 0.000 claims description 51
- 230000006870 function Effects 0.000 claims description 29
- 239000013598 vector Substances 0.000 claims description 29
- 238000012545 processing Methods 0.000 claims description 27
- 230000006835 compression Effects 0.000 claims description 20
- 238000007906 compression Methods 0.000 claims description 20
- 239000011159 matrix material Substances 0.000 claims description 20
- 238000010801 machine learning Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 6
- 230000004913 activation Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000004891 communication Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 238000003491 array Methods 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000000802 evaporation-induced self-assembly Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000001502 supplementing effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
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
- G06N3/0455—Auto-encoder networks; Encoder-decoder 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a construction method and a data processing method of an encoder neural network model, wherein a multi-source data sample is input into a hiding module through an input module of an initial encoder neural network model, and a fusion result is output; respectively inputting the fusion result into a prediction module and a measurement module, and outputting the prediction result and the measurement result; and determining a loss value based on the loss function, an actual tag value corresponding to the multi-source data sample, a prediction result and a measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model. According to the method, the initial encoder neural network model can be optimally trained based on the multi-source data samples, fusion of multi-source data is achieved, and the loss function comprehensively considers the prediction result and the measurement result because the measurement result can represent the difference between the fusion result and the multi-source data samples, so that the prediction precision of the encoder neural network model can be improved based on the loss function.
Description
Technical Field
The invention relates to the technical field of data processing, in particular to a construction method and a data processing method of an encoder neural network model.
Background
With the deep application of the big data machine learning method in a wider field, the faced data sources are more and more complex and various, which brings challenges to the construction of a big data machine learning model. The encoder theory is a novel calculation model which is focused in the fields of neural networks and artificial intelligence in recent years, and has the advantage of realizing denoising, dimension reduction and semantic association of data by means of encoder architecture. In traditional encoder machine learning, the type of a considered data source is generally single, and vectors generated by an encoder are often directly used for supervised prediction tasks, such as supervised classification, supervised prediction and the like, so that the generated vectors may not well reflect information of original input data, and the supervised prediction precision of a constructed encoder neural network is poor, and therefore, the optimization training of the encoder neural network based on multi-source data is difficult to realize in the related technology.
Disclosure of Invention
The invention aims to provide a construction method and a data processing method of an encoder neural network model, so as to realize the optimal training of the encoder neural network based on multi-source data.
The invention provides a method for constructing an encoder neural network model, which comprises the following steps: acquiring a multisource data sample and an initial encoder neural network model; the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module; inputting the multi-source data sample to a hiding module through an input module, and outputting a fusion result; inputting the fusion result into a prediction module, and outputting a prediction result; inputting the fusion result to a measurement module, and outputting a measurement result; the measurement result is used for representing the difference between the fusion result and the multi-source data sample; and determining a loss value based on a preset loss function, an actual tag value, a prediction result and a measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model.
Further, the hiding module comprises an encoding neural network layer and a decoding neural network layer; the multi-source data sample is input to the hiding module through the input module, and the step of outputting the fusion result comprises the following steps: inputting the multi-source data sample into the coding neural network layer through the input module so as to compress the multi-source data sample through the coding neural network layer to obtain a compression result; inputting the compression result into the decoding neural network layer, so as to decompress the compression result through the decoding neural network layer, and obtaining a fusion result.
Further, the multi-source data samples are represented in the form of a multi-source data matrix; the loss function is expressed as follows:
；
wherein y represents an actual tag value; y is pre Representing the prediction result;representing a binary norm operator; />For adjusting the factor superparameter, X represents a multi-source data matrix; />Representing a fusion result; />Representing the measurement result.
Further, the method comprises the steps of,the expression is as follows:
；
wherein ,representing a multisource data matrix->Is>A row vector; />Representing a fusion result; />Representation->And->A covariance matrix between the two; n represents a multisource data matrix->Is a vector number of rows of (a).
Further, the multi-source data samples include multi-scale data, multi-feature data, and/or multi-semantic data.
The invention provides a data processing method, which is applied to equipment configured with an encoder neural network model; the encoder neural network model is obtained by training any one of the methods; the method comprises the following steps: acquiring data to be processed; inputting data to be processed into an encoder neural network model, and outputting a data processing result; if the data to be processed is multi-source data, the data processing result is a result obtained by carrying out semantic association and fusion based on the multi-source data.
The invention provides a device for constructing an encoder neural network model, which comprises: the first acquisition module is used for acquiring a multi-source data sample and an initial encoder neural network model; the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module; the first output module is used for inputting the multi-source data sample to the hiding module through the input module and outputting a fusion result; the second output module is used for inputting the fusion result to the prediction module and outputting the prediction result; inputting the fusion result to a measurement module, and outputting a measurement result; the measurement result is used for representing the difference between the fusion result and the multi-source data sample; the training module is used for determining a loss value based on a preset loss function, an actual label value, a prediction result and a measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model.
The invention provides a data processing device, which is applied to equipment provided with an encoder neural network model; the encoder neural network model is obtained by training any one of the methods; the device comprises: the second acquisition module is used for acquiring data to be processed; the third output module is used for inputting the data to be processed into the encoder neural network model and outputting a data processing result; if the data to be processed is multi-source data, the data processing result is a result obtained by carrying out semantic association and fusion based on the multi-source data.
The electronic device provided by the invention comprises a processor and a memory, wherein the memory stores machine executable instructions capable of being executed by the processor, and the processor executes the machine executable instructions to realize the method for constructing the encoder neural network model or the data processing method.
The present invention provides a machine-readable storage medium storing machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement a method of constructing an encoder neural network model of any one of the above, or a data processing method of the above.
The invention provides a construction method and a data processing method of an encoder neural network model, which are used for acquiring a multi-source data sample and an initial encoder neural network model; the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module; inputting the multi-source data sample to a hiding module through an input module, and outputting a fusion result; inputting the fusion result into a prediction module, and outputting a prediction result; inputting the fusion result to a measurement module, and outputting a measurement result; the measurement result is used for representing the difference between the fusion result and the multi-source data sample; and determining a loss value based on a preset loss function, an actual tag value, a prediction result and a measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model. According to the method, the initial encoder neural network model can be optimally trained based on the multi-source data samples, fusion of multi-source data is achieved, and the loss function comprehensively considers the prediction result and the measurement result because the measurement result can represent the difference between the fusion result and the multi-source data samples, so that the prediction precision of the encoder neural network model can be improved based on the loss function.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the present invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flowchart of a method for constructing an encoder neural network model according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of an architecture of an encoder neural network model for a supervised prediction task according to an embodiment of the present invention;
FIG. 3 is a schematic structural diagram of a device for constructing an encoder neural network model according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a data processing apparatus according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
The technical solutions of the present invention will be clearly and completely described in connection with the embodiments, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
In recent years, the machine learning field oriented to big data is vigorously developed, and the great achievements of the machine learning field are widely applied to the fields of computer vision, natural language processing, knowledge patterns and the like. However, as the big data machine learning method is deeply applied in a wider field, the data sources facing the big data machine learning method are more and more complex and various, which brings challenges to the construction of the big data machine learning model. The challenges are embodied in that the data features are more and more complex, the data scale is more and more wide, and the multi-source heterogeneous characteristics of the data are more and more remarkable, for example, in the field of intelligent prediction of the aerodynamic profile characteristics of an automobile, the data used for intelligent prediction model training comprise multi-source data such as design data, simulation data and aerodynamic test data, and the multi-scale and multi-feature information such as time scale, space scale, time sequence characteristics and geometric characteristics is covered. How to fully couple the characteristics, the scale and the heterogeneous information of the data and effectively realize multi-source data fusion, thereby helping the construction of a machine learning algorithm and becoming a research hot spot in recent years.
Encoder theory is a novel calculation model which is focused in the fields of neural networks and artificial intelligence in recent years. The encoder is a special unsupervised neural network computing architecture, and the core architecture comprises an input layer network, a hidden layer network and an output layer network, and the encoding, decoding and re-outputting of the input information are realized through parameter transmission. The advantage of the encoder is that denoising, dimension reduction and semantic association of data are achieved by means of the encoder architecture. However, the related art lacks a method for realizing the optimized training of the encoder neural network based on the multi-source data, and based on the method, the embodiment of the invention provides a method for constructing the encoder neural network model and a data processing method, and the technology can be applied to a scene requiring fusion application of the multi-source data.
For the sake of understanding the present embodiment, first, a method for constructing an encoder neural network model according to an embodiment of the present invention is described, as shown in fig. 1, where the method includes the following steps:
step S102, a multisource data sample and an initial encoder neural network model are obtained; the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module.
The multi-source data sample comprises multi-scale data, multi-feature data and/or multi-semantic data; the multi-scale, multi-feature and multi-semantic are descriptions of the data, and the multi-scale refers to the range of the data in the dimensions of space, time and the like, such as the spatial range of the description of the data, the time sequence length of the data and the like; the multi-feature refers to the attribute of the description data characteristics contained in the data; multiple semantics refer to embedded semantic associations between different features in the data. The initial encoder neural network model may include an input module, a hiding module, and an output module, where the output module may specifically be composed of a prediction module and a measurement module; the above-mentioned actual tag value may be understood as a real result under an actual working condition determined based on a multi-source data sample in an application scenario, for example, taking the field of intelligent prediction of aerodynamic profile characteristics of an automobile as an example, the input multi-source data sample includes: calculating working conditions, constraint conditions, geometric characteristics of objects and the like, wherein the actual label value is the true value of aerodynamic force under the actual working conditions, and the aerodynamic force is the force acting on the automobile when the automobile and the air relatively move.
Step S104, the multi-source data sample is input to the hiding module through the input module, and a fusion result is output.
The hiding module can be used for fusing and interacting semantic information embedded in the multi-source data sample to realize semantic association of each input data in the multi-source data sample; in actual implementation, the multi-source data sample can be input into an input module of the initial encoder neural network model, and the multi-source data sample is input into a hiding module through the input module so as to output a fusion result through the hiding module; continuing with the field of intelligent prediction of the aerodynamic shape characteristics of the automobile as an example, semantic association, fusion and the like can be carried out on the input multi-source data samples, and fusion vectors can be output.
Step S106, inputting the fusion result into a prediction module, and outputting a prediction result; inputting the fusion result to a measurement module, and outputting a measurement result; the measurement result is used for representing the difference between the fusion result and the multi-source data sample.
The prediction module can be used for task prediction; the measurement module can be used for measuring information difference between the fusion result and the multi-source data sample; in actual implementation, the fusion result output by the hiding module can be respectively input to the prediction module and the measurement module, and the prediction result is output through the prediction module; outputting a measurement result through a measurement module; continuing taking the intelligent prediction field of the aerodynamic profile characteristics of the automobile as an example, the output fusion vector can be input into a prediction module, the prediction value of aerodynamic force is output, the fusion vector is input into a measurement module, the measurement result is output, the measurement result can represent the similarity degree between the fusion vector and the input multi-source data sample, and the difference can be judged according to the similarity degree.
And S108, determining a loss value based on a preset loss function, an actual label value, a prediction result and a measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model.
The method for constructing the encoder neural network model comprises the steps of obtaining a multi-source data sample and an initial encoder neural network model; the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module; inputting the multi-source data sample to a hiding module through an input module, and outputting a fusion result; inputting the fusion result into a prediction module, and outputting a prediction result; inputting the fusion result to a measurement module, and outputting a measurement result; the measurement result is used for representing the difference between the fusion result and the multi-source data sample; and determining a loss value based on a preset loss function, an actual tag value, a prediction result and a measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model. According to the method, the initial encoder neural network model can be optimally trained based on the multi-source data samples, fusion of multi-source data is achieved, and the loss function comprehensively considers the prediction result and the measurement result because the measurement result can represent the difference between the fusion result and the multi-source data samples, so that the prediction precision of the encoder neural network model can be improved based on the loss function.
The embodiment of the invention also provides another method for constructing the encoder neural network model, which is realized on the basis of the method of the embodiment, wherein the hiding module comprises an encoding neural network layer and a decoding neural network layer; the multi-source data samples are represented in the form of a multi-source data matrix; the method comprises the following steps:
step one, acquiring a multisource data sample and an initial encoder neural network model; the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module;
in actual implementation, it is possible to defineData of species origin, denoted respectively +.>，/>，Wherein->The data comprises->The number of features may be different for all data.
For the followingThe present disclosure optionally constructs data moments in such a way that the data are listed sequentiallyThe array aligns data by supplementing 0 elements at the end of the data with insufficient length. Alternatively, assume dataCharacteristic number of->Maximum, then data matrix->Can be expressed as
，
wherein ,is>Behavioral data->Filling the vector after 0 element to guarantee length and +.>The vector length remains consistent.
Therefore, the integration of multi-scale, multi-feature and multi-semantic data is completed, and the collection of multi-source data is realized.
And secondly, inputting the multi-source data sample into the coding neural network layer through the input module so as to compress the multi-source data sample through the coding neural network layer, thereby obtaining a compression result.
The input of the input module is multisource data matrix of multisource collection. The input of the coding neural network layer is a multi-source data matrix received by an input module>BraidingThe output of the code neural network layer is the input of the decoding neural network layer, and the coding neural network layer is used for receiving the multisource data matrix +.>Compression is performed to compress the longer-dimension input vector into a low-dimension encoded vector with highly condensed information.
And thirdly, inputting the compression result into a decoding neural network layer to decompress the compression result through the decoding neural network layer, so as to obtain a fusion result.
The output of the decoding neural network layer is then the input of the output layer (including the prediction module and the metrics module). And the decoding neural network layer decompresses the compression result and reconverts the compression result into the coding vector with sparse and scattered information and longer dimension. Semantic information in the input multi-source data samples is collected, interacted and recombined in the neural network neurons through compression operation of the coding neural network layer and decoding operation of the decoding neural network layer, so that semantic association in the input multi-source data samples is realized.
For convenience of explanation, the modules are hiddenfRepresentation, coding neural network layerf 1 Representation, decoding neural network layerf 2 Denoted as example, coding neural network layerf 1 Output compressed result to output vectorExpressed as an example +.>The computational process of (1) is expressed as
；
wherein ,the activation function is represented as a function of the activation,W 1 、b 1 coding neural network layer->Is a parameter of (a).
Decoding neural network layersOutput fusion result to output vector +.>Expressed as an example +.>The computational process of (1) is expressed as
；
wherein ,the activation function is represented as a function of the activation,W 2 、b 2 coding neural network layer->Is a parameter of (a).
(Vector)Multisource data matrix which is multisource aggregate>And the output value after passing through a hiding module of the encoder neural network model. During the calculation, data of different sources, scales, features, i.e. data +.>Is fused by means of a non-linear encoding-decoding calculation of the encoder neural network and is represented by a vector +.>And outputting.
Inputting the fusion result to a prediction module, and outputting the prediction result; inputting the fusion result to a measurement module, and outputting a measurement result; the measurement result is used for representing the difference between the fusion result and the multi-source data sample.
Assume an input multi-source data matrixThe actual label is->. The output layer of the present disclosure includes two parts: prediction Module->And metric Module->。
Specifically, the prediction moduleFor carrying out prediction tasks, optionally prediction module->Can be constructed by using a multi-layer neural network, prediction module +.>Output of +.>The calculation process of (2) can be expressed as:
；
wherein ,representing an activation function->And->For prediction module/>Is a parameter of (a).
Metrology moduleFor measuring multisource data matrix->Fusion result->Information differences of (3). Due to the original multisource data matrix->Fusion result->Is different in vector size dimensions, optionally by means of the Mahalanobis distance operator>By computing the original multisource data matrix->Fusion result->Similarity between the two information differences are judged, and the measurement module is +.>The scale after the data change can be ensured to meet the requirement, and the scale can be understood as a specific numerical value, namely the fluctuation of the data is within an allowable range.
The measurement resultCan be expressed as follows:
；
wherein ,representing a multisource data matrix->Is>A row vector; />Representing a fusion result; />Representation->And->A covariance matrix between the two; n represents a multisource data matrix->Is a vector number of rows of (a).
In conventional machine learning of an encoder, a vector generated by the encoder is often directly used for a supervised prediction task, such as supervised classification, supervised prediction, and the like, but a constraint on the generated vector is absent, so that the generated vector may not well reflect information of original input data. In order to solve the above problems, the present disclosure considers the coding effect of the encoder while considering the actual prediction effect, i.e. the information amount contained in the fusion vector generated by the encoder is as close to the original multi-source data matrix as possibleIs used for the information amount of the (a).
And fifthly, determining a loss value based on a preset loss function, an actual label value, a prediction result and a measurement result, and training an initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model.
Based on the above prediction moduleAnd metric Module->The above-mentioned loss function can be expressed as follows:
；
wherein ,yrepresenting the actual tag value;y pre representing the prediction result;representing a binary norm operator; />For adjusting the factor superparameter, a Mahalanobis distance operator is used for controlling>Term pair loss function->And has a value between 0 and 1. X represents a multi-source data matrix; />Representing a fusion result; />Representing the measurement result.
For easy understanding, referring to the schematic architecture diagram of an encoder neural network model facing the task of supervised prediction shown in fig. 2, a multi-source data matrix X composed of multi-source data samples is sent to a hidden layer (corresponding to the above-mentioned hidden module) through an input layer (corresponding to the above-mentioned input module), where the hidden module includes an encoding neural network layerf 1 Decoding neural network layerf 2 Composition by coding neural network layerf 1 Encoding compression of multi-source data matrixAnd obtaining a compression resultBy decoding neural network layersf 2 For the compression result->Decoding to obtain fusion result->The method comprises the steps of carrying out a first treatment on the surface of the The output layer comprises a prediction layer (corresponding to the prediction module) and a measurement layer (corresponding to the measurement module), and the fusion result is +.>Input to the prediction layer to obtain a supervision predicted valuey pre (corresponding to the above predicted results), fusion result +.>Input to the measurement layer to obtain information measurement value(corresponding to the measurement result), assuming that the actual tag value corresponding to the multi-source data matrix X is y, the loss function isThe method comprises the steps of carrying out a first treatment on the surface of the According to the loss function, a loss value can be determined, and the initial encoder neural network model is trained based on the loss value until the loss value converges, so that the encoder neural network model is obtained.
According to the method for constructing the encoder neural network model, based on the collected multi-source data, the encoder neural network is constructed, and the compression and decoding operations of the encoder are utilized to fuse and interact semantic information embedded in the data, so that semantic association in input data is realized. The method and the device are oriented to the supervision and prediction task, and realize the effects of ensuring the supervision and prediction precision and ensuring the fusion vector generated by the encoder to be close to the information quantity of the original data by constructing a loss function training framework comprehensively considering the coding effect and the prediction effect.
The method for constructing the encoder neural network model comprises the following steps: integrating multi-scale, multi-feature and multi-semantic data to realize the collection of multi-source data; constructing an encoder neural network model based on the collected data to realize fusion of multi-source data; and constructing a cross-parameter-scale loss function training architecture for the supervision prediction task to realize the optimization training of the encoder neural network model. The method solves the problems of fusion of multi-feature, multi-scale and multi-source heterogeneous data and supervised learning prediction calculation for model training. The advantages of the encoder are fully utilized, and fusion of multi-feature, multi-scale and multi-source heterogeneous data is realized through the encoder, so that semantic features among the data are effectively extracted, and the supervised learning task is assisted. The semantic association and data fusion of multi-source data are realized from the viewpoint of constructing an encoder neural network model, and the construction of a loss function is realized from the viewpoint of comprehensively considering the double effects of coding and prediction.
The embodiment of the invention provides a data processing method, which is applied to equipment provided with an encoder neural network model; the encoder neural network model is obtained by training any one of the methods; the method comprises the following steps:
and step A, obtaining data to be processed.
Step B, inputting the data to be processed into an encoder neural network model, and outputting a data processing result; if the data to be processed is multi-source data, the data processing result is a result obtained by carrying out semantic association and fusion based on the multi-source data.
The data to be processed can be single-source data or multi-source data, and can be specifically selected according to actual requirements, for example, in the field of intelligent prediction of aerodynamic appearance characteristics of an automobile, the aerodynamic force of the automobile needs to be predicted based on various data, and the data to be processed is specifically multi-source data, for example, length, width, height, appearance, radian angle, working condition (pitch angle, translational angle and the like) of the automobile, time-varying simulation data and the like, and the predicted aerodynamic force value of the automobile can be output through the encoder neural network model, and the aerodynamic force value is output after semantic association and fusion are carried out on each data in the data to be processed, namely, the aerodynamic force value is a value obtained by comprehensively considering the multi-source data in the data to be processed.
The data processing method comprises the steps of inputting the acquired data to be processed into an encoder neural network model, and outputting a data processing result; if the data to be processed is multi-source data, the data processing result is a result obtained by carrying out semantic association and fusion based on the multi-source data. According to the method, semantic information embedded in data to be processed is fused and interacted through the encoder neural network model, so that semantic association in input data is realized, and a data processing result which is more fit with reality can be obtained.
The embodiment of the invention provides a device for constructing an encoder neural network model, as shown in fig. 3, the device comprises: a first acquisition module 30 for acquiring a multi-source data sample and an initial encoder neural network model; the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module; the first output module 31 is configured to input the multi-source data sample to the hiding module through the input module, and output a fusion result; a second output module 32, configured to input the fusion result to the prediction module, and output a prediction result; inputting the fusion result to a measurement module, and outputting a measurement result; the measurement result is used for representing the difference between the fusion result and the multi-source data sample; the training module 33 is configured to determine a loss value based on a preset loss function, an actual tag value, a prediction result, and a measurement result, and train the initial encoder neural network model based on the loss value until the loss value converges, so as to obtain the encoder neural network model.
According to the construction device of the encoder neural network model, the initial encoder neural network model can be optimally trained based on the multi-source data samples, fusion of multi-source data is achieved, and the prediction result and the measurement result are comprehensively considered by the loss function because the measurement result can represent the difference between the fusion result and the multi-source data samples, so that the prediction precision of the encoder neural network model can be improved based on the loss function.
Further, the first output module 31 is further configured to: inputting the multi-source data sample into the coding neural network layer through the input module so as to compress the multi-source data sample through the coding neural network layer to obtain a compression result; inputting the compression result into the decoding neural network layer, so as to decompress the compression result through the decoding neural network layer, and obtaining a fusion result.
Further, the multi-source data samples are represented in the form of a multi-source data matrix; the loss function is expressed as follows:
；
wherein y represents an actual tag value; y is pre Representing the prediction result;representing a binary norm operator; />For adjusting the factor superparameter, X represents a multi-source data matrix; />Representing a fusion result; />Representing the measurement result.
Further, the method comprises the steps of,the expression is as follows:
；
wherein ,representing a multisource data matrix->Is>A row vector; />Representing a fusion result; />Representation->And->A covariance matrix between the two; n represents a multisource data matrix->Is a vector number of rows of (a).
Further, the multi-source data samples include multi-scale data, multi-feature data, and/or multi-semantic data.
The implementation principle and the generated technical effects of the device for constructing the encoder neural network model provided by the embodiment of the invention are the same as those of the embodiment of the method for constructing the encoder neural network model, and for the sake of brief description, reference may be made to corresponding contents in the embodiment of the method for constructing the encoder neural network model.
The embodiment of the invention provides a data processing device, which is arranged on equipment provided with an encoder neural network model; the encoder neural network model is obtained by training the method; as shown in fig. 4, the apparatus includes: a second acquiring module 40, configured to acquire data to be processed; a third output module 41, configured to input data to be processed into the encoder neural network model, and output a data processing result; if the data to be processed is multi-source data, the data processing result is a result obtained by carrying out semantic association and fusion based on the multi-source data.
According to the data processing device, the semantic information embedded in the data to be processed is fused and interacted through the encoder neural network model, so that semantic association in the input data is realized, and a data processing result which is more fit with reality can be obtained.
The embodiment of the present invention further provides an electronic device, as shown in fig. 5, where the electronic device includes a processor 130 and a memory 131, where the memory 131 stores machine executable instructions that can be executed by the processor 130, and the processor 130 executes the machine executable instructions to implement the method for constructing the encoder neural network model, or the data processing method.
Further, the electronic device shown in fig. 5 further includes a bus 132 and a communication interface 133, and the processor 130, the communication interface 133, and the memory 131 are connected through the bus 132.
The memory 131 may include a high-speed random access memory (RAM, random Access Memory), and may further include a non-volatile memory (non-volatile memory), such as at least one magnetic disk memory. The communication connection between the system network element and at least one other network element is implemented via at least one communication interface 133 (which may be wired or wireless), and may use the internet, a wide area network, a local network, a metropolitan area network, etc. Bus 132 may be an ISA bus, a PCI bus, an EISA bus, or the like. The buses may be classified as address buses, data buses, control buses, etc. For ease of illustration, only one bi-directional arrow is shown in FIG. 5, but not only one bus or type of bus.
The processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above method may be performed by integrated logic circuitry in hardware or instructions in software in processor 130. The processor 130 may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU for short), a network processor (Network Processor, NP for short), etc.; but also digital signal processors (Digital Signal Processor, DSP for short), application specific integrated circuits (Application Specific Integrated Circuit, ASIC for short), field-programmable gate arrays (Field-Programmable Gate Array, FPGA for short) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The disclosed methods, steps, and logic blocks in the embodiments of the present invention may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present invention may be embodied directly in the execution of a hardware decoding processor, or in the execution of a combination of hardware and software modules in a decoding processor. The software modules may be located in a random access memory, flash memory, read only memory, programmable read only memory, or electrically erasable programmable memory, registers, etc. as well known in the art. The storage medium is located in the memory 131, and the processor 130 reads the information in the memory 131, and in combination with its hardware, performs the steps of the method of the foregoing embodiment.
The embodiment of the invention also provides a machine-readable storage medium, which stores machine-executable instructions that, when being called and executed by a processor, cause the processor to implement the method for constructing the encoder neural network model or the data processing method, and the specific implementation can be referred to the method embodiment and will not be described herein.
The computer program product of the method for constructing the encoder neural network model and the method for processing data provided by the embodiments of the present invention includes a computer readable storage medium storing program codes, and the instructions included in the program codes may be used to execute the method described in the foregoing method embodiments, and specific implementation may refer to the method embodiments and will not be repeated herein.
The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.
Claims (10)
1. A method for constructing an encoder neural network model, the method comprising:
acquiring a multisource data sample and an initial encoder neural network model; wherein the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module;
inputting the multi-source data sample to the hiding module through the input module, and outputting a fusion result;
inputting the fusion result to the prediction module, and outputting a prediction result; inputting the fusion result to the measurement module, and outputting a measurement result; wherein the measurement result is used for representing the difference between the fusion result and the multi-source data sample;
and determining a loss value based on a preset loss function, the actual label value, the prediction result and the measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model.
2. The method of claim 1, wherein the concealment module comprises an encoding neural network layer and a decoding neural network layer; the multi-source data sample is input to the hiding module through the input module, and the step of outputting a fusion result comprises the following steps:
inputting the multi-source data sample to the coding neural network layer through the input module so as to compress the multi-source data sample through the coding neural network layer to obtain a compression result;
and inputting the compression result to the decoding neural network layer so as to decompress the compression result through the decoding neural network layer to obtain the fusion result.
3. The method of claim 1, wherein the multi-source data samples are represented in the form of a multi-source data matrix; the loss function is expressed as follows:
；
wherein y represents an actual tag value; y is pre Representing the prediction result;representing a binary norm operator; />For adjusting a factor superparameter, X represents the multi-source data matrix; />Representing the fusion result; />Representing the measurement result.
4. The method of claim 1, wherein theThe expression is as follows:
；
wherein ,representing a multisource data matrix->Is>A row vector; />Representing the fusion result; />Representation->And->A covariance matrix between the two; n represents a multisource data matrix->Is a vector number of rows of (a).
5. The method according to claim 1, wherein the multi-source data samples comprise multi-scale data, multi-feature data and/or multi-semantic data.
6. A data processing method, characterized in that the method is applied to a device configured with an encoder neural network model; the encoder neural network model is an encoder neural network model trained by the method according to any one of claims 1-5; the method comprises the following steps:
acquiring data to be processed;
inputting the data to be processed into the encoder neural network model, and outputting a data processing result; and if the data to be processed is multi-source data, the data processing result is a result obtained by carrying out semantic association and fusion on the multi-source data.
7. An apparatus for constructing an encoder neural network model, the apparatus comprising:
the first acquisition module is used for acquiring a multi-source data sample and an initial encoder neural network model; wherein the multi-source data sample has a corresponding actual tag value; the initial encoder neural network model includes: the system comprises an input module, a hiding module, a prediction module and a measurement module;
the first output module is used for inputting the multi-source data sample to the hiding module through the input module and outputting a fusion result;
the second output module is used for inputting the fusion result to the prediction module and outputting a prediction result; inputting the fusion result to the measurement module, and outputting a measurement result; wherein the measurement result is used for representing the difference between the fusion result and the multi-source data sample;
and the training module is used for determining a loss value based on a preset loss function, the actual label value, the prediction result and the measurement result, and training the initial encoder neural network model based on the loss value until the loss value converges to obtain the encoder neural network model.
8. A data processing apparatus, characterized in that the method is applied to a device configured with an encoder neural network model; the encoder neural network model is an encoder neural network model trained by the method according to any one of claims 1-5; the device comprises:
the second acquisition module is used for acquiring data to be processed;
the third output module is used for inputting the data to be processed into the encoder neural network model and outputting a data processing result; and if the data to be processed is multi-source data, the data processing result is a result obtained by carrying out semantic association and fusion on the multi-source data.
9. An electronic device comprising a processor and a memory, the memory storing machine executable instructions executable by the processor to implement the method of constructing an encoder neural network model of any one of claims 1-5 or the method of data processing of claim 6.
10. A machine-readable storage medium storing machine-executable instructions which, when invoked and executed by a processor, cause the processor to implement the method of constructing an encoder neural network model of any one of claims 1 to 5 or the data processing method of claim 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311182240.3A CN116911354A (en) | 2023-09-14 | 2023-09-14 | Encoder neural network model construction method and data processing method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311182240.3A CN116911354A (en) | 2023-09-14 | 2023-09-14 | Encoder neural network model construction method and data processing method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116911354A true CN116911354A (en) | 2023-10-20 |
Family
ID=88353513
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311182240.3A Pending CN116911354A (en) | 2023-09-14 | 2023-09-14 | Encoder neural network model construction method and data processing method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116911354A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113807447A (en) * | 2021-09-23 | 2021-12-17 | 兰州理工大学 | Multi-source heterogeneous data fusion method based on FC-SAE |
| CN114399064A (en) * | 2021-12-31 | 2022-04-26 | 淮阴工学院 | Equipment health index construction method based on multi-source sensor data fusion |
| CN115359313A (en) * | 2022-08-11 | 2022-11-18 | 中国人民解放军国防科技大学 | Multi-mode semi-supervised characterization learning method of hypergraph variation self-encoder |
| CN115376638A (en) * | 2022-08-22 | 2022-11-22 | 中国电子科技集团公司第二十八研究所 | Physiological characteristic data analysis method based on multi-source health perception data fusion |
| CN115512144A (en) * | 2022-08-31 | 2022-12-23 | 电子科技大学 | Automatic XRF spectrogram classification method based on convolution self-encoder |
-
2023
- 2023-09-14 CN CN202311182240.3A patent/CN116911354A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113807447A (en) * | 2021-09-23 | 2021-12-17 | 兰州理工大学 | Multi-source heterogeneous data fusion method based on FC-SAE |
| CN114399064A (en) * | 2021-12-31 | 2022-04-26 | 淮阴工学院 | Equipment health index construction method based on multi-source sensor data fusion |
| CN115359313A (en) * | 2022-08-11 | 2022-11-18 | 中国人民解放军国防科技大学 | Multi-mode semi-supervised characterization learning method of hypergraph variation self-encoder |
| CN115376638A (en) * | 2022-08-22 | 2022-11-22 | 中国电子科技集团公司第二十八研究所 | Physiological characteristic data analysis method based on multi-source health perception data fusion |
| CN115512144A (en) * | 2022-08-31 | 2022-12-23 | 电子科技大学 | Automatic XRF spectrogram classification method based on convolution self-encoder |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yu et al. | Pointr: Diverse point cloud completion with geometry-aware transformers | |
| Zhang et al. | Joint task-recursive learning for semantic segmentation and depth estimation | |
| US20210224640A1 (en) | Neural network circuit device, neural network processingmethod, and neural network execution program | |
| US20190044535A1 (en) | Systems and methods for compressing parameters of learned parameter systems | |
| US20240020541A1 (en) | Model training method and apparatus | |
| US20240135174A1 (en) | Data processing method, and neural network model training method and apparatus | |
| CN116151845A (en) | Product full life cycle management system and method based on industrial Internet of things technology | |
| CN110781686A (en) | Statement similarity calculation method and device and computer equipment | |
| CN111241271B (en) | Text emotion classification method and device and electronic equipment | |
| CN112712079A (en) | Character recognition method and device based on progressive coding and electronic equipment | |
| CN110688897A (en) | Pedestrian re-identification method and device based on joint judgment and generation learning | |
| Han et al. | L-Net: lightweight and fast object detector-based ShuffleNetV2 | |
| CN114998583A (en) | Image processing method, image processing apparatus, device, and storage medium | |
| CN112364166A (en) | Method for establishing relation extraction model and relation extraction method | |
| Lehtimäki et al. | Accelerating neural odes using model order reduction | |
| CN114692624A (en) | Information extraction method and device based on multitask migration and electronic equipment | |
| CN116911354A (en) | Encoder neural network model construction method and data processing method | |
| Ge et al. | Fine-tuning vision foundation model for crack segmentation in civil infrastructures | |
| CN117274689A (en) | Detection method and system for detecting defects of packaging box | |
| CN116467513A (en) | Attention mechanism-based multi-mode knowledge graph recommendation method, device and medium | |
| CN115499635A (en) | Data compression processing method and device | |
| Patrick et al. | Max-pooled fast learning Gabor capsule network | |
| CN110493088B (en) | Mobile internet traffic classification method based on URL | |
| CN114299204A (en) | Three-dimensional cartoon character model generation method and device | |
| CN113343807A (en) | Target detection method and device for complex scene under reconstruction guidance |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20231020 |
|
| RJ01 | Rejection of invention patent application after publication |