WO2023231753A1 - Neural network training method, data processing method, and device - Google Patents

Abstract

A neural network training method, a data processing method, and a device. In the method, artificial intelligence technology can be used for decision making. The method comprises: according to feature information of a first training sample, generating, by means of a feature acquisition network, first feature information and second feature information which correspond to the first training sample, wherein the first feature information comprises a feature of information, associated with a target attribute, in the first training sample; and training the feature acquisition network, wherein a first loss function item indicates the similarity between predicted category information and expected category information, the expected category information indicates a correct category for the information, associated with the target attribute, in the first training sample, and the purpose of performing training by using a second loss function item comprises reducing the similarity between the first feature information and the second feature information. Features of information associated with and not associated with a target attribute, which leads to a deviation, are respectively extracted, so as to improve the accuracy of finally-obtained prediction decision information.

Description

A neural network training method, data processing method and equipment

This application claims the priority of the Chinese patent application submitted to the China Patent Office on May 31, 2022, with the application number 202210613415.0 and the invention title "A neural network training method, data processing method and equipment", and its entire content incorporated herein by reference.

Technical field

This application relates to the field of artificial intelligence, and in particular to a neural network training method, data processing method and equipment.

Background technique

With the development of the Artificial Intelligence (AI) era, neural networks are increasingly used to help people make decisions. As an example, for example, in a news recommendation system, a neural network can be used to determine whether to recommend specific news; as another example, for example, in an education recommendation system, a neural network can be used to determine whether to recommend a specific course; as another example, for example, Neural networks are used to determine whether to grant a loan to a specific applicant, and so on.

However, because the training data we collect may have biases, the predicted decision-making information output by the trained neural network during the inference phase will also be biased, or even amplify this bias. As an example, in an education recommendation system, since users tend to search for courses based on the name of the supplier, the collected training data contains more feedback data for courses provided by suppliers targeting the general public, and more feedback data for courses provided by niche suppliers. There is less feedback data on courses provided by suppliers. The neural network trained based on such data will consider the course suppliers when determining the recommended courses. That is, the neural network learns the characteristics of the training data during the training process. deviation. In the inference stage, the probability of courses provided by public suppliers being recommended will be greatly increased, but the quality of the courses provided by the specific suppliers mentioned above is not necessarily good, that is, the predictive decision-making information output by the neural network in the inference stage will also be biased. .

Based on the above description, it can be seen that the predictive decision-making information output by the neural network currently used for decision-making is biased, and a solution that can improve the predictive decision-making information output by the neural network urgently needs to be introduced.

Contents of the invention

Embodiments of the present application provide a neural network training method, a data processing method and a device to respectively obtain the characteristics of the information associated with the target attribute in the input data, and the characteristics of the information not associated with the target attribute in the input data, In this way, the user can determine the target attribute based on the reasons that cause the deviation of the prediction decision information in the current task, and extract the characteristics of the information associated with the target attribute that causes the deviation and the characteristics of the information not associated with the target attribute, which is beneficial to reducing the The difficulty of the subsequent decision-making process is conducive to improving the accuracy of the final predictive decision-making information.

In order to solve the above technical problems, the embodiments of this application provide the following technical solutions:

In the first aspect, embodiments of the present application provide a neural network training method that can use artificial intelligence technology to make decisions. The method includes: the training device inputs the first training sample into a feature acquisition network, and uses the feature acquisition network to perform the first training Feature extraction is performed on the sample to obtain the feature information of the first training sample; according to the feature information of the first training sample, the first feature information and the second feature information corresponding to the first training sample are generated through the feature acquisition network; wherein, with the first The first characteristic information corresponding to the training sample includes the characteristics of the information associated with the target attribute in the first training sample; and the first characteristic information The second feature information corresponding to the training sample includes features of information in the first training sample that are not associated with each target attribute, that is, the training target includes the obtained first feature information and the second feature information includes different information. Further, technicians can determine target attributes based on factors in the current task that cause bias in predictive decision information. As an example, for example, if a trained neural network is used to determine whether to recommend a certain course, and factors that lead to bias in the prediction decision information include the supplier providing the course, then at least one target attribute used in the training phase of the neural network may include Supplier, if the first training sample is a course, the information associated with the target attribute in the first training sample may include a watermark indicating the supplier of the course, an introduction to the supplier in the course cover, or other information; as another For example, if a trained neural network is used to determine whether a face in an image has curly hair, and the factors that lead to bias in the prediction decision information include the gender of the person in the image, then at least one goal is used in the training phase of the neural network. The attributes may include the gender of the user. The information associated with the target attribute in the face image (i.e., the first training sample) may include image information of the neck part in the face image. The image information of the neck part may be used to determine whether the user has Adam's apple. wait.

The training device performs a classification operation according to the first feature information corresponding to the first training sample to obtain prediction category information. The prediction category information indicates the prediction category of the first feature information corresponding to the first training sample, that is, the first prediction category information indicates The prediction category corresponding to the information associated with the target attribute in the first training sample, the prediction category is included in multiple categories corresponding to the target attribute; according to the first loss function, the feature acquisition network is trained to obtain the feature acquisition after training Network; as an example, if a neural network is used to determine whether a certain course is recommended, and the target attribute is a supplier, then the multiple categories corresponding to the target attribute can include supplier A, supplier B, supplier C, and supplier D. etc.; as another example, if a neural network is used to determine whether the person in the image has curly hair, and the target attribute is gender, then the multiple categories corresponding to the target attribute may include male and female.

Wherein, the first loss function includes a first loss function term and a second loss function term. The first loss function term indicates the similarity between the predicted category information and the expected category information. The expected category information indicates the similarity between the first training sample and the target attribute. Correct category of associated information, the purpose of using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information; further, the second loss function term can directly calculate the first feature information and The similarity between the second feature information, or the second loss function term can also be the similarity between other information.

In this implementation, the feature information of the first training sample is decomposed through the feature acquisition network to obtain the first feature information and the second feature information corresponding to the first training sample. The first feature information corresponding to the first training sample includes Characteristics of the information associated with the target attribute in the first training sample; performing a classification operation according to the first feature information corresponding to the first training sample to obtain predicted category information, where the predicted category information indicates the first feature information corresponding to the first training sample The prediction category is included in multiple categories corresponding to the target attribute; according to the first loss function, the feature acquisition network is trained to obtain the trained feature acquisition network; the first loss function includes the first loss function term and The second loss function term, the first loss function term indicates the similarity between the predicted category information and the expected category information, the expected category information indicates the correct category of the information associated with the target attribute in the first training sample, the second loss function term is used The purpose of training includes reducing the similarity between the first feature information and the second feature information. Through the above solution, the trained feature acquisition network can respectively obtain the characteristics of the information associated with the target attribute in the input data, and the characteristics of the information not associated with the target attribute in the input data, so that the user can make predictions based on the current task. Determine the target attribute based on the reasons for the deviation of the decision-making information, and extract the characteristics of the information associated with the target attribute that caused the deviation and the characteristics of the information not associated with the target attribute, which will help reduce the difficulty of the subsequent decision-making process and improve the quality of the decision-making process. most The accuracy of the final forecast decision-making information.

In a possible implementation of the first aspect, the method further includes: the training device combines the first feature information and the second feature information corresponding to the first training sample to obtain the first combined feature; the aforementioned "combination" operation The method can be any one or more of the following operations: splicing, addition or other types of operations, etc.; input the first combined features into the first classification network to obtain the output of the first classification network corresponding to the first training sample First predictive decision information.

The training device trains the feature acquisition network according to the first loss function, including: the training device trains the feature acquisition network and the first classification network according to the first loss function, where the first loss function also includes a third loss function term , the third loss function term indicates the similarity between the first predicted decision information and the expected decision information corresponding to the first training sample.

In this implementation, in the process of training the feature acquisition network, the purpose of training the feature acquisition network not only includes accurately obtaining the characteristics of the information associated with the target attribute from the feature information of the first training sample, but also The first feature information and the second feature information are combined to obtain the first combined feature, and a third loss function term is introduced. The purpose of using the third loss function term for training includes improving the prediction decision information based on the first combined feature. The accuracy rate is conducive to further improving the accuracy rate of the prediction decision information obtained in the reasoning stage.

In a possible implementation of the first aspect, the training device trains the feature acquisition network and the first classification network, and after obtaining the trained feature acquisition network and the trained first classification network, the method further includes: training device acquisition third feature information, and combine the second feature information and the third feature information corresponding to the second training sample to obtain the second combined feature; wherein, the third feature information and the first feature corresponding to the second training sample The data size of the messages is the same and the data content is different. Further, if the first feature information is specifically expressed as an N-dimensional tensor, the third feature information is also expressed as an N-dimensional tensor, and the third feature information and the first feature information are in each of the aforementioned N dimensions. The lengths are the same. As an example, if the first feature information is specifically expressed as a vector, then the third feature information is also expressed as a vector, and the length of the third feature information and the first feature information are the same; if the first feature information is specifically expressed as a matrix, then the third feature information is specifically expressed as a vector. The feature information is also represented as a matrix, and the length and width of the third feature information and the first feature information are the same; the difference in data content of the third feature information and the first feature information corresponding to the second training sample refers to: The data contents of the third feature information and the first feature information corresponding to the second training sample are not exactly the same, that is, it suffices that there are different data in the third feature information and the first feature information corresponding to the second training sample.

The training device inputs the second combined features into the trained first classification network to obtain the second prediction decision information corresponding to the second training sample output by the trained first classification network; inputs the second combined features into the second classification network to obtain the third prediction decision information corresponding to the second training sample output by the second classification network; the training device trains the second classification network according to the second loss function to obtain the trained second classification network. The feature acquisition network and the trained second classification network belong to the same target neural network; "second prediction decision information" and "third prediction decision information" are both a kind of decision information, and "second prediction decision information" and "third prediction decision information" What is indicated by "Three Predictive Decision Information" depends on the type of target task performed by the target neural network.

Wherein, the second loss function includes a fourth loss function term and a fifth loss function term. The fourth loss function term indicates the similarity between the second prediction decision information and the expected decision information corresponding to the second training sample. The fifth loss function term indicates the similarity between the second prediction decision information and the expected decision information corresponding to the second training sample. The function indicates the degree of similarity between the second prediction decision information and the third prediction decision information.

In this implementation, if the data size of the third feature information and the first feature information are the same and the data content is different, then The combined features obtained by combining the first feature information and the second feature information corresponding to the second training sample are compared with the second combined features, that is, the characteristics of the information associated with the target attribute in the second training sample occur. Change, but the training goal also includes generating the original expected decision-making information of the second training sample, that is, non-existent training data is added during the training phase, and the training goal includes information based on different categories of target attributes, and the expected decision-making information can be obtained ; Through the above solution, it not only increases the diversity of training data; it also helps reduce the dependence of the trained neural network on the information associated with the target attributes in the input data, and pays more attention to the information related to the tasks performed, which is beneficial to This improves the accuracy of the output prediction decision information and is conducive to improving the fairness of the prediction decision information obtained for groups pointed to by different categories of target attributes.

In a possible implementation of the first aspect, the training device obtains the third feature information, including: the training device generates the first feature information corresponding to the second training sample through the trained feature acquisition network; The corresponding first feature information and disturbance information are weighted and summed to obtain the third feature information, and the weight value of the disturbance information is adjustable. Further, the disturbance information may be information randomly generated by the training equipment, or may include the gradient corresponding to the first loss function term. The disturbance information may also be obtained through other methods, etc., and is not exhaustive here.

In this implementation, the characteristic information of the second training sample is first decomposed, and then the characteristics of the information associated with the target attribute are intervened, thereby obtaining a training sample that is opposite to the actual situation of the second training sample. If the weight value of the disturbance information exceeds is larger, the lower the similarity between the third feature information and the first feature information, the more conducive to improving the fairness of the obtained prediction decision information; then the obtained prediction decision information can be balanced by adjusting the weight value of the disturbance information accuracy and fairness.

In a possible implementation of the first aspect, the method further includes: the training device performs a classification operation according to the second feature information corresponding to the first training sample, and obtains fourth prediction decision information corresponding to the first training sample; wherein, The first loss function also includes a sixth loss function term, the sixth loss function term indicates the similarity between the fourth predicted decision information and the expected decision information corresponding to the first training sample, and the second loss function term indicates the first loss function The similarity between the gradient corresponding to the term and the gradient corresponding to the sixth loss function term.

In this implementation, the gradient corresponding to the first loss function term is to enable the acquisition of the first feature information to more accurately reflect the characteristics of the information associated with the target attribute, and the gradient corresponding to the sixth loss function term is to enable the obtained first feature information to more accurately reflect the characteristics of the information associated with the target attribute. Based on the obtained second feature information, the interference of the information related to the target attribute included in the first training sample can be eliminated, so that more accurate prediction decision information can be generated. The second loss function term adopts the gradient sum corresponding to a loss function term. The similarity between the gradients corresponding to the sixth loss function term can further improve the efficiency of updating the weight parameters of the feature acquisition network, and is conducive to improving the first feature information and the second feature information generated by the feature acquisition network after training. accuracy.

In a possible implementation of the first aspect, the method is applied to any of the following scenarios: determining whether to recommend the object pointed by the first training sample, determining whether the object in the first training sample is in the target state, or determining whether to agree with the first The training sample points to the applicant's request. Correspondingly, various "decision information" in this solution can be used to indicate any of the following information: whether to recommend the object pointed to by the first training sample, whether the object in the first training sample is in the target state, whether to agree to the first training The sample points to the applicant's request. This implementation provides a variety of application scenarios for this method, which improves the implementation flexibility of this solution.

In the second aspect, embodiments of this application provide a data processing method that can use artificial intelligence technology to make decisions. The method includes: the execution device inputs the data to be processed into the feature acquisition network, performs feature extraction on the data to be processed through the feature acquisition network, and obtains the feature information of the data to be processed; and generates the first feature information through the feature acquisition network according to the feature information of the data to be processed. and second characteristic information, where the first characteristic information includes characteristics of information associated with the target attribute in the data to be processed. The execution device combines the first feature information and the second feature information to obtain the first combined feature; inputs the first combined feature into the classification network to obtain prediction decision information output by the classification network. Among them, the feature acquisition network is trained using a first loss function. The first loss function includes a first loss function term and a second loss function term. The first loss function term indicates the similarity between the predicted category information and the expected category information. The predicted The category information indicates the predicted category of the information associated with the target attribute in the data input to the feature acquisition network. The expected category information indicates the correct category of the information associated with the target attribute in the data input to the feature acquisition network. The second loss function term is used for training. The purpose includes reducing the similarity between the first feature information and the second feature information.

In a possible implementation of the second aspect, the first loss function further includes a third loss function term, and the third loss function term indicates the relationship between the first prediction decision information and the expected decision information corresponding to the data of the input feature acquisition network. The similarity of the first prediction decision information indicates the decision corresponding to the data input to the feature acquisition network.

In the second aspect of this application, the execution device can also be used to execute the steps performed by the training device in the first aspect and each possible implementation manner of the first aspect. The specific implementation manner and noun of the steps in each possible implementation manner of the second aspect Please refer to the first aspect for its meaning and beneficial effects, and will not be repeated here.

In the third aspect, embodiments of the present application provide a neural network training device that can use artificial intelligence technology to make decisions. The neural network training device includes: a feature extraction module for inputting the first training sample into the feature acquisition network, Feature extraction is performed on the first training sample through the feature acquisition network to obtain the feature information of the first training sample; the generation module is used to generate the first training sample corresponding to the first training sample through the feature acquisition network based on the feature information of the first training sample. Feature information and second feature information. The first feature information corresponding to the first training sample includes characteristics of the information associated with the target attribute in the first training sample; A feature information performs a classification operation to obtain predicted category information. The predicted category information indicates the predicted category of the first feature information corresponding to the first training sample. The predicted category is included in multiple categories corresponding to the target attribute; the training module is used to According to the first loss function, the feature acquisition network is trained to obtain the trained feature acquisition network; wherein the first loss function includes a first loss function term and a second loss function term, and the first loss function term indicates the predicted category information and The similarity between expected category information, the expected category information indicates the correct category of the information associated with the target attribute in the first training sample, and the purpose of using the second loss function term for training includes reducing the difference between the first feature information and the second feature information. similarity between.

In the third aspect of this application, the neural network training device can also be used to perform the steps performed by the training device in the first aspect and each possible implementation of the first aspect, and the specific implementation of the steps in each possible implementation of the third aspect. For the methods, meanings of nouns and the beneficial effects, please refer to the first aspect and will not be repeated here.

In the fourth aspect, embodiments of the present application provide a data processing device that can use artificial intelligence technology to make decisions. The data processing device includes: a feature extraction module for inputting the data to be processed into the feature acquisition network, and through the feature acquisition The network performs feature extraction on the data to be processed to obtain the feature information of the data to be processed; the generation module is used to generate first feature information and second feature information through the feature acquisition network according to the feature information of the data to be processed. The first feature information includes the feature information to be processed. Process the characteristics of the information associated with the target attribute in the data; the combination module is used to combine the first feature information and the second feature information to obtain the first combined feature; the classification module is used to input the first combined feature into the classification network to obtain the prediction decision information output by the classification network; wherein, the feature acquisition network is trained using a first loss function, the first loss function includes a first loss function term and a second loss function term, and the first loss function term indicates the prediction category information The similarity between the predicted category information and the expected category information. The predicted category information indicates the predicted category of the information associated with the target attribute in the data input to the feature acquisition network. The expected category information indicates the correctness of the information associated with the target attribute in the data input to the feature acquisition network. Category, the purpose of using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information.

In the fourth aspect of the present application, the data processing device can also be used to execute the second aspect and the steps performed by the execution device in each possible implementation of the second aspect, and the specific implementation of the steps in each possible implementation of the fourth aspect. , the meaning of nouns and the beneficial effects they bring can be referred to the first aspect, and will not be repeated here.

In a fifth aspect, embodiments of the present application provide a computer program product. The computer program product includes a program. When the program is run on a computer, it causes the computer to execute the neural network training method described in the first aspect, or causes the computer to perform the neural network training method described in the first aspect. The computer executes the data processing method described in the second aspect above.

In a sixth aspect, embodiments of the present application provide a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to execute the neural processing described in the first aspect. A network training method, or a method for causing a computer to perform the data processing method described in the second aspect above.

In the seventh aspect, embodiments of the present application provide a training device, including a processor and a memory. The processor is coupled to the memory. The memory is used to store programs; the processor is used to execute the program in the memory, so that the training device executes the above The training method of the neural network described in the first aspect.

In the eighth aspect, embodiments of the present application provide an execution device, including a processor and a memory. The processor is coupled to the memory. The memory is used to store programs; the processor is used to execute the program in the memory, so that the execution device executes the above The data processing method described in the second aspect.

In a ninth aspect, the present application provides a chip system, which includes a processor and is used to support an execution device or a communication device to implement the functions involved in the above aspects, for example, sending or processing data involved in the above methods and /or information. In a possible design, the chip system further includes a memory, and the memory is used to store necessary program instructions and data for execution devices or communication devices. The chip system may be composed of chips, or may include chips and other discrete devices.

Description of the drawings

Figure 1 is a schematic structural diagram of the artificial intelligence main framework provided by the embodiment of the present application;

Figure 2a is a system architecture diagram of the data processing system provided by the embodiment of the present application;

Figure 2b is a schematic flow chart of a neural network training method provided by an embodiment of the present application;

Figure 3 is a schematic flow chart of a neural network training method provided by an embodiment of the present application;

Figure 4 is a schematic flow chart of a neural network training method provided by an embodiment of the present application;

Figure 5 is a schematic diagram comparing the first feature information and the second feature information in the neural network training method provided by the embodiment of the present application;

Figure 6 is a schematic flow chart of a neural network training method provided by an embodiment of the present application;

Figure 7 is a schematic flow chart of a neural network training method provided by an embodiment of the present application;

Figure 8 is a schematic flow chart of the data processing method provided by the embodiment of the present application;

Figure 9 is a schematic diagram of the beneficial effects of the neural network training method provided by the embodiment of the present application;

Figure 10 is a schematic structural diagram of a neural network training device provided by an embodiment of the present application;

Figure 11 is a schematic structural diagram of a data processing device provided by an embodiment of the present application;

Figure 12 is a schematic structural diagram of an execution device provided by an embodiment of the present application;

Figure 13 is another structural schematic diagram of the training equipment provided by the embodiment of the present application;

Figure 14 is a schematic structural diagram of a chip provided by an embodiment of the present application.

Detailed ways

The embodiments of the present application are described below with reference to the accompanying drawings. Persons of ordinary skill in the art know that with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

The terms "first", "second", etc. in the description and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms so used are interchangeable under appropriate circumstances, and are merely a way of distinguishing objects with the same attributes in describing the embodiments of the present application. Furthermore, the terms "include" and "having" and any variations thereof, are intended to cover non-exclusive inclusions, such that a process, method, system, product or apparatus comprising a series of elements need not be limited to those elements, but may include not explicitly other elements specifically listed or inherent to such processes, methods, products or equipment.

First, the overall workflow of the artificial intelligence system is described. Please refer to Figure 1. Figure 1 shows a structural schematic diagram of the artificial intelligence main framework. The following is from the "intelligent information chain" (horizontal axis) and "IT value chain" ( The above artificial intelligence theme framework is elaborated on the two dimensions of vertical axis). Among them, the "intelligent information chain" reflects a series of processes from data acquisition to processing. For example, it can be the general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, intelligent execution and output. In this process, the data has gone through the condensation process of "data-information-knowledge-wisdom". The "IT value chain" reflects the value that artificial intelligence brings to the information technology industry, from the underlying infrastructure of human intelligence and information (providing and processing technology implementation) to the systematic industrial ecological process.

(1)Infrastructure

Infrastructure provides computing power support for artificial intelligence systems, enables communication with the external world, and supports it through basic platforms. Communicate with the outside through sensors; computing power is provided by a smart chip, which can specifically use a central processing unit (CPU), an embedded neural network processor (neural-network processing unit, NPU), a graphics processor ( Graphics processing unit (GPU), application specific integrated circuit (ASIC) or field programmable gate array (FPGA) and other hardware acceleration chips; the basic platform includes distributed computing framework and network and other related platforms Guarantee and support can include cloud storage and computing, interconnection networks, etc. For example, sensors communicate with the outside world to obtain data, which are provided to smart chips in the distributed computing system provided by the basic platform for calculation.

(2)Data

Data from the upper layer of the infrastructure is used to represent data sources in the field of artificial intelligence. The data involves graphics, images, voice, and text, as well as IoT data of traditional devices, including business data of existing systems as well as force, displacement, Sensing data such as liquid level, temperature, humidity, etc.

(3)Data processing

Data processing usually includes data training, machine learning, deep learning, search, reasoning, decision-making and other methods.

Among them, machine learning and deep learning can perform symbolic and formal intelligent information modeling, extraction, preprocessing, training, etc. on data.

Reasoning refers to the process of simulating human intelligent reasoning in computers or intelligent systems, using formal information to perform machine thinking and problem solving based on reasoning control strategies. Typical functions are search and matching.

Decision-making refers to the process of decision-making after intelligent information is reasoned, and usually provides functions such as classification, sorting, and prediction.

(4) General ability

After the data is processed as mentioned above, some general capabilities can be formed based on the results of further data processing, such as algorithms or a general system, such as translation, text analysis, computer vision processing, speech recognition, and image processing. identification, etc.

(5) Intelligent products and industry applications

Intelligent products and industry applications refer to the products and applications of artificial intelligence systems in various fields. They are the encapsulation of overall artificial intelligence solutions, productizing intelligent information decision-making and realizing practical applications. Its application fields mainly include: intelligent terminals, intelligent manufacturing, Smart transportation, smart home, smart healthcare, smart security, autonomous driving, smart city, etc.

The embodiments of the present application can be applied to various application fields in the field of artificial intelligence. Specifically, artificial intelligence technology can be used to solve decision-making problems in various application fields. As an example, in the field of smart cities, the predictive decision information output by the neural network can indicate whether to recommend specific news provided by a certain supplier; or, the predictive decision information output by the neural network can indicate whether to recommend specific news provided by a certain supplier. specific courses offered; alternatively, the predictive decision information output by the neural network can indicate whether to recommend a specific movie offered by a certain piece of paper.

As another example, for example, in the field of smart home, the prediction decision information output by the neural network can indicate whether the person in the image has curly hair; or, the prediction decision information output by the neural network can indicate whether the person in the image is smiling; or, The predictive decision-making information output by the neural network can indicate whether the person in the image is attractive or not.

As another example, for example, the prediction decision information output by the neural network can indicate whether to approve a specific applicant's loan request, etc. The application scenarios of the embodiments of this application are not exhaustive here.

In order to improve the accuracy of the prediction decision information output by the trained neural network, embodiments of the present application provide a neural network training method. Before introducing the foregoing method, please refer to Figure 2a. Figure 2a is a diagram of the present application. A system architecture diagram of the data processing system provided by the embodiment. In Figure 2a, the data processing system 200 includes a training device 210, a database 220, an execution device 230, a data storage system 240 and a client device 250. The execution device 230 includes a computing device Module 231.

The database 220 stores a training data set, the training device 210 generates the first model/rule 201, and uses the training data set to iteratively train the first model/rule 201 to obtain the trained first model/rule 201. The first model/rule 201 may be embodied as a neural network, or may be embodied as a non-neural network model. In the embodiment of this application, only the first model/rule 201 expressed as a neural network is used as an example for explanation. Further, the first model/rule 201 may include a neural network for feature acquisition on input data.

Specifically, please refer to Figure 2b. Figure 2b is a flowchart of a neural network training method provided by an embodiment of the present application. Schematic diagram. A1. The training device 210 inputs the first training sample into the feature acquisition network (that is, an example of the first model/rule 201), and performs feature extraction on the first training sample through the feature acquisition network to obtain feature information of the first training sample. A2. The training device 210 decomposes the feature information of the first training sample through the feature acquisition network to obtain the first feature information and the second feature information corresponding to the first training sample. The first feature information corresponding to the first training sample includes Characteristics of the information associated with the target attribute in the first training sample. A3. The training device 210 performs a classification operation according to the first feature information corresponding to the first training sample to obtain predicted category information. The predicted category information indicates the predicted category of the first feature information corresponding to the first training sample. The predicted category is included in the target in various categories corresponding to attributes. A4. The training device 210 trains the feature acquisition network according to the first loss function to obtain the trained feature acquisition network. The first loss function includes a first loss function term and a second loss function term. The first loss function term indicates prediction. The similarity between the category information and the expected category information. The expected category information indicates the correct category of the information associated with the target attribute in the first training sample. The purpose of using the second loss function term for training includes reducing the first feature information and the second similarity between feature information.

The trained first model/rules 201 obtained by the training device 210 will be deployed to the execution device 230. The execution device 230 can call the data, codes, etc. in the data storage system 240, or store the data, instructions, etc. in the data storage. In system 240. The data storage system 240 may be placed in the execution device 230 , or the data storage system 240 may be an external memory relative to the execution device 230 .

In some embodiments of the present application, please refer to Figure 2a. The execution device 230 and the client device 250 may be independent devices. The execution device 230 is configured with an input/output (I/O) interface for data interaction with the client device 250. The "user" can input the data to be processed through the client device 250. The client device 250 sends the data to be processed to the execution device 230 through the I/O interface. The execution device 230 generates and After the prediction decision information corresponding to the data to be processed is obtained, the aforementioned prediction decision information can be returned to the client device 250 through the I/O interface and provided to the user.

It is worth noting that Figure 2a is only an architectural schematic diagram of the data processing system provided by the embodiment of the present invention, and the positional relationship between the devices, devices, modules, etc. shown in the figure does not constitute any limitation. For example, in other embodiments of the present application, the execution device 230 may be configured in the client device 250. As an example, when the client device is a mobile phone or tablet, the execution device 230 may be the host processor (Host) of the mobile phone or tablet. A module in the CPU for array image processing. The execution device 230 can also be a graphics processing unit (GPU) or a neural network processor (NPU) in a mobile phone or tablet. The GPU or NPU serves as a co-processor. Loaded to the main processor, the main processor allocates tasks.

In conjunction with the above description, the specific implementation process of the training phase and inference phase of the neural network provided by the embodiment of the present application will be described below.

1. Training stage

In the embodiment of the present application, specifically, please refer to Figure 3. Figure 3 is a schematic flow chart of the neural network training method provided by the embodiment of the present application. The neural network training method provided by the embodiment of the present application may include:

301. The training device inputs the first training sample into the feature acquisition network, performs feature extraction on the first training sample through the feature acquisition network, and obtains feature information of the first training sample.

In the embodiment of this application, a training data set is deployed on the training device, and the training device can sample from the training data set At least one first training sample is input into a feature acquisition network, and feature extraction is performed on the first training sample through the feature acquisition network to obtain feature information of the first training sample.

Wherein, the feature acquisition network may include a first neural network module for feature extraction. The first neural network module may specifically adopt a convolutional neural network, a recurrent neural network, a residual neural network or other types of neural networks. Specifically, it may The selection is made based on the data type of the first training sample. As an example, the first neural network module can use residual neural network (residual neural network, ResNet)-18, ResNet-34 or other types of neural networks, etc., which are not mentioned here. Do exhaustion.

302. The training device generates first feature information and second feature information corresponding to the first training sample through the feature acquisition network based on the feature information of the first training sample. The first feature information corresponding to the first training sample includes the first training sample. Characteristics of the information in the sample associated with the target attribute.

In the embodiment of the present application, the training device can generate at least one first feature information and second feature information corresponding to the first training sample through the feature acquisition network according to the feature information of the first training sample; wherein, corresponding to the first training sample Each of the first feature information includes features of information associated with a target attribute in the first training sample, and the second feature information corresponding to the first training sample includes information that is not associated with each target attribute in the first training sample. The characteristics, that is, the training target includes different information in the obtained first characteristic information and second characteristic information.

Further, technicians can determine target attributes based on factors in the current task that cause bias in predictive decision information. As an example, for example, if a trained neural network is used to determine whether to recommend a certain course, and factors that cause bias in the prediction decision information include the supplier of the course, then at least one target attribute used in the training phase of the neural network may include Supplier, if the first training sample is a course, the information associated with the target attribute in the first training sample may include a watermark indicating the supplier of the course, an introduction to the supplier in the course cover, or other information; as another For example, if a trained neural network is used to determine whether a face in an image has curly hair, and the factors that cause bias in the predicted decision information include the gender of the person in the image, then at least one goal is used in the training phase of the neural network. The attributes may include the gender of the user. The information associated with the target attribute in the face image (i.e., the first training sample) may include the image information of the neck part in the face image. The image information of the neck part may be used to determine whether the user has Adam's apple. etc. It should be understood that the examples here are only for the convenience of understanding the two concepts of "target attribute" and "information associated with the target attribute in the training sample" and are not used to limit this solution.

Specifically, in one implementation, the feature acquisition network may include at least a second neural network module and a third neural network module that correspond to at least one target attribute. The training device generates the feature information of the first training sample. , a first feature information corresponding to the first training sample can be obtained from the feature information of the first training sample through each second neural network module, and the first feature information corresponding to the first training sample can be obtained from the feature information of the first training sample through the third neural network module. A second feature information corresponding to the first training sample is obtained.

In another implementation, the feature acquisition network may include a complete second neural network module. After the training device generates the feature information of the first training sample, it may input the feature information of the first training sample into the second neural network module. , performing a decomposition operation through the second neural network module to obtain at least one first feature information and one second feature information corresponding to the first training sample output by the second neural network module.

It should be noted that the feature acquisition network can also be embodied in other structural forms. The description here is only used to prove the feasibility of this solution and is not used to limit this solution.

303. The training device performs a classification operation according to the first feature information corresponding to the first training sample to obtain first prediction category information. The first prediction category information indicates the prediction category of the first feature information corresponding to the first training sample. The prediction category Included in various categories corresponding to the target attribute.

In the embodiment of the present application, the training device can input the first feature information corresponding to the first training sample into the first classifier, and perform the classification operation through the first classifier to obtain the first predicted category information generated by the first classifier; Wherein, the first prediction category information indicates the prediction category of the first feature information corresponding to the first training sample, that is, the first prediction category information indicates the prediction category corresponding to the information associated with the target attribute in the first training sample, and the prediction The category is included in various categories corresponding to the target attribute.

As an example, if a neural network is used to determine whether a certain course is recommended, and the target attribute is supplier, then the multiple categories corresponding to the target attribute may include supplier A, supplier B, supplier C, supplier D, etc.; As another example, for example, a neural network is used to determine whether the person in the image has curly hair, and the target attribute is gender, then the multiple categories corresponding to the target attribute may include male and female; as another example, for example, a neural network is used to determine whether the person in the image has curly hair. Whether a certain movie is recommended, the first training sample includes the user's rating of the movie and multiple attribute information of the movie itself. The target attribute can include the producer of the movie, and the multiple categories corresponding to the target attribute can include A system. Film producers, film producers B, film producers C, and film producers D, etc. I will not list them all here.

304. The training device performs a classification operation based on the second feature information corresponding to the first training sample, and obtains fourth prediction decision information corresponding to the first training sample.

In some embodiments of the present application, the training device may also input second feature information corresponding to the first training sample into a second classifier, and generate fourth prediction decision information corresponding to the first training sample through the second classifier.

Among them, the trained feature acquisition network is used to perform the target task, the "fourth prediction decision information" is the prediction decision information output by the first classifier, and the content indicated by the "fourth prediction decision information" depends on the target task. Further, the trained feature acquisition network can be applied to any of the following scenarios: determining whether to recommend the object pointed by the first training sample, determining whether the object in the first training sample is in the target state, or determining whether to agree with the object pointed by the first training sample. the applicant's request.

Correspondingly, the "fourth prediction decision information" can be used to indicate any of the following information: whether to recommend the object pointed by the first training sample, whether the object in the first training sample is in the target state, whether to agree with the object pointed by the first training sample The applicant's request or "fourth prediction decision information" may also be used to indicate other types of information and so on.

Further, as an example, for example, the object pointed by the first training sample may be news, courses or other types of objects, etc., for example, the target status may be smile, curly hair, attractive or other types of status, etc., such as the applicant's request. It can be a loan request, a promotion request, or other types of requests, etc. The examples here are only for the convenience of understanding this program and are not used to limit this program. It should be noted that the content of the information indicated by the “fourth prediction decision information” is determined based on the content of the target task, and is not exhaustive here. Multiple application scenarios of this method are provided, which improves the implementation flexibility of this solution.

305. The training device combines the first feature information and the second feature information corresponding to the first training sample to obtain the first combined feature.

In this embodiment of the present application, the training device may combine the first feature information and the second feature information corresponding to the first training sample to obtain the first combined feature. Among them, the aforementioned "combination" operation method can be any of the following or Various operations: splicing, addition or other types of operations, etc., not exhaustive here.

Specifically, after obtaining the first feature information and the second feature information corresponding to the first training sample, the training device can directly perform a combination operation on the aforementioned first feature information and the second feature information; it can also perform a combination operation on the first training sample and the first feature information. The corresponding first feature information and/or second feature information are preprocessed, and then the above combination operation is performed. The aforementioned preprocessing operations may include normalization processing, processing through activation functions, multiplication with preset weight values or other processing methods, etc., which are not limited here.

306. The training device inputs the first combined features into the first classification network to obtain the first prediction decision information corresponding to the first training sample output by the first classification network.

In the embodiment of the present application, the training device can input the first combined features into the first classification network to obtain the first prediction decision information corresponding to the first training sample output by the first classification network; the meaning of "first prediction decision information" The meaning is similar to the "fourth prediction decision information", but the difference is that the "first prediction decision information" is generated by the first classification network, and the "fourth prediction decision information" is generated by the second classifier.

307. The training device trains the feature acquisition network according to the first loss function, where the first loss function includes a first loss function term and a second loss function term, and the first loss function term indicates the first prediction category information and the first loss function term. The similarity between expected category information, the first expected category information indicates the correct category of the information associated with the target attribute in the first training sample, and the purpose of using the second loss function term for training includes reducing the first feature information and the second feature similarity between information.

In the embodiment of the present application, the training device can iteratively train the feature acquisition network and the first classifier according to the first loss function until the convergence conditions are met, and a trained feature acquisition network is obtained. The aforementioned convergence conditions may include the number of iterative training reaching a preset number, meeting the convergence conditions of the first loss function or other convergence conditions, etc., which are not exhaustive here.

The first loss function may include a first loss function term and a second loss function term, the first loss function term indicates the similarity between the first predicted category information and the first expected category information, and the first expected category information indicates the first training The correct category of the information associated with the target attribute in the sample, and the purpose of using the first loss function term for training includes improving the similarity between the first predicted category information and the first expected category information. The purpose of using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information.

Further, the first loss function term can adopt cosine similarity, L1 similarity, L2 similarity or other types of similarity between the first predicted category information and the first expected category information, or the first loss function term can Based on the Euclidean distance, cosine distance, Mahalanobis distance or other types of distance between the first predicted category information and the first expected category information, the greater the distance between the first predicted category information and the first expected category information, the first The smaller the similarity between the predicted category information and the first expected category information, it should be noted that the example of the first loss function term here is only for the convenience of understanding the first loss function term and is not used to limit this solution.

Specifically, steps 304 to 306 are all optional steps. If steps 304 to 306 are not executed, then after repeatedly executing steps 301 to 303 and step 307 multiple times, the trained feature acquisition network can be obtained, and the trained feature acquisition network can be The subsequent feature acquisition network is deployed to the execution device.

If steps 304 to 306 are executed, step 307 may include: the training device may iteratively train the feature acquisition network, the first classifier, the second classifier and the first classification network according to the first loss function until the convergence condition is met. software to obtain the trained feature acquisition network and the trained first classification network.

More specifically, after generating the function value of the first loss function, the training device performs gradient derivation on the function value of the first loss function, and reversely updates the feature acquisition network, the first classifier, the second classifier and the first classifier. The weight parameters of the classification network are used to complete a training of the feature acquisition network and the first classification network.

Wherein, the first loss function may also include a third loss function term and a sixth loss function term, and the third loss function term indicates the similarity between the first predicted decision information and the expected decision information corresponding to the first training sample; The six loss function terms indicate the similarity between the fourth predicted decision information and the expected decision information corresponding to the first training sample; "the similarity between the first predicted decision information and the expected decision information corresponding to the first training sample" For the calculation method of "the similarity between the fourth predicted decision information and the expected decision information corresponding to the first training sample", please refer to the calculation method of "the similarity between the first predicted category information and the first expected category information" , will not be described in detail here.

In this implementation, the second loss function term can directly calculate the similarity between the first feature information and the second feature information, or the second loss function term can also use the gradient corresponding to the first loss function term and the sixth The similarity between the gradients corresponding to the loss function term; for the calculation method of the similarity between the two gradients, please refer to the calculation method of "similarity between the first predicted category information and the first expected category information", here No further details will be given.

In order to further understand this solution, the calculation formula of the first loss function is shown below. Here, the similarity between two gradients is taken as the cosine similarity between the two gradients as an example.

Among them, L ₁ represents the first loss function, n represents the training device to obtain n training samples (that is, a batch of training samples) from the training data set to train the feature acquisition network, Represents the second loss function term; Represents the second feature information (that is, the feature of the information in the training sample that is not associated with the sensitive attribute), represents the first feature information (that is, the characteristics of the information associated with sensitive attributes in the training sample), g represents the first classification network, _yi represents the expected decision information corresponding to the first training sample, and _Li represents the third loss function term , that is, the similarity between the first prediction decision information output by the first classification network and _yi ; g _y represents the second classifier, represents the sixth loss function term, that is, the similarity between the fourth predicted decision information and the expected decision information corresponding to the first training sample; g _a represents the first classifier, a _i represents the target attribute in the training sample the correct category of the associated information (i.e. the first desired category of information), Represents the first loss function term, indicating the similarity between the first predicted category information and _ai ; Represents the sixth loss function term corresponding gradient, Represents the gradient corresponding to the first loss function term, represents the second loss function term. Here, the cosine similarity between the gradient corresponding to the first loss function term and the gradient corresponding to the sixth loss function term is used as an example; β ₁ , β ₂ and β ₃ respectively represent three weight value, it should be understood that the examples of specific implementation methods of the first loss function here are only for convenience of understanding this solution and are not used to limit this solution.

If step 304 is executed and steps 305 and 306 are not executed, step 307 may include: the training device may iteratively train the feature acquisition network, the first classifier, and the second classifier according to the first loss function until the convergence condition is met. , get the trained feature acquisition network.

Wherein, the first loss function may include a first loss function term, a second loss function term and a sixth loss function term; further, the second loss function term may directly calculate the similarity between the first feature information and the second feature information. Alternatively, the similarity between the gradient corresponding to the first loss function term and the gradient corresponding to the sixth loss function term can also be calculated.

In the embodiment of the present application, the gradient corresponding to the first loss function term is to enable the acquisition of the first feature information to more accurately reflect the characteristics of the information associated with the target attribute. The gradient corresponding to the sixth loss function term is In order to eliminate the interference of information related to the target attribute included in the first training sample based on the obtained second feature information, so as to generate more accurate prediction decision information, the second loss function term adopts the gradient corresponding to a loss function term. The similarity between the gradient corresponding to the sixth loss function term can further improve the efficiency of updating the weight parameters of the feature acquisition network, and is conducive to improving the first feature information and second features generated by the trained feature acquisition network. Accuracy of information.

If step 304 is not performed and steps 305 and 306 are performed, step 307 may include: the training device may iteratively train the feature acquisition network, the first classifier, and the first classification network according to the first loss function until the convergence condition is met. , obtain the trained feature acquisition network and the trained first classification network.

Wherein, the first loss function may include a first loss function term, a second loss function term and a third loss function term; the second loss function term in this implementation may calculate the difference between the first feature information and the second feature information. Similarity.

In the embodiment of the present application, in the process of training the feature acquisition network, the purpose of training the feature acquisition network not only includes accurately obtaining the features of the information associated with the target attribute from the feature information of the first training sample, but also The first feature information and the second feature information are combined to obtain the first combined feature, and a third loss function term is introduced. The purpose of using the third loss function term for training includes improving the prediction decision based on the first combined feature. The accuracy of the information is conducive to further improving the accuracy of the predictive decision-making information obtained in the reasoning stage.

In order to understand this solution more intuitively, please refer to Figures 4 and 5. Figure 4 is a schematic flowchart of a neural network training method provided by an embodiment of the present application, and Figure 5 is a neural network training method provided by an embodiment of the present application. A comparison diagram of the first feature information and the second feature information. Refer to Figure 4 first. In Figure 4, the prediction decision information output by the neural network indicates whether the person in the image is smiling is used as an example. The training device can obtain the first training sample of a batch (patch) (in Figure 4, four The first training sample is taken as an example), and the expected decision information corresponding to each first training sample is obtained (in Figure 4, the first one without a smile and the last three smiles are taken as an example). The training device inputs the four first training samples into the feature extraction network one after another to obtain the feature information of each first training sample, performs a decomposition operation on the feature information of each first training sample, and obtains the feature information corresponding to each first training sample. The first feature information and the second feature information, each first training The first feature information and the second feature information corresponding to the training sample are spliced, and the first combined feature corresponding to each first training sample can be obtained. The training device inputs the first combined features corresponding to each first training sample into the first classification network to obtain the first prediction decision information corresponding to each first training sample.

The training device may also perform a classification operation based on the first feature information corresponding to each first training sample to obtain the first prediction category information corresponding to each first feature information; Prediction category information, first expected category information corresponding to each first feature information, second feature information corresponding to each first training sample, first prediction decision information corresponding to each first training sample, and each third The expected decision information corresponding to a training sample is generated to generate the function value of the first loss function, perform gradient derivation of the function value of the first loss function, and reversely update the weight parameters of the feature extraction network and the first classification network to complete Multiple trainings of the feature extraction network and the first classification network. It should be understood that the example in Figure 4 is only for convenience of understanding this solution and is not used to limit this solution.

Please continue to refer to Figure 5. Figure 5 is an image obtained after visualizing the first feature information and the second feature information. In Figure 5, the prediction decision information output by the neural network indicates whether the person in the image has curly hair. The target attribute is Taking gender as an example, as shown in Figure 5, the second feature information corresponding to the first training sample (that is, the features of the information that is not associated with the target attribute in the first training sample) carries more of the hair in the image The characteristic information of the area, the first characteristic information corresponding to the first training sample (that is, the characteristics of the information associated with the target attribute in the first training sample) has less characteristic information of the hair area, and carries more information related to the image. Characteristics of gender-related information; comparing the second characteristic information and the first characteristic information corresponding to the first training sample, the focus of the two is obviously different. It should be understood that the example in Figure 5 is only for the convenience of understanding this solution and is not used for Limited to this plan.

308. The training device generates second feature information corresponding to the second training sample through the trained feature acquisition network.

In some embodiments of the present application, the training device can obtain the trained feature acquisition network after iteratively training the feature acquisition network according to the first loss function. The training device can also obtain the second training sample from the training data set, and through the training The final feature acquisition network (that is, the trained feature acquisition network obtained in steps 301 to 307) generates the second feature information corresponding to the second training sample; optionally, the training device can generate the second feature information corresponding to the second training sample. first characteristic information and second characteristic information.

For the specific implementation of the above steps, please refer to the description in step 302, "First feature information and second feature information corresponding to the second training sample" and "First feature information and second feature information corresponding to the first training sample" "The concept is similar, the difference is that the first training sample and the second training sample are different training samples.

Optionally, after acquiring the second feature information corresponding to the second training sample, the training device can also perform a classification operation based on the second feature information corresponding to the second training sample to obtain the fifth feature information corresponding to the second training sample. Prediction decision information; the concepts of "fifth prediction decision information corresponding to the second training sample" and "fourth prediction decision information corresponding to the first training sample" are similar. For the specific acquisition method of "fifth prediction decision information", please refer to The description of the specific acquisition method of the "fourth prediction decision information" in the above steps will not be repeated here.

309. The training device obtains the third feature information. The third feature information and the first feature information corresponding to the second training sample have the same data size and different data contents.

In some embodiments of the present application, the training device can also obtain third feature information; wherein the third feature information and the first feature information corresponding to the second training sample have the same data size and different data contents.

Further, if the first feature information is specifically expressed as an N-dimensional tensor, the third feature information is also expressed as an N-dimensional tensor. tensor, and the third feature information and the first feature information have the same length in each of the aforementioned N dimensions. As an example, if the first feature information is specifically expressed as a vector, then the third feature information is also expressed as a vector, and the length of the third feature information and the first feature information are the same; if the first feature information is specifically expressed as a matrix, then the third feature information is specifically expressed as a vector. The feature information is also represented as a matrix, and the length and width of the third feature information and the first feature information are the same, so an exhaustive list is not included here.

The difference in data content between the third feature information and the first feature information corresponding to the second training sample means that the data content of the third feature information and the first feature information corresponding to the second training sample are not exactly the same, that is, the third feature information and the first feature information corresponding to the second training sample are different in data content. It suffices that there are different data in the three feature information and the first feature information corresponding to the second training sample.

Specifically, in one implementation, the training device can obtain the first feature information corresponding to the second training sample, perform a weighted summation of the first feature information and the disturbance information corresponding to the second training sample, and obtain the third feature. information; where the weight value of the disturbance information can be variable or fixed.

Further, the disturbance information may be information randomly generated by the training equipment, or may include the gradient corresponding to the first loss function term. The disturbance information may also be obtained through other methods, etc., and is not exhaustive here.

In order to understand this solution more intuitively, an example of the calculation formula of the third characteristic information is disclosed as follows:

in, represents the third characteristic information, represents the first characteristic information, represents the disturbance information, α ₁ represents the weight of the disturbance information, α ₁ is uniformly sampled from [0, λ], Represents the gradient corresponding to the first loss function term, Represents the first loss function term, represents the predicted category information corresponding to the first feature information output by the first classifier, and a _i represents the expected category information corresponding to the first feature information (that is, the expected category information of the information associated with the target attribute in the second training sample) , represents the second norm of the gradient corresponding to the first loss function term; further, λ can be constant or adjustable. It should be understood that the example in formula (6) is only for the convenience of understanding this solution, It is not used to limit this plan.

In another implementation, the training device may also obtain the first feature information from the third training sample, and determine the first feature information corresponding to the third training sample as the third feature corresponding to the second training sample. information; the correct category of the information associated with the target attribute in the third training sample, and the correct category of the information associated with the target attribute in the second training sample are different.

In another implementation, the training device can randomly obtain a piece of third feature information, and the first feature information corresponding to the second training sample has the same data size as the aforementioned randomly obtained third feature information. It should be noted that the training device can also obtain the third feature information in other ways, and this list will not be exhaustive here.

Optionally, after acquiring the first feature information corresponding to the second training sample, the training device can also perform a classification operation based on the first feature information corresponding to the second training sample to obtain the second prediction category information. The second prediction The category information indicates the prediction category of the first feature information corresponding to the second training sample; the concepts of "second prediction category information" and "first prediction category information" are similar. For the specific acquisition method of "second prediction category information", please refer to The description of the specific acquisition method of the "first prediction category information" in the above steps will not be repeated here.

310. The training device combines the second feature information and the third feature information corresponding to the second training sample to obtain the second combined feature.

311. The training device inputs the second combined features into the trained first classification network, and obtains the second prediction decision information corresponding to the second training sample output by the trained first classification network.

In the embodiment of this application, the specific implementation of steps 310 and 311 can refer to the description in steps 305 and 306 above. The difference is that the "second feature information corresponding to the first training sample" in steps 305 and 306 is replaced by step "Second feature information corresponding to the second training sample" in steps 310 and 311, replace "first feature information corresponding to the first training sample" in steps 305 and 306 with "third feature information corresponding to the first training sample" in steps 310 and 311 The meanings of "feature information" and "second prediction decision information" are similar to the meanings of the above-mentioned "fourth prediction decision information", and will not be described again here.

312. The training device inputs the second combined features into the second classification network, and obtains the third prediction decision information output by the second classification network corresponding to the second training sample.

In some embodiments of the present application, the training device can input the second combined features into the second classification network to obtain the third prediction decision information corresponding to the second training sample output by the second classification network, "third prediction decision information" The meaning is similar to the meaning of the "fourth prediction decision information" mentioned above, and will not be described again here.

313. The training device trains the second classification network according to the second loss function, where the second loss function includes a fourth loss function term and a fifth loss function term, and the fourth loss function term indicates the second prediction decision information and the fifth loss function term. The similarity between the expected decision information corresponding to the two training samples, and the fifth loss function indicates the similarity between the second prediction decision information and the third prediction decision information.

In some embodiments of the present application, the training device can also keep the weight parameters of the first classification network unchanged, and iteratively train the second classification network according to the second loss function until the convergence conditions are met to obtain the trained second classification network. , the trained feature acquisition network and the trained second classification network belong to the same target neural network, and the target neural network will be arranged on the execution device to perform the target task. The aforementioned target task may be any of the following tasks: determining whether to recommend the object pointed to by the first training sample, determining whether the object in the first training sample is in the target state, determining whether to agree to the request of the applicant pointed to by the first training sample, or other tasks. type of tasks.

Optionally, the training device can also keep the weight parameters of the first classification network unchanged, and iteratively train the feature acquisition network and the second classification network according to the second loss function until the convergence conditions are met, to obtain the trained second classification network. and the feature acquisition network after retraining, the second classification network after training and the feature acquisition network after retraining belong to the above-mentioned target neural network.

Wherein, the second loss function at least includes a fourth loss function term and a fifth loss function term, the fourth loss function term indicates the similarity between the second prediction decision information and the expected decision information corresponding to the second training sample, and the fifth loss function term indicates the similarity between the second prediction decision information and the expected decision information corresponding to the second training sample. The loss function indicates the similarity between the second prediction decision information and the third prediction decision information.

Further, the fourth loss function term may specifically adopt cosine similarity, L1 similarity, L2 similarity or other types of similarity between the second predicted decision information and the expected decision information corresponding to the second training sample, or , the fourth loss function term can be obtained based on the Euclidean distance, cosine distance, Mahalanobis distance or other types of distance between the second predicted decision information and the expected decision information corresponding to the second training sample, etc., no exhaustive list will be made here. . For the specific expression form of the fifth loss function term, please refer to the specific expression form of the fourth loss function term, and will not be described in detail here.

Optionally, the second loss function may also include a first loss function term and a second loss function term; for "first loss function For the description of "loss function term and second loss function term", please refer to the description in step 307. It should be noted that in step 307, the first loss function term and the second loss function term are both calculated based on the first training sample. Step 313 is calculated based on the second training sample.

Further optionally, the second loss function may also include a seventh loss function term, the seventh loss function term indicates the similarity between the fifth predicted decision information and the expected decision information corresponding to the second training sample; then the second loss The function term may adopt the similarity between the gradient corresponding to the first loss function term and the gradient corresponding to the seventh loss function term.

To further understand this solution, an example of the calculation formula of the second loss function is disclosed below:

Among them, L ₂ represents the second loss function, and n represents the training device that obtains n training samples (that is, a batch of training samples) from the training data set to train the feature acquisition network and the second classification network. Represents the fourth loss function term, Represents the fifth loss function term, represents the third characteristic information, represents the third prediction decision information corresponding to the second training sample output by the second classification network, _yi represents the expected decision information corresponding to the second training sample, represents the second prediction decision information corresponding to the second training sample output by the trained first classification network, γ represents the weight value, and the meaning of other characters in equation (7) can be found in the above introduction to equation (1), here No further details will be given.

In order to understand this solution more intuitively, please refer to Figures 6 and 7. Figures 6 and 7 are two schematic flow charts of neural network training methods provided by embodiments of the present application. Figure 6 mainly shows the process of training the feature acquisition network and the second classification network included in the target neural network in the second training stage. In Figure 6, the prediction decision information output by the neural network indicates whether the person in the image is smiling is used as an example. The training device can obtain a batch of second training samples (in Figure 4, four second training samples are For example), and obtain the expected decision information corresponding to each second training sample (in Figure 4, the first one without a smile and the last three smiles are taken as an example). The training device inputs the four second training samples into the feature extraction network one after another to obtain the feature information of each second training sample, performs a decomposition operation on the feature information of each second training sample, and obtains the feature information corresponding to each second training sample. first characteristic information and second characteristic information.

The training device combines the first feature information corresponding to each second training sample and the disturbance information to obtain the third feature information corresponding to each second training sample, and combines the third feature information corresponding to each second training sample By splicing the information and the second feature information, the second combined features corresponding to each second training sample can be obtained.

The training device separates the second combined features corresponding to each second training sample into the first classification network and the second classification network after training, and obtains the second classification network output by the trained first classification network corresponding to the second training sample. Prediction decision information, and third prediction decision information corresponding to the second training sample output by the second classification network.

The training device may generate the second prediction decision information according to the second prediction decision information corresponding to each second training sample, the third prediction decision information corresponding to each second training sample, and the expected decision information corresponding to each second training sample. The function value of the loss function is reversely derived from the function value of the second loss function, and the weight parameters of the feature extraction network and the second classification network are updated to complete multiple trainings of the feature extraction network and the first classification network. It should be noted that the specific meaning of the second loss function can be referred to the above description. The example in Figure 6 is only for convenience of understanding this solution and is not used to limit this solution.

In the embodiment of the present application, if the data size of the third feature information and the first feature information are the same and the data content is different, then the combined feature obtained by combining the first feature information and the second feature information corresponding to the second training sample, Compared with the second combined features, that is, the changes in the characteristics of the information associated with the target attribute in the second training sample, but the goal of training also includes generating the original expected decision information of the second training sample, that is, in the training phase Non-existent training data is added, and the training target includes information based on different categories of target attributes to obtain the expected decision-making information; through the above solution, not only the diversity of training data is increased; it is also beneficial to reducing the impact of the neural network after training. The dependence of the information associated with the target attributes in the input data, paying more attention to the information related to the tasks performed, is conducive to improving the accuracy of the output prediction decision information, and is conducive to improving the groups pointed to by different categories of target attributes. The fairness of the obtained prediction decision-making information.

In the embodiment of the present application, the characteristic information of the second training sample is first decomposed, and then the characteristics of the information associated with the target attribute are intervened, thereby obtaining a training sample that is opposite to the actual situation of the second training sample. If the weight value of the perturbation information is The larger the value, the lower the similarity between the third feature information and the first feature information, which is more conducive to improving the fairness of the obtained prediction decision information; then the obtained prediction decision can be balanced by adjusting the weight value of the disturbance information Accuracy and fairness of information.

In order to understand this solution more intuitively, please refer to Figure 7 , which is a schematic flowchart of a neural network training method provided by an embodiment of the present application. As shown in Figure 7, the neural network training method provided by the embodiment of the present application can be divided into a first training stage and a second training stage. In the first training stage, the training device inputs the first training sample into the feature extraction network, The extracted feature information is decomposed to obtain the first feature information and the second feature information corresponding to the first training sample, and the first feature information and the second feature information corresponding to the first training sample are spliced to obtain the first feature information. A combined feature; the training device can input the first combined feature into the first classification network to obtain the first prediction decision information output by the first classification network corresponding to the first training sample; the training device can obtain according to the aforementioned steps information, generate the function value of the first loss function, and update the weight parameters of the feature acquisition network and the first classification network. The training device repeatedly performs the aforementioned steps to iteratively train the feature acquisition network and the first classification network to obtain the trained feature acquisition network and the trained first classification network.

In the second training phase, the training device inputs the second training sample into the feature extraction network, decomposes the extracted feature information to obtain the first feature information and the second feature information corresponding to the second training sample, and converts the second training sample into the feature extraction network. The first feature information and disturbance information corresponding to the training sample are weighted and summed to obtain the third feature information, and the second feature information and the third feature information corresponding to the second training sample are spliced to obtain the second combined feature.

The training device inputs the second combined features into the second classification network and the trained first classification network respectively, obtains the third prediction decision information output by the second classification network and corresponds to the second training sample, and obtains the trained first The second prediction decision information corresponding to the second training sample output by the classification network; the training device can, based on the obtained aforementioned information, Generate the function value of the second loss function, keep the weight parameters of the first classification network unchanged, and update the weight parameters of the feature acquisition network and the first classification network. The training device repeatedly performs the aforementioned steps to iteratively train the feature acquisition network and the second classification network to obtain the retrained feature acquisition network and the trained second classification network. It should be understood that the example in Figure 7 is only for convenience of understanding this solution and is not used to limit this solution.

In the embodiment of the present application, the trained feature acquisition network can respectively obtain the characteristics of the information associated with the target attribute in the input data, and the characteristics of the information not associated with the target attribute in the input data, so that the user can obtain the characteristics of the information based on the current task. Determine the target attributes that cause deviations in the prediction decision information, and extract the characteristics of the information associated with the target attributes that cause the deviations and the characteristics of the information not associated with the target attributes, which will help reduce the difficulty of the subsequent decision-making process and have It is beneficial to improve the accuracy of the final prediction decision information.

2. Reasoning stage

In the embodiment of the present application, specifically, please refer to Figure 8. Figure 8 is a schematic flow chart of the data processing method provided by the embodiment of the present application. The data processing method provided by the embodiment of the present application may include:

801. The execution device inputs the data to be processed into the feature acquisition network, extracts features from the data to be processed through the feature acquisition network, and obtains the feature information of the data to be processed.

802. The execution device generates first feature information and second feature information through the feature acquisition network based on the feature information of the data to be processed. The first feature information includes features of the information associated with the target attribute in the data to be processed.

In the embodiment of this application, the specific implementation of steps 801 and 802 can refer to the description of steps 301 and 302 in the corresponding embodiment of Figure 3 above. The difference is that the "first training sample" in steps 301 and 302 is replaced with step 801 and "data to be processed" in step 802. The specific meaning of each noun in steps 801 and 802 can be referred to the description in the corresponding embodiment of Figure 3 above, and will not be described again here.

803. The execution device combines the first feature information and the second feature information to obtain the first combined feature.

804. The execution device inputs the first combined features into the classification network to obtain prediction decision information output by the classification network. The feature acquisition network is trained using a first loss function, and the first loss function at least includes a first loss function term and a second loss function. Loss function term, the first loss function term indicates the similarity between the predicted category information and the expected category information, the predicted category information indicates the predicted category of the information associated with the target attribute in the data of the input feature acquisition network, and the expected category information indicates the input feature The purpose of obtaining the correct category of information associated with the target attribute in the network data and using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information.

In the embodiment of this application, the specific implementation of steps 803 and 804 can refer to the description of steps 305 and 306 in the corresponding embodiment of Figure 3 above. The difference is that the "first training sample" in steps 305 and 306 is replaced with step 803 and "data to be processed" in step 804. For the specific meanings of each noun in steps 803 and 804, please refer to the description in the corresponding embodiment of Figure 3 above, and will not be described again here.

Among them, the feature acquisition network is trained using the first loss function. For the specific form of the feature acquisition network, please refer to the description in the corresponding embodiment of Figure 3; the classification network can specifically use the first classification network in the corresponding embodiment of Figure 3, or it can also be used Figure 3 corresponds to the second classification network in the embodiment, and may also be a trained classification network obtained by other methods, etc., and is not exhaustive here.

Optionally, the feature acquisition network and the classification network are trained using the first loss function and the second loss function. For the meaning of the first loss function and the second loss function, please refer to the description in the corresponding embodiment in Figure 3, which will not be done here. Repeat. For the specific meaning of each loss function term in the "first loss function and the second loss function", you can also refer to the description in the corresponding embodiment in Figure 3, which will not be described here.

In the embodiment of the present application, the characteristics of the information associated with the target attribute in the data to be processed are obtained respectively, and the characteristics of the information in the data to be processed that are not associated with the target attribute are obtained, because the user can predict and make decisions based on the information in the current task. Determine the cause of the deviation to determine the target attribute, and extract the characteristics of the information associated with the target attribute that caused the deviation and the characteristics of the information not associated with the target attribute, which is beneficial to the difficulty of the classification network in the process of generating predictive decision-making information. It is beneficial to improve the accuracy of the final prediction decision information.

Next, the beneficial effects brought by the embodiments of the present application will be demonstrated in combination with experimental data. Figure 9 is a schematic diagram of the beneficial effects of the neural network training method provided by the embodiments of the present application. Figure 9 shows two sub-schematic diagrams on the left and right. In Figure 9, the target attribute is the film producer. According to the number of films produced, the producers are divided into mass producers and niche producers. After training A neural network is used to determine whether to recommend a certain movie. The left sub-schematic diagram and the right sub-schematic diagram of Figure 9 both use a polyline to show the accuracy and fairness of the prediction decision information generated by the trained neural network obtained through the embodiment of the present application, and use three polylines and a triangle to show the comparison. The group's method obtains the accuracy and fairness of the predictive decision-making information generated by the trained neural network. Among them, the higher the score of the accuracy index, the higher the accuracy of the generated prediction decision information; the lower the score of the fairness index, the better the fairness of the generated prediction decision information on the target attribute.

In some embodiments of the present application, there will be a second training phase. In the second training phase, the third feature information will be obtained, and the second feature information corresponding to the second training sample will be combined with the third feature information to obtain Construct training samples that don't actually exist. The rightmost point in the left sub-schematic diagram and the right sub-schematic diagram of Figure 9 represents the accuracy and fairness of the prediction decision information without the second training stage. As shown in the figure, if the second training stage is not used, whether it is the left diagram of Figure 9 or the right diagram of Figure 9, the prediction decision information obtained based on the method of the embodiment of the present application has the highest accuracy. If the second training stage is adopted, the prediction decision information obtained based on the method of the embodiment of the present application has the highest accuracy and good fairness.

On the basis of the embodiments corresponding to Figures 1 to 9, in order to better implement the above solutions of the embodiments of the present application, relevant equipment for implementing the above solutions is also provided below. Specifically referring to Figure 10, Figure 10 is a schematic structural diagram of a neural network training device provided by an embodiment of the present application. The neural network training device 1000 includes: a feature extraction module 1001 for inputting the first training sample into the feature acquisition network, Feature extraction is performed on the first training sample through the feature acquisition network to obtain feature information of the first training sample; the generation module 1002 is configured to generate a third training sample corresponding to the first training sample through the feature acquisition network based on the feature information of the first training sample. A feature information and a second feature information. The first feature information corresponding to the first training sample includes the characteristics of the information associated with the target attribute in the first training sample; the first classification module 1003 is used to classify the information according to the first training sample corresponding to the first training sample. Perform a classification operation on the first feature information to obtain predicted category information. The predicted category information indicates the predicted category of the first feature information corresponding to the first training sample. The predicted category is included in multiple categories corresponding to the target attribute; training module 1004 , used to train the feature acquisition network according to the first loss function to obtain the trained feature acquisition network; wherein the first loss function includes a first loss function term and a second loss function term, and the first loss function term indicates prediction The similarity between the category information and the expected category information, which indicates the correct category of the information associated with the target attribute in the first training sample, is performed using the second loss function term The purpose of training includes reducing the similarity between the first feature information and the second feature information.

In one possible design, the neural network training device 1000 further includes: a combination module for combining the first feature information and the second feature information corresponding to the first training sample to obtain the first combined feature; The second classification module is used to input the first combined features into the first classification network to obtain the first prediction decision information corresponding to the first training sample output by the first classification network; the training module 1004 is specifically used to calculate the first loss function according to the first loss function. , train the feature acquisition network and the first classification network, where the first loss function also includes a third loss function term, and the third loss function term indicates the relationship between the first prediction decision information and the expected decision information corresponding to the first training sample. similarity between.

In one possible design, the neural network training device 1000 further includes: an acquisition module, configured to acquire third feature information, where the data size of the third feature information and the first feature information corresponding to the second training sample are the same. And the data content is different; the combination module is also used to combine the second feature information and the third feature information corresponding to the second training sample to obtain the second combined feature; the second classification module is also used to combine the second combination The latter features are input into the trained first classification network to obtain the second prediction decision information corresponding to the second training sample output by the trained first classification network; the second classification module is also used to input the second combined features into the third The second classification network obtains the third prediction decision information corresponding to the second training sample output by the second classification network; the training module 1004 is specifically used to train the second classification network according to the second loss function to obtain the trained third For the two-classification network, the trained feature acquisition network and the trained second classification network belong to the same target neural network; among them, the second loss function includes the fourth loss function term and the fifth loss function term, and the fourth loss function term indicates The similarity between the second prediction decision information and the expected decision information corresponding to the second training sample, and the fifth loss function indicates the similarity between the second prediction decision information and the third prediction decision information.

In one possible design, the acquisition module is specifically used to: generate first feature information corresponding to the second training sample through the trained feature acquisition network; combine the first feature information and disturbance information corresponding to the second training sample Perform weighted summation to obtain the third feature information, and the weight value of the disturbance information is adjustable.

In a possible design, the second classification module is also used to perform a classification operation based on the second feature information corresponding to the first training sample to obtain the fourth prediction decision information corresponding to the first training sample, wherein the first The loss function also includes a sixth loss function term, the sixth loss function term indicates the similarity between the fourth predicted decision information and the expected decision information corresponding to the first training sample, and the second loss function term indicates the first loss function term. The similarity between the corresponding gradient and the gradient corresponding to the sixth loss function term.

In one possible design, the neural network training device 1000 is applied to any of the following scenarios: determining whether to recommend the object pointed to by the first training sample, determining whether the object in the first training sample is in the target state, or determining whether to agree with the first training sample. A training sample points to the applicant's request.

It should be noted that the information interaction and execution process between each module/unit in the neural network training device 1000 are based on the same concept as the various method embodiments corresponding to Figures 2b to 7 in this application. For specific content, please refer to The descriptions in the method embodiments shown above in this application will not be repeated here.

The embodiment of the present application also provides a data processing device. Please refer to Figure 11. Figure 11 is a schematic structural diagram of the data processing device provided by the embodiment of the present application. The data processing device 1100 includes: a feature extraction module 1101. The data to be processed is input into the feature acquisition network, and the features of the data to be processed are extracted through the feature acquisition network to obtain the feature information of the data to be processed; the generation module 1102 is used to generate the first first step through the feature acquisition network according to the feature information of the data to be processed. Feature information and second feature information. The first feature information includes the relationship between the target attributes in the data to be processed. characteristics of the associated information; the combination module 1103 is used to combine the first feature information and the second feature information to obtain the first combined features; the classification module 1104 is used to input the first combined features into the classification network to obtain the classification Prediction decision information output by the network; wherein, the feature acquisition network is trained using a first loss function. The first loss function includes a first loss function term and a second loss function term. The first loss function term indicates the predicted category information and the expected category information. The similarity between the two, the predicted category information indicates the predicted category of the information associated with the target attribute in the data input to the feature acquisition network, and the expected category information indicates the correct category of the information associated with the target attribute in the data input to the feature acquisition network, using the The purpose of training the two loss function terms includes reducing the similarity between the first feature information and the second feature information.

In a possible design, the first loss function also includes a third loss function term, and the third loss function term indicates the similarity between the first predicted decision information and the expected decision information corresponding to the data of the input feature acquisition network, The first prediction decision information indicates a decision corresponding to the data input to the feature acquisition network.

It should be noted that the information interaction, execution process, etc. between the modules/units in the data processing device 1100 are based on the same concept as the various method embodiments corresponding to Figure 8 in this application. For specific content, please refer to the foregoing description of this application. The descriptions in the method embodiments shown are not repeated here.

Next, an execution device provided by an embodiment of the present application will be introduced. Please refer to FIG. 12 . FIG. 12 is a schematic structural diagram of an execution device provided by an embodiment of the present application. Specifically, the execution device 1200 includes: a receiver 1201, a transmitter 1202, a processor 1203 and a memory 1204 (the number of processors 1203 in the execution device 1200 may be one or more, one processor is taken as an example in Figure 12) , wherein the processor 1203 may include an application processor 12031 and a communication processor 12032. In some embodiments of the present application, the receiver 1201, the transmitter 1202, the processor 1203, and the memory 1204 may be connected through a bus or other means.

Memory 1204 may include read-only memory and random access memory and provides instructions and data to processor 1203 . A portion of memory 1204 may also include non-volatile random access memory (NVRAM). The memory 1204 stores processor and operating instructions, executable modules or data structures, or a subset thereof, or an extended set thereof, where the operating instructions may include various operating instructions for implementing various operations.

The processor 1203 controls the execution of operations of the device. In specific applications, various components of the execution device are coupled together through a bus system. In addition to the data bus, the bus system may also include a power bus, a control bus, a status signal bus, etc. However, for the sake of clarity, various buses are called bus systems in the figure.

The methods disclosed in the above embodiments of the present application can be applied to the processor 1203 or implemented by the processor 1203. The processor 1203 may be an integrated circuit chip with signal processing capabilities. During the implementation process, each step of the above method can be completed by instructions in the form of hardware integrated logic circuits or software in the processor 1203 . The above-mentioned processor 1203 can be a general-purpose processor, a digital signal processing (DSP), a microprocessor or a microcontroller, and can further include an application specific integrated circuit (ASIC), a field programmable Gate array (field-programmable gate array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The processor 1203 can implement or execute the various methods, steps and logical block diagrams disclosed in the embodiments of this application. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly The implementation is implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other mature storage media in this field. The storage medium is located in the memory 1204. The processor 1203 reads the information in the memory 1204 and completes the steps of the above method in combination with its hardware.

The receiver 1201 may be configured to receive input numeric or character information and generate signal inputs related to performing relevant settings and functional controls of the device. The transmitter 1202 can be used to output numeric or character information through the first interface; the transmitter 1202 can also be used to send instructions to the disk group through the first interface to modify the data in the disk group; the transmitter 1202 can also include a display device such as a display screen .

In this embodiment of the present application, the application processor 12031 in the processor 1203 is used to execute the data processing method executed by the execution device in the corresponding embodiment of FIG. 8 . It should be noted that the specific manner in which the application processor 12031 performs each step in the data processing method is based on the same concept as the various method embodiments corresponding to Figure 8 in this application, and the technical effects it brings correspond to Figure 8 in this application. The various method embodiments are the same. For specific details, please refer to the descriptions in the method embodiments shown above in this application, and will not be described again here.

The embodiment of the present application also provides a training device. Please refer to Figure 13. Figure 13 is a schematic structural diagram of the training device provided by the embodiment of the present application. Specifically, the training device 1300 is implemented by one or more servers. The training device 1300 There may be relatively large differences due to different configurations or performance, and may include one or more central processing units (CPU) 1322 (for example, one or more processors) and memory 1332, one or more storage applications Storage medium 1330 for program 1342 or data 1344 (eg, one or more mass storage devices). Among them, the memory 1332 and the storage medium 1330 may be short-term storage or persistent storage. The program stored in the storage medium 1330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the training device. Furthermore, the central processor 1322 may be configured to communicate with the storage medium 1330 and execute a series of instruction operations in the storage medium 1330 on the training device 1300 .

The training device 1300 may also include one or more power supplies 1326, one or more wired or wireless network interfaces 1350, one or more input and output interfaces 1358, and/or, one or more operating systems 1341, such as Windows Server™, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM and so on.

In this embodiment of the present application, the central processor 1322 is used to execute the neural network training method executed by the training device in the corresponding embodiments of Figures 2b to 7. It should be noted that the specific manner in which the central processor 1322 executes each step in the neural network training method is based on the same concept as the various method embodiments corresponding to Figures 2b to 7 in this application, and the technical effects it brings are the same as those in this application. The respective method embodiments corresponding to Figures 2b to 7 are the same. For details, please refer to the descriptions in the method embodiments shown above in this application, and will not be described again here.

An embodiment of the present application also provides a computer program product that, when run on a computer, causes the computer to perform the steps performed by the execution device in the method described in the embodiment shown in FIG. 8, or causes the computer to perform the following: The steps performed by the training device in the method described in the embodiments shown in Figures 2b to 7 are mentioned above.

Embodiments of the present application also provide a computer-readable storage medium. The computer-readable storage medium stores a program for signal processing. When it is run on a computer, it causes the computer to execute the embodiment shown in Figure 8. The steps performed by the execution device in the described method, or causing the computer to perform the steps shown in the aforementioned Figures 2b to 7 The example describes the steps performed by training the device in the method.

The neural network training device, data processing device, execution device or training device provided by the embodiment of the present application may specifically be a chip. The chip includes: a processing unit and a communication unit. The processing unit may be, for example, a processor. The communication unit For example, it can be an input/output interface, a pin or a circuit, etc. The processing unit can execute the computer execution instructions stored in the storage unit, so that the chip executes the neural network training method described in the embodiments shown in FIGS. 2b to 7, or to cause the chip in the training device to execute the neural network training method shown in FIG. 8. The embodiment describes the training method of neural network. Optionally, the storage unit is a storage unit within the chip, such as a register, cache, etc. The storage unit may also be a storage unit located outside the chip in the wireless access device, such as Read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (random access memory, RAM), etc.

Specifically, please refer to Figure 14. Figure 14 is a structural schematic diagram of a chip provided by an embodiment of the present application. The chip can be represented as a neural network processor NPU 140. The NPU 140 serves as a co-processor and is mounted to the main CPU (Host). CPU), tasks are allocated by the Host CPU. The core part of the NPU is the arithmetic circuit 1403. The arithmetic circuit 1403 is controlled by the controller 1404 to extract the matrix data in the memory and perform multiplication operations.

In some implementations, the computing circuit 1403 internally includes multiple processing units (Process Engine, PE). In some implementations, arithmetic circuit 1403 is a two-dimensional systolic array. The arithmetic circuit 1403 may also be a one-dimensional systolic array or other electronic circuit capable of performing mathematical operations such as multiplication and addition. In some implementations, arithmetic circuit 1403 is a general-purpose matrix processor.

For example, assume there is an input matrix A, a weight matrix B, and an output matrix C. The arithmetic circuit obtains the corresponding data of matrix B from the weight memory 1402 and caches it on each PE in the arithmetic circuit. The operation circuit takes matrix A data and matrix B from the input memory 1401 to perform matrix operations, and the partial result or final result of the matrix is stored in an accumulator (accumulator) 1408 .

The unified memory 1406 is used to store input data and output data. The weight data directly passes through the storage unit access controller (Direct Memory Access Controller, DMAC) 1405, and the DMAC is transferred to the weight memory 1402. Input data is also transferred to unified memory 1406 via DMAC.

BIU is the Bus Interface Unit, that is, the bus interface unit 1410, which is used for the interaction between the AXI bus and the DMAC and the Instruction Fetch Buffer (IFB) 1409.

The bus interface unit 1410 (Bus Interface Unit, BIU for short) is used to fetch the memory 1409 to obtain instructions from the external memory, and is also used for the storage unit access controller 1405 to obtain the original data of the input matrix A or the weight matrix B from the external memory.

DMAC is mainly used to transfer the input data in the external memory DDR to the unified memory 1406 or the weight data to the weight memory 1402 or the input data to the input memory 1401 .

The vector calculation unit 1407 includes multiple arithmetic processing units, and if necessary, further processes the output of the arithmetic circuit, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc. Mainly used for non-convolutional/fully connected layer network calculations in neural networks, such as Batch Normalization, pixel-level summation, upsampling of feature planes, etc.

In some implementations, vector calculation unit 1407 can store the processed output vectors to unified memory 1406 . For example, the vector calculation unit 1407 can apply a linear function and/or a nonlinear function to the output of the operation circuit 1403, such as linear interpolation on the feature plane extracted by the convolution layer, or a vector of accumulated values, to generate an activation value. In some implementations, vector calculation unit 1407 generates normalized values, pixel-wise summed values, or both. In some implementations, the processed output vector can be used as an activation input to the arithmetic circuit 1403, such as for use in a subsequent layer in a neural network.

The instruction fetch buffer 1409 connected to the controller 1404 is used to store instructions used by the controller 1404;

The unified memory 1406, input memory 1401, weight memory 1402 and instruction fetch memory 1409 are all On-Chip memories. External memory is private to the NPU hardware architecture.

Among them, the operations of each layer in the neural network shown in the above embodiments can be performed by the operation circuit 1403 or the vector calculation unit 1407.

The processor mentioned in any of the above places may be a general central processing unit, a microprocessor, an ASIC, or one or more integrated circuits used to control program execution of the method of the first aspect.

In addition, it should be noted that the device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physically separate. The physical unit can be located in one place, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the drawings of the device embodiments provided in this application, the connection relationship between modules indicates that there are communication connections between them, which can be specifically implemented as one or more communication buses or signal lines.

Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by software plus necessary general hardware. Of course, it can also be implemented by dedicated hardware including dedicated integrated circuits, dedicated CPUs, dedicated memories, Special components, etc. to achieve. In general, all functions performed by computer programs can be easily implemented with corresponding hardware. Moreover, the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits or special-purpose circuits. circuit etc. However, for this application, software program implementation is a better implementation in most cases. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product in essence or that contributes to the existing technology. The computer software product is stored in a readable storage medium, such as a computer floppy disk. , U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk, etc., including several instructions to cause a computer device (which can be a personal computer, training device, or network device, etc.) to execute the steps described in various embodiments of this application. method.

In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product.

The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another, for example, the computer instructions may be transferred from a website, computer, training device, or data The center transmits data to the network through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) to another website, computer, training device or data center. The computer-readable storage medium may be any available medium that a computer can store, or a data storage device such as a training device or a data center integrated with one or more available media. The available media may be magnetic media (eg, floppy disk, hard disk, magnetic tape), optical media (eg, DVD), or semiconductor media (eg, solid state disk (Solid State Disk, SSD)), etc.

Claims (20)

A neural network training method, characterized in that the method includes: Input the first training sample into the feature acquisition network, perform feature extraction on the first training sample through the feature acquisition network, and obtain the feature information of the first training sample; According to the characteristic information of the first training sample, the first characteristic information and the second characteristic information corresponding to the first training sample are generated through the characteristic acquisition network, and the first characteristic information corresponding to the first training sample is The first feature information includes features of information associated with the target attribute in the first training sample; Perform a classification operation according to the first feature information corresponding to the first training sample to obtain predicted category information, where the predicted category information indicates the predicted category of the first feature information corresponding to the first training sample, The predicted category is included in multiple categories corresponding to the target attribute; According to the first loss function, the feature acquisition network is trained to obtain a trained feature acquisition network; Wherein, the first loss function includes a first loss function term and a second loss function term, the first loss function term indicates the similarity between the predicted category information and the expected category information, and the expected category information indicates The correct category of the information associated with the target attribute in the first training sample. The purpose of using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information. Spend. The method of claim 1, further comprising: Combine the first feature information and the second feature information corresponding to the first training sample to obtain a first combined feature; Input the first combined features into the first classification network to obtain the first prediction decision information output by the first classification network corresponding to the first training sample; Training the feature acquisition network according to the first loss function includes: The feature acquisition network and the first classification network are trained according to the first loss function, wherein the first loss function also includes a third loss function term, and the third loss function term indicates the The similarity between the first predicted decision information and the expected decision information corresponding to the first training sample. The method according to claim 2, characterized in that, after training the feature acquisition network and the first classification network to obtain the trained feature acquisition network and the trained first classification network, The method also includes: Obtain third feature information, and combine the second feature information corresponding to the second training sample and the third feature information to obtain a second combined feature, wherein the third feature information and the third feature information are The data size of the first feature information corresponding to the second training sample is the same and the data content is different; Input the second combined features into the trained first classification network to obtain second prediction decision information corresponding to the second training sample output by the trained first classification network; Input the second combined features into a second classification network to obtain third prediction decision information output by the second classification network corresponding to the second training sample; According to the second loss function, the second classification network is trained to obtain a trained second classification network, and the trained feature acquisition network and the trained second classification network belong to the same target neural network; Wherein, the second loss function includes a fourth loss function term and a fifth loss function term, and the fourth loss function The term indicates the similarity between the second prediction decision information and the expected decision information corresponding to the second training sample, and the fifth loss function indicates the second prediction decision information and the third prediction decision information. similarity between them. The method according to claim 3, characterized in that said obtaining the third characteristic information includes: Generate the first feature information corresponding to the second training sample through the trained feature acquisition network; The first feature information and disturbance information corresponding to the second training sample are weighted and summed to obtain the third feature information, and the weight value of the disturbance information is adjustable. The method according to any one of claims 1 to 4, characterized in that the method further includes: Perform a classification operation according to the second feature information corresponding to the first training sample to obtain fourth prediction decision information corresponding to the first training sample, wherein the first loss function also includes a sixth loss function term, The sixth loss function term indicates the similarity between the fourth predicted decision information and the expected decision information corresponding to the first training sample, and the second loss function term indicates the similarity of the first loss function term. The similarity between the corresponding gradient and the gradient corresponding to the sixth loss function term. The method according to any one of claims 1 to 4, characterized in that the method is applied to any of the following scenarios: determining whether to recommend the object pointed by the first training sample, determining whether the object pointed to by the first training sample is Determine whether the object is in the target state or whether to agree to the request of the applicant pointed to by the first training sample. A data processing method, characterized in that the method includes: Input the data to be processed into a feature acquisition network, perform feature extraction on the data to be processed through the feature acquisition network, and obtain the feature information of the data to be processed; According to the characteristic information of the data to be processed, first characteristic information and second characteristic information are generated through the characteristic acquisition network, where the first characteristic information includes characteristics of the information associated with the target attribute in the data to be processed; Combine the first feature information and the second feature information to obtain a first combined feature; Input the first combined features into a classification network to obtain prediction decision information output by the classification network; Wherein, the feature acquisition network is trained using a first loss function. The first loss function includes a first loss function term and a second loss function term. The first loss function term indicates the relationship between predicted category information and expected category information. The predicted category information indicates the predicted category of the information associated with the target attribute in the data input to the feature acquisition network, and the expected category information indicates the predicted category of the information associated with the target attribute in the data input to the feature acquisition network. The correct category of the associated information, the purpose of using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information. The method according to claim 7, characterized in that the first loss function further includes a third loss function term, the third loss function term indicates the first prediction decision information and the data input to the feature acquisition network. Similarity between corresponding desired decision information, the first predicted decision information indicating a decision corresponding to the data input to the feature acquisition network. A neural network training device, characterized in that the device includes: A feature extraction module, configured to input the first training sample into a feature acquisition network, perform feature extraction on the first training sample through the feature acquisition network, and obtain feature information of the first training sample; a generation module, configured to generate first feature information and second feature information corresponding to the first training sample through the feature acquisition network according to the feature information of the first training sample, and the first feature information corresponding to the first training sample. The sample corresponds to The first feature information includes features of information associated with the target attribute in the first training sample; A first classification module configured to perform a classification operation based on the first feature information corresponding to the first training sample to obtain prediction category information, where the prediction category information indicates the third feature information corresponding to the first training sample. A predicted category of feature information, the predicted category being included in multiple categories corresponding to the target attribute; A training module, configured to train the feature acquisition network according to the first loss function to obtain a trained feature acquisition network; Wherein, the first loss function includes a first loss function term and a second loss function term, the first loss function term indicates the similarity between the predicted category information and the expected category information, and the expected category information indicates The correct category of the information associated with the target attribute in the first training sample. The purpose of using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information. Spend. The device of claim 9, further comprising: A combination module, configured to combine the first feature information and the second feature information corresponding to the first training sample to obtain a first combined feature; A second classification module, configured to input the first combined features into a first classification network to obtain the first prediction decision information output by the first classification network corresponding to the first training sample; The training module is specifically used to train the feature acquisition network and the first classification network according to the first loss function, wherein the first loss function also includes a third loss function term, and the The third loss function term indicates the similarity between the first predicted decision information and the expected decision information corresponding to the first training sample. The device according to claim 10, characterized in that the device further includes: An acquisition module, configured to acquire third feature information, wherein the data size of the third feature information and the first feature information corresponding to the second training sample are the same and the data content is different; The combination module is also used to combine the second feature information corresponding to the second training sample and the third feature information to obtain a second combined feature; The second classification module is also used to input the second combined features into the trained first classification network to obtain the output of the trained first classification network corresponding to the second training sample. second prediction decision information; The second classification module is also used to input the second combined features into a second classification network to obtain the third prediction decision information output by the second classification network corresponding to the second training sample; The training module is specifically used to train the second classification network according to the second loss function to obtain the trained second classification network, the trained feature acquisition network and the trained second classification network. The network belongs to the same target neural network; Wherein, the second loss function includes a fourth loss function term and a fifth loss function term, and the fourth loss function term indicates the relationship between the second prediction decision information and the expected decision information corresponding to the second training sample. The fifth loss function indicates the similarity between the second prediction decision information and the third prediction decision information. The device according to claim 11, characterized in that the acquisition module is specifically used for: Generate the first feature information corresponding to the second training sample through the trained feature acquisition network; The first feature information and disturbance information corresponding to the second training sample are weighted and summed to obtain the third feature information, and the weight value of the disturbance information is adjustable. The device according to any one of claims 9 to 12, characterized in that, The second classification module is also configured to perform a classification operation based on the second feature information corresponding to the first training sample to obtain fourth prediction decision information corresponding to the first training sample, wherein the first The loss function also includes a sixth loss function term, the sixth loss function term indicates the similarity between the fourth predicted decision information and the expected decision information corresponding to the first training sample, and the second loss function The term indicates the similarity between the gradient corresponding to the first loss function term and the gradient corresponding to the sixth loss function term. The device according to any one of claims 9 to 12, characterized in that the device is applied in any of the following scenarios: determining whether to recommend the object pointed by the first training sample, determining whether the object pointed to by the first training sample is Determine whether the object is in the target state or whether to agree to the request of the applicant pointed to by the first training sample. A data processing device, characterized in that the device includes: A feature extraction module, used to input the data to be processed into a feature acquisition network, perform feature extraction on the data to be processed through the feature acquisition network, and obtain the feature information of the data to be processed; Generating module, configured to generate first feature information and second feature information through the feature acquisition network according to the feature information of the data to be processed, where the first feature information includes the feature information associated with the target attribute in the data to be processed. Characteristics of information; A combination module, used to combine the first feature information and the second feature information to obtain the first combined feature; A classification module, configured to input the first combined features into a classification network to obtain prediction decision information output by the classification network; Wherein, the feature acquisition network is trained using a first loss function. The first loss function includes a first loss function term and a second loss function term. The first loss function term indicates the relationship between predicted category information and expected category information. The predicted category information indicates the predicted category of the information associated with the target attribute in the data input to the feature acquisition network, and the expected category information indicates the predicted category of the information associated with the target attribute in the data input to the feature acquisition network. The correct category of the associated information, the purpose of using the second loss function term for training includes reducing the similarity between the first feature information and the second feature information. The device according to claim 15, characterized in that the first loss function further includes a third loss function term, the third loss function term indicates the first prediction decision information and the data input to the feature acquisition network. Similarity between corresponding desired decision information, the first predicted decision information indicating a decision corresponding to the data input to the feature acquisition network. A computer program product, characterized in that the computer program product includes a program that, when the program is run on a computer, causes the computer to perform the method according to any one of claims 1 to 6, or causes the computer to The method of claim 7 or 8 is performed. A computer-readable storage medium, characterized in that a program is stored in the computer-readable storage medium. When the program is run on a computer, it causes the computer to execute the method described in any one of claims 1 to 6. method, or causing the computer to perform the method as claimed in claim 7 or 8. A training device, characterized by comprising a processor and a memory, the processor being coupled to the memory, The memory is used to store programs; The processor is configured to execute the program in the memory, so that the training device executes the method according to any one of claims 1 to 6. An execution device, characterized by comprising a processor and a memory, the processor being coupled to the memory, The memory is used to store programs; The processor is configured to execute the program in the memory, so that the execution device executes the method according to claim 7 or 8.