WO2019228358A1 - 深度神经网络的训练方法和装置 - Google Patents

深度神经网络的训练方法和装置 Download PDF

Info

Publication number
WO2019228358A1
WO2019228358A1 PCT/CN2019/088846 CN2019088846W WO2019228358A1 WO 2019228358 A1 WO2019228358 A1 WO 2019228358A1 CN 2019088846 W CN2019088846 W CN 2019088846W WO 2019228358 A1 WO2019228358 A1 WO 2019228358A1
Authority
WO
WIPO (PCT)
Prior art keywords
domain
data
target
training
sample data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/088846
Other languages
English (en)
French (fr)
Inventor
张炜晨
欧阳万里
徐东
李文
吴小飞
刘健庄
钱莉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Sydney
Huawei Technologies Co Ltd
Original Assignee
University of Sydney
Huawei Technologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Sydney, Huawei Technologies Co Ltd filed Critical University of Sydney
Priority to EP19812148.5A priority Critical patent/EP3757905A4/en
Publication of WO2019228358A1 publication Critical patent/WO2019228358A1/zh
Priority to US17/033,316 priority patent/US20210012198A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/211—Selection of the most significant subset of features
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning

Definitions

  • the present invention relates to the field of machine learning, and in particular, to a training method and device based on an adversarial network in the field of transfer learning.
  • Artificial intelligence is a theory, method, technology, and application system that uses digital computers or digital computer-controlled machines to simulate, extend, and extend human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.
  • artificial intelligence is a branch of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence.
  • Artificial intelligence is the study of the design principles and implementation methods of various intelligent machines, so that the machines have functions of perception, reasoning and decision-making.
  • Research in the field of artificial intelligence includes robotics, natural language processing, computer vision, decision and reasoning, human-computer interaction, recommendation and search, basic AI theory, and more.
  • Deep learning has been a key driving force for the development of the field of artificial intelligence in recent years, especially in the various tasks of computer vision, such as object classification / detection / recognition / segmentation, which has achieved impressive results; however, the deep learning Success depends on large amounts of labeled data.
  • labeling large amounts of data is an extremely time-consuming and labor-intensive task.
  • task models trained based on publicly available data sets or labeled data in the source domain can be directly applied to task prediction in the target domain.
  • the target domain is relative to the source domain, and the target domain Generally there is no labeled data or insufficient labeled data.
  • the publicly available datasets and labeled data in the source domain can be called source domain data.
  • the unlabeled data in the target domain can be called target domain data. . Because the distribution of the target domain data and the source domain data is not the same, the effect of directly using a model trained on the source domain data is not good.
  • Unsupervised domain adaptation is a typical transfer learning method that can be used to solve the above problems. Different from directly using the model trained on the source domain data for task prediction in the target domain, the unsupervised domain adaptation method not only uses the source domain data for training, but also fuses the unlabeled target domain data into the training to make the trained model It has better prediction effect on the target field data. At present, the unsupervised domain adaptation method with relatively good performance in the prior art is an unsupervised domain adaptation method based on domain adversarial. As shown in FIG.
  • the feature is that while learning the image classification task, the domain invariant feature is learned using the domain discriminator (full English name: Domain Discriminator) and the gradient direction (Gradient Reversal) method.
  • the main steps are: (1) Features extracted using a Convolutional Neural Network Feature Extractor (CNN Feature Extractor) are used in addition to the image classifier to build a domain classifier.
  • CNN Feature Extractor Convolutional Neural Network Feature Extractor
  • the domain classifier can You can output domain categories for the input features; (2) use the gradient inversion method to modify the gradient direction during the backpropagation, so that the features learned by the convolutional neural network feature extractor have domain invariance; (3) the above The convolutional neural network feature extractor and the obtained classifier are used for image classification prediction in the target domain.
  • the present application provides a training method based on a cooperative adversarial network, which can retain low-level features with domain discrimination, thereby improving the accuracy of the task model. It further provides a method for increasing collaborative domain confrontation, using the data in the target domain to train the task model, and improving the adaptability of the trained task model in the target domain.
  • the present application provides a training method for deep neural networks.
  • the training method is applied to the field of transfer learning, and specifically, a task model trained based on data in the source domain is applied to the prediction of data in the target domain.
  • the training method includes: Extract the low-level features and high-level features corresponding to the sample data in the source domain data and target domain data input to the deep neural network.
  • the target domain data is different from the source domain data, that is, the data distribution of the two is inconsistent; based on The high-level features of each sample data in the source domain data and the target domain data and the corresponding domain labels.
  • the first loss corresponding to each sample data is calculated by the first loss function; based on the lower layers of each sample data in the source domain data and the target domain data.
  • the second loss corresponding to each sample data is calculated by the second loss function; based on the high-level features of the sample data in the source field data and the corresponding sample labels, the third loss function is used to calculate the source field data.
  • the third loss corresponding to the sample data; First loss obtained in the above, loss of the second and third loss parameter update target depth of each module neural network.
  • the update is to update the parameters through loss back propagation. In the back propagation, the gradient of the first loss needs to go through the gradient reverse operation. The purpose of the gradient reverse operation is to realize the reverse conduction gradient to make the loss larger.
  • the high-level features can be made invariant and the low-level features can be distinguished from the domain. This improves the accuracy of the trained model applied to the target domain. .
  • the target deep neural network includes a feature extraction module, a task module, a domain invariant feature module, and a domain distinguishing feature module.
  • the feature extraction module includes at least one low-level feature network layer and a high-level feature network. Layer, at least one of the low-level feature network layers can be used to extract low-level features, the high-level feature network layer is used to extract high-level features, and the domain invariant feature module is used to enhance the high-level features extracted by the feature extraction module. Denaturation, domain discriminant feature module is used to enhance the domain discrimination of low-level features extracted by the feature extraction module;
  • the parameters for updating the target deep neural network according to the first loss, the second loss, and the third loss include: first calculating the total loss according to the first loss, the second loss, and the third loss; and then updating the feature extraction module based on the total loss.
  • the total loss can be the sum of the first loss, the second loss, and the third loss of a sample data. It may also be the sum of multiple first losses, multiple second losses, and multiple third losses of multiple sample data. Each loss is specifically used as a parameter of the corresponding module in the target neural network during the back propagation process.
  • the first loss updates the parameters of the domain invariant feature module and the feature extraction module through back propagation
  • the second loss updates the parameters of the domain distinguishing feature module and the feature extraction module are updated by back propagation.
  • the third loss updates the parameters of the task module and the feature extraction module through back propagation. The loss is generally obtained by further obtaining the corresponding gradient in the back-propagation to update the parameters of the relative module.
  • the first loss corresponding to each sample data is calculated through the first loss function. Including: inputting the high-level features of each sample data in the source domain data and the target domain data into the domain invariance feature module to obtain a first result corresponding to each sample data; according to the first domain data corresponding to each sample data in the source domain data and the target domain data A result and a corresponding field label are used to calculate a first loss corresponding to each sample data through a first loss function.
  • the second loss corresponding to each sample data is calculated through the second loss function, including: combining the source domain data and the target domain data.
  • the low-level feature input domain distinguishing feature module of each sample data obtains the second result corresponding to each sample data; according to the second result corresponding to each sample data in the source domain data and the target domain data and the corresponding domain label, the second loss is passed The function calculates the second loss corresponding to each sample data.
  • the third loss corresponding to the sample data in the source domain data is calculated by the third loss function, including: the high-level features of the sample data in the source domain data
  • the input task module obtains the third result corresponding to the sample data in the source domain data; based on the third result corresponding to the sample data in the source domain data and the corresponding sample label, the third loss function is used to calculate the corresponding sample data in the source domain data.
  • Third loss is used to calculate the corresponding sample data in the source domain data.
  • the domain invariance feature module further includes: a gradient inversion module; the training method further includes: performing gradient inversion on the gradient of the first loss through the gradient inversion module.
  • the gradient direction can realize the gradient of reverse conduction of the first loss, so that the calculation loss of the first loss function becomes larger, so that high-level features have domain-invariant features.
  • the training method further includes: inputting high-level features of the sample data in the target domain data into the task module to obtain corresponding prediction sample labels and corresponding confidence degrees; according to the samples in the target domain data Confidence of the data
  • the target field training sample data is selected from the target field data, and the target field training sample data is the sample data corresponding to the preset confidence condition in the target field data.
  • Using the target domain data for training the task model can further improve the classification accuracy of the task model on the data in the target domain.
  • the training method further includes: setting a weight of the training data of the target domain according to a first result corresponding to the training sample data of the target domain.
  • the distribution of the target domain training sample data is closer to the source domain image data and the target domain image data, which is more helpful for the training of the image classification model.
  • setting the weight can make the above-mentioned target domain training sample data that is not easily distinguished by the domain account for a larger weight in training.
  • setting the weight of the target domain training sample data according to the first result corresponding to the target domain training sample data includes: according to the similarity between the first result corresponding to the target domain training sample data and the domain label , Set the weight of the training data in the target domain, and the similarity indicates the difference between the first result and the domain label.
  • setting the weight of the target domain training sample data includes: calculating the first corresponding to the target domain training sample data. A first difference between the result and the domain label of the source domain, and a second difference between the first result corresponding to the training sample data of the target domain and the domain label of the target domain; if the absolute value of the first difference is greater than the second difference.
  • the absolute value of the target field training data is set to a small value, such as a value less than 0.5; otherwise, the target field training sample data is set to a larger value, such as a value greater than 0.5.
  • the target field training sample data is set. Is the maximum weight (for example, 1).
  • the first field has a label value of 0, the second field has a label value of 1, and the middle value refers to 0.5 or a value in a floating range of 0.5.
  • the first field label value is the value corresponding to the field label of the source field, and the second field label value is the value corresponding to the field label of the target field.
  • the training method before the above-mentioned training sample data of the target domain is selected from the target domain data based on the confidence corresponding to the sample data in the target domain data, the training method further includes: according to the accuracy of the task model
  • the adaptive threshold is set.
  • the task model includes a feature extraction module and a task module.
  • the adaptive threshold is positively related to the accuracy of the task model.
  • the preset condition is that the confidence is greater than or equal to the adaptive threshold.
  • the adaptive threshold is calculated by the following logical function:
  • T c is an adaptive threshold
  • A is the accuracy of the task model
  • ⁇ c is a hyperparameter used to control the inclination of the logic function.
  • the training method further includes: extracting low-level features and high-level features of target sample training sample data through a feature extraction module; based on high-level features and corresponding field labels of target sample training sample data, The first loss function calculates the first loss corresponding to the training data in the target field; based on the low-level features of the training data in the target field and the corresponding field labels, the second loss corresponding to the training data in the target field is calculated through the second loss function; based on the target High-level features of the domain training sample data and corresponding prediction sample labels.
  • the third loss function is used to calculate the third loss corresponding to the training data in the target domain; according to the first loss, the second loss, and the third loss corresponding to the training data in the target domain Calculate the total loss corresponding to the training data in the target field, where the gradient of the first loss corresponding to the training data in the target field is reversed by the gradient; update according to the total loss corresponding to the training data in the target field and the weight of the training data in the target field.
  • the parameter extraction module, the task module parameters, and the parameter field distinguishing feature domain invariant feature module module parameters.
  • calculating the first loss corresponding to the target domain training sample data through the first loss function includes: training the target domain
  • the high-level feature input domain invariance feature module of the sample data obtains the first result corresponding to the training data of the target domain; according to the first result corresponding to the training data of the target domain and the corresponding field label, the training sample of the target domain is calculated by the first loss function The first loss corresponding to the data;
  • calculating the second loss corresponding to the target domain training sample data through the second loss function includes: inputting the low-level features of the target domain training sample data into the domain distinguishing feature module to obtain A second result corresponding to the training data in the target field; and a second loss corresponding to the training data in the target field according to the second result corresponding to the training data in the target field and the corresponding field label;
  • a third loss function is used to calculate the third loss corresponding to the target domain training sample data, including: entering the high-level features of the target domain training sample data into the task module to obtain the target domain The third result corresponding to the training sample data; based on the third result corresponding to the training sample data in the target domain and the corresponding prediction sample label, a third loss corresponding to the training sample data in the target domain is calculated by a third loss function.
  • the present application provides a training device.
  • the training device includes a memory and a processor coupled to the memory; the memory is used to store instructions, and the processor is used to execute instructions; wherein, when the processor executes the instructions, the first aspect is performed And the method described in the possible implementation of the first aspect.
  • the present application provides a computer-readable storage medium, where the computer-readable storage stores a computer program that, when executed by a processor, implements the first aspect described above and a possible implementation manner of the first aspect method.
  • the present application provides a computer program product including code for performing the methods described in the first aspect and possible implementations of the first aspect.
  • the present application provides a training device including a functional unit for performing the foregoing first aspect and the method described in a possible implementation manner of the first aspect.
  • the present application provides an enhanced cooperative adversarial network constructed based on a convolutional neural network CNN.
  • the enhanced cooperative adversarial network includes low-level features and high-level features for extracting sample data from source domain data and target domain data.
  • Feature extraction module the data distribution of the target domain data and the source domain data is different;
  • a task module for receiving the high-level features output by the feature extraction module and calculating the third loss corresponding to each sample data through the third loss function, the third loss
  • a domain invariance module for updating the parameters of the feature extraction module and the task module; for receiving high-level features output by the feature extraction module and calculating the first loss corresponding to each sample data through the first loss function; the first loss is used for updating
  • the parameters of the feature extraction module and the domain invariance module make the high-level features output by the feature extraction module have domain invariance; used to receive the low-level features output by the feature extraction module and calculate the second loss corresponding to each sample data through the second loss function.
  • Domain discrimination module the second loss is used
  • the enhanced cooperative adversarial network further includes: a sample data selection module for selecting target training data from the target domain data from the target domain data according to the confidence corresponding to the sample data in the target domain data,
  • the confidence level corresponding to the sample data in the target domain data is obtained by inputting high-level features of the sample data in the target domain data into the task module.
  • the target domain training sample data is the sample data whose corresponding confidence level in the target domain data satisfies a preset condition.
  • the sample data selection module is further configured to set an adaptive threshold according to the accuracy of the task model.
  • the task model includes a feature extraction module and a task module.
  • the adaptive threshold is positively related to the accuracy of the task model. ; Wherein, the preset condition is that the confidence is greater than or equal to the adaptive threshold.
  • the enhanced cooperative adversarial network further includes a weight setting module for setting a weight of the target domain training sample data according to a first result corresponding to the target domain training sample data.
  • the weight setting module is specifically configured to set the weight of the training data of the target field according to the similarity between the first result corresponding to the training data of the target field and the field label; The difference between a result and the field label.
  • the weight setting module is specifically configured to calculate a first difference between a first result corresponding to training sample data in a target domain and a domain label of a source domain, and a corresponding value of training sample data in the target domain.
  • the second difference between the first result and the domain label of the target domain; if the absolute value of the first difference is greater than the absolute value of the second difference, then set the weight of the training sample data in the target domain to a smaller value, otherwise, set The weight of the training data of the target domain is a large value.
  • the foregoing weight setting module is specifically configured to: if the first result corresponding to the training data in the target domain is an intermediate value in a range from the first domain label value to the second domain label value range , Set the weight of the training sample data in the target domain to the maximum value, for example 1, the first domain label value is the value corresponding to the domain label of the source domain, and the second domain label value is the value corresponding to the domain label of the target domain.
  • the intermediate value refer to the related description of the first aspect, which is not repeated here.
  • the present application provides a method for setting training data weights based on a cooperative adversarial network.
  • the cooperative adversarial network includes at least a feature extraction module, a task module, a domain invariance module, and may also include a domain discrimination module. Reference may be made to the related description of the sixth aspect above, which is not repeated here.
  • the weight setting method includes: inputting high-level features of sample data in the target domain data into a task module to obtain corresponding prediction sample labels and corresponding confidence degrees; and selecting targets from the target domain data according to the corresponding confidence degrees of the sample data in the target domain data.
  • the target domain training sample data is the corresponding sample data in the target domain data whose confidence level meets the preset conditions; the high-level features of the sample data in the target domain data are input into the domain invariance module to obtain the target domain training sample data corresponding The first result of; the weight of the training data of the target domain is set according to the first result corresponding to the training data of the target domain.
  • setting the weight of the target domain training sample data according to the first result corresponding to the target domain training sample data specifically includes: according to the first domain corresponding to the target domain training sample data, the first result is similar to the domain label. Degree, which sets the weight of the training data in the target domain, and the similarity indicates the difference between the first result and the domain label.
  • setting the weight of the target domain training sample data includes: calculating the first corresponding to the target domain training sample data. A first difference between the result and the domain label of the source domain, and a second difference between the first result corresponding to the training sample data of the target domain and the domain label of the target domain; if the absolute value of the first difference is greater than the second difference.
  • the absolute value of the target field training data is set to a small value, such as a value less than 0.5; otherwise, the target field training sample data is set to a larger value, such as a value greater than 0.5.
  • the target field training sample data is set Is the maximum weight (for example, 1).
  • the first field has a label value of 0, the second field has a label value of 1, and the middle value refers to 0.5 or a value in a floating range of 0.5.
  • the first field label value is the value corresponding to the field label of the source field, and the second field label value is the value corresponding to the field label of the target field.
  • the weight setting method before selecting the target field training sample data from the target field data according to the confidence level corresponding to the sample data in the target field data, further includes: The accuracy sets an adaptive threshold.
  • the task model includes a feature extraction module and a task module.
  • the adaptive threshold is positively related to the accuracy of the task model.
  • the preset condition is that the confidence is greater than or equal to the adaptive threshold.
  • the above adaptive threshold is calculated by the following logical function:
  • T c is an adaptive threshold
  • A is the accuracy of the task model
  • ⁇ c is a hyperparameter used to control the inclination of the logic function.
  • the present application provides a device including a memory and a processor coupled to the memory; the memory is used to store instructions, and the processor is used to execute instructions; wherein, when the processor executes the instructions, the seventh aspect and the first aspect are executed. Methods described in seven possible implementations.
  • the present application provides a computer-readable storage medium, where the computer readable stores a computer program, and the computer program, when executed by a processor, implements the seventh aspect and the possible implementation manners described in the seventh aspect. method.
  • the present application provides a computer program product including code for performing the methods described in the seventh aspect and the possible implementation manners of the seventh aspect.
  • the present application provides a weight setting device, and the weight setting device includes a functional unit for performing the methods described in the seventh aspect and the possible implementation manners of the seventh aspect.
  • the training method provided in the embodiment of the present application establishes a domain invariance loss function and a domain discriminative loss function based on the high-level features and the low-level features, respectively, while ensuring the domain-invariant features of the high-level features while retaining the domain distinguishing features in the low-level features. , Can improve the accuracy of the training task model applied to the target domain for prediction.
  • FIG. 1 is a schematic diagram of a method for adaptively training an image classifier based on an unsupervised domain according to an embodiment of the present invention
  • FIG. 2 is a schematic diagram of an artificial intelligence main body frame provided by an embodiment of the present invention.
  • FIG. 3 is a schematic diagram of comparison of image data of people and vehicles in different cities according to an embodiment of the present invention
  • FIG. 4 is a schematic diagram of face image data comparison in different regions according to an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of a training system architecture according to an embodiment of the present invention.
  • FIG. 6 is a schematic diagram of a feature extraction unit according to an embodiment of the present invention.
  • FIG. 7 is a schematic diagram of a feature extraction CNN provided by an embodiment of the present invention.
  • FIG. 8 is a schematic diagram of a domain invariant feature unit according to an embodiment of the present invention.
  • FIG. 9 is a schematic structural diagram of a training device according to an embodiment of the present invention.
  • FIG. 10 is a schematic structural diagram of another training device according to an embodiment of the present invention.
  • FIG. 11 is a schematic diagram of a cloud-end system architecture according to an embodiment of the present invention.
  • FIG. 12 is a flowchart of a training method according to an embodiment of the present invention.
  • FIG. 13 is a schematic diagram of a training method based on a cooperative adversarial network according to an embodiment of the present invention.
  • FIG. 14 is a schematic diagram of a weight setting curve provided by an embodiment of the present invention.
  • FIG. 15 is a schematic diagram of a chip hardware structure according to an embodiment of the present invention.
  • 16 is a schematic structural diagram of a training device according to an embodiment of the present invention.
  • FIG. 17A is a test result on Office-31 provided by an embodiment of the present invention.
  • FIG. 17B is a test result on ImageCLEF-DA according to an embodiment of the present invention.
  • FIG. 2 shows a schematic diagram of an artificial intelligence main body frame, which describes the overall workflow of the artificial intelligence system and is suitable for general artificial intelligence field requirements.
  • 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 undergone the condensed process of "data-information-knowledge-wisdom".
  • the "IT value chain” reflects the value that artificial intelligence brings to the information technology industry, from the low-level infrastructure of human intelligence, information (the provision and processing technology implementation) to the system's industrial ecological process.
  • Infrastructure provides computing power support for artificial intelligence systems, enables communication with the outside world, and supports it through basic platforms.
  • sensors communicate with external sources to obtain data, which is provided to smart chips in the distributed computing system provided by the basic platform for calculation.
  • the data in the upper layer of the infrastructure is used to represent data sources in the field of artificial intelligence.
  • the data involves graphics, images, voice, text, and IoT data of traditional devices, including business data of existing systems and perceptual data such as force, displacement, liquid level, temperature, and humidity.
  • Data processing usually includes data training, machine learning, deep learning, search, reasoning, decision making and other methods.
  • machine learning and deep learning can symbolize and formalize data for intelligent information modeling, extraction, preprocessing, training, and so on.
  • Reasoning refers to the process of simulating human's intelligent reasoning in a computer or an intelligent system, using formal information to perform machine thinking and solving problems according to inference control strategies. Typical functions are search and match.
  • Decision-making refers to the process of making decisions after intelligent information is inferred, and usually provides functions such as classification, ranking, and prediction.
  • some general capabilities can be formed based on the results of data processing, such as algorithms or a general system, such as translation, text analysis, computer vision processing, speech recognition, and images. Identification and so on.
  • Intelligent products and industry applications refer to the products and applications of artificial intelligence systems in various fields. They are the packaging of the overall artificial intelligence solution, productizing intelligent information decision-making, and implementing applications. Its application areas include: intelligent manufacturing, intelligent transportation, Smart home, smart medical, smart security, autonomous driving, safe city, smart terminal, etc.
  • Unsupervised domain adaptation is a typical method of transfer learning.
  • Task models are trained based on the data in the source and target domains.
  • the trained task models are used to implement recognition, classification, segmentation, and detection of objects in the target domain. , Where the data in the source domain is labeled and the data in the target domain is unlabeled, and the distribution of the data in the two domains is different. It should be noted that, in this application, "data in the source domain” and “data in the source domain”, “data in the target domain” and “data in the target domain” usually have the same meaning.
  • Domain-invariant features refer to the common features of data in different domains, and the features extracted from data in different domains have a consistent distribution.
  • Domain distinguishing features Refers to the features in the data in a specific domain, and the features extracted from the data in different domains have different distributions.
  • This application describes a training method for a neural network, which is applied to the training of a task / prediction model (hereinafter referred to as a task model) in the field of transfer learning. Specifically, it can be applied to training various task models built based on deep neural networks, including but not limited to classification models, recognition models, segmentation models, and detection models.
  • the task model obtained through the training method described in this application can be widely applied to a variety of specific application scenarios such as AI photography, autonomous driving, safe cities, etc., to achieve intelligent application scenarios.
  • the detection of people and vehicles is a basic unit in an automatic driving perception system.
  • the accuracy of human and vehicle detection is related to the safety of autonomous vehicles. Whether the pedestrians and vehicles around the vehicle can be accurately detected depends on whether the detection model for human and vehicle detection has high accuracy.
  • the high accuracy detection model depends on Extensive labeled car / vehicle image / video data. Labeling data is another huge project. In order to achieve the accuracy requirements of autonomous driving, it is almost necessary to label different data for different cities, which is difficult to achieve.
  • the migration of human and vehicle detection models is the most commonly used method, that is, the detection model trained based on the labeled human / car image / video data in area A is directly applied to the vehicle with no or insufficient labeled people Person / car detection in scene B of image / video data, where area A is the source area, area B is the target area, data in area A is the source area data with labels, and data in area B is the target area without labels data.
  • area A is the source area
  • area B is the target area
  • data in area A is the source area data with labels
  • data in area B is the target area without labels data.
  • the race, living habits, architectural style, climatic environment, transportation facilities, and data collection equipment of different cities may vary greatly, that is, the distribution of data is different, and it is difficult to guarantee the autonomous driving. Precision required.
  • the four images on the left collect image data from a collection device in a city in Europe, and the four images on the right are image data collected by a collection device in a city in Asia. It can be seen that pedestrian skin, clothing, and posture There are obvious differences, and there are also obvious differences in the appearance of urban buildings and traffic. If the detection model trained on the image / video data of one city in FIG. 3 is applied to another city scene in FIG. 3, the accuracy of the detection model will inevitably be greatly reduced.
  • the training method described in this application uses labeled data and unlabeled data to jointly train a task model, that is, jointly uses labeled human / vehicle image / video data in area A and labeled human / vehicle image / video data in area B to train together.
  • the detection model for the detection of people and vehicles can greatly improve the accuracy of the detection model trained based on the image / video data of people and vehicles in area A when applied to the detection of people and vehicles in area B.
  • face recognition often involves the identification of people in different countries and regions, and the face data of people in different countries and regions will have large distribution differences.
  • the European Caucasian face data has training labels as the source domain data, that is labeled face data
  • the African black people's face data without training labels as the target domain data that is, not labeled Face data. Due to the large differences in skin colors and facial contours of white and black people, the distribution of face data is different; however, even if the face data of black people is unlabeled data, the training method described in this application
  • the obtained face recognition model can also improve the accuracy of face recognition for black people.
  • An embodiment of the present invention provides a deep neural network training system architecture 100.
  • the system architecture 100 includes at least a training device 110 and a database 120, and further includes a data acquisition device 130, a client device 140, and a data storage system 150.
  • the data collection device 130 is configured to collect data and store the collected data (for example, pictures / videos / audios) into the database 120 as training data.
  • the database 120 is used to maintain and store training data.
  • the training data stored in the database 120 includes source domain data and target domain data.
  • the source domain data can be understood as labeled data
  • the target domain data can be understood as unlabeled data.
  • the source domain and The target field is a relative concept of the transfer learning field. For details, see FIG. 3 and FIG. 4 for a description of understanding the source field, the target field, the source field data, and the target field data. The above concepts can be understood by those skilled in the art.
  • the training device 110 interacts with the database 120 and obtains required training data from the database 120 for training a task model.
  • the task model includes a feature extraction module and a task module.
  • the feature extraction module may be the feature extraction unit 111, or may be used after training.
  • a deep neural network constructed by the parameters of the feature extraction unit 111; similarly, the task module may be the task unit 112, or a model constructed using the parameters of the trained task unit 112, such as a function model, a neural network model, and the like.
  • the task model obtained by the training device 110 through training may be applied to the client device 140, and may also output a prediction result in response to the client device 140 request.
  • the client device 140 is an autonomous driving vehicle, and the training device 110 trains a human-vehicle detection model according to the training data in the database 120.
  • the human-vehicle detection model obtained by the training device 110 can complete the person.
  • the vehicle is inspected and fed back to the autonomous vehicle.
  • the trained person-car detection model can be arranged on the autonomous vehicle or in the cloud.
  • the specific form is not limited.
  • the client device 140 can also be used as a data collection device for the database 120 to expand the database when needed.
  • the training device 110 includes a feature extraction unit 111, a task unit 112, a domain invariant feature unit 113, a domain distinguishing feature unit 114, and an I / O interface 115.
  • the / O interface 115 is used for the training device 110 to interact with external devices.
  • the feature extraction unit 111 is used to extract low-level features and high-level features of the input data.
  • the feature extraction single unit 111 includes a low-level feature extraction sub-unit 1111 and a high-level feature extraction sub-unit 1112.
  • the low-level feature extraction sub-unit 1111 is used for
  • the high-level feature extraction subunit 1112 is used to extract the high-level features of the input data.
  • the data is input to the low-level feature extraction sub-unit 1111 to obtain data representing low-level features
  • the data representing the low-level features is input to the high-level feature extraction sub-unit 1112 to obtain data representing high-level features, that is, the high-level features are further based on the low-level features. Processed features.
  • the feature extraction unit 111 may be implemented by software, hardware (for example, a circuit) or a combination of software and hardware (for example, a processor call code). It is common to implement the function of the feature extraction unit 111 through a neural network.
  • the function of the feature extraction unit 111 is implemented by a Convosutionas Neuras Network (CNN). As shown in FIG. 7, the feature extraction CNN includes multiple Convolution layer. Feature extraction of input data can be achieved through convolution calculations.
  • the last convolution layer of multiple convolution layers can be called a high-level convolution layer, as a high-level feature extraction subunit 1112 for extracting high-level features; other
  • the convolutional layer may be called a low-level convolutional layer, and as a low-level feature extraction subunit 1111 is used to extract low-level features.
  • Each low-level convolutional layer can output a low-level feature, that is, after a data input is used as the CNN of the feature extraction unit 111, a high-level feature and at least one low-level feature can be output.
  • the number of low-level features can be set according to the actual training needs and formulated.
  • the specific output is used as a low-level feature convolutional layer for outputting low-level features as the low-level feature extraction subunit 1111.
  • Convosutionas Neuras Network is a deep neural network with a convolutional structure.
  • the convolutional neural network includes a feature extractor composed of a convolutional layer and a sub-sampling layer.
  • the feature extractor can be regarded as a filter, and the convolution process can be regarded as a convolution using a trainable filter and an input image or a convolution feature map.
  • a convolution layer is a neuron layer in a convolutional neural network that performs convolution processing on input signals. In the convolutional layer of a convolutional neural network, a neuron can only be connected to some of the neighboring layer neurons.
  • a convolution layer usually contains several feature planes, and each feature plane can be composed of some rectangularly arranged neural units.
  • Neural units in the same feature plane share weights, and the weights shared here are convolution kernels. Sharing weights can be understood as the way of extracting image information has nothing to do with location. The underlying principle is that the statistical information of one part of the image is the same as the other parts. That means that the image information learned in one part can also be used in another part. So for all positions on the image, we can use the same learned image information. In the same convolution layer, multiple convolution kernels can be used to extract different image information. Generally, the more the number of convolution kernels, the richer the image information reflected by the convolution operation.
  • the convolution kernel can be initialized in the form of a matrix of random size. During the training process of the convolutional neural network, the convolution kernel can obtain reasonable weights through learning. In addition, the direct benefit of sharing weights is to reduce the connections between the layers of the convolutional neural network, while reducing the risk of overfitting.
  • Convolutional neural networks can use the backpropagation (BP) algorithm to modify the size of the parameters in the initial super-resolution model during the training process, which makes the reconstruction error loss of the super-resolution model smaller and smaller.
  • BP backpropagation
  • the input signal is forwardly transmitted until the output will generate an error loss, and the parameters in the initial super-resolution model are updated by back-propagating the error loss information, thereby converging the error loss.
  • the back-propagation algorithm is a back-propagation motion dominated by error loss, and aims to obtain the optimal parameters of the super-resolution model, such as the weight matrix.
  • the input of the task unit 112 is the high-level features output by the high-level feature extraction sub-unit 1112, specifically the high-level features output by the labeled source domain data through the feature extraction unit 111, and the output is a label.
  • the trained task unit 112 and the feature extraction unit 111 can be used as a task model, and the task model can be used for prediction tasks in the target domain.
  • the input of the domain-invariant feature unit 113 is a high-level feature output by the high-level feature extraction sub-unit 1112, and the output is a field (source field or target field) label to which the corresponding data belongs.
  • the domain invariance feature unit 113 includes a domain distinguishing feature subunit 1131 and a gradient inversion subunit 1132.
  • the gradient inversion subunit 1132 can perform gradient inversion on the back-propagated gradient, so that the domain distinguishes feature subunits.
  • the error (ie, loss) between the field label and the real field label output by the unit 1131 becomes larger.
  • the domain invariance feature unit 113 can realize that the high-level features output by the feature extraction unit 111 have domain invariance, that is, it is difficult to reduce the high-level features output by the feature extraction unit 111 or it is impossible to distinguish the domains.
  • the input of the domain distinguishing feature unit 114 is the low-level feature output by the low-level feature extraction sub-unit 1111, and the output is the domain label to which the corresponding data belongs.
  • the domain distinguishing feature unit 114 can make the low-level features output by the feature extracting unit 111 easily distinguish the domain, thereby being domain distinguishable.
  • both the domain distinguishing feature unit 114 and the domain distinguishing feature subunit 1131 can target the domain to which the input feature output belongs.
  • a gradient inversion sub-unit 1132 is also included.
  • the domain distinguishing feature unit 114 and the feature extracting unit 111 can constitute a domain distinguishing model.
  • the gradient inversion subunit 1132 is ignored, and the domain distinguishing feature subunit 1131 and the feature extracting unit 111 in the domain invariant feature unit 113 can also be used. Form a domain differentiation model.
  • the training device 110 has the structure shown in FIG. 9.
  • the training device 110 includes a feature extraction unit 111, a task unit 112, a domain distinguishing feature unit 113 ′, a gradient inversion unit 114 ′, and an I / O interface 115.
  • the domain distinguishing feature unit 113 ′ and the gradient inversion unit 114 ′ are equivalent to the domain invariant feature unit 113 and the domain distinguishing feature unit 114 of the training device 110 in FIG. 5.
  • the task unit 112, the domain invariant feature unit 113 and the domain distinguishing feature unit 114, and the domain distinguishing feature unit 113 'and the gradient inversion unit 114' may be called by software, hardware (for example, a circuit), or software and hardware (for example, a processor). Code) combined implementation, which can be implemented by vector matrices, functions, neural networks, etc. without limitation.
  • the task unit 112, the domain invariant feature unit 113, and the domain distinguishing feature unit 114 all include a loss function for calculating the loss of the output value and the true value, and the loss is used to update the parameters in each unit. The specific update details are in the technical field. As far as the technical staff can understand, I won't go into details.
  • the training device 110 includes a domain invariant feature unit 113 and a domain distinguishing feature unit 114.
  • the low-level features output by the feature extraction unit 111 can be distinguished by the domain, and the output high-level features. It has domain invariance, and the high-level features are further obtained based on the low-level features, so that the high-level features can still retain the domain-distinctive features, and further use in the task model can improve the prediction accuracy.
  • the training device 110 further includes a sample data selection unit 116.
  • the sample data selection unit 116 is configured to select data that meets the conditions from the target domain data as training sample data for training performed by the training device 110.
  • the sample data selection unit 116 specifically includes a selection subunit 1161 and a weight setting subunit 1162.
  • the selection subunit 1161 is configured to select data that meets the conditions from the target domain data according to the accuracy of the task model and add corresponding labels as training sample data.
  • the weight setting subunit 1162 is used to set weights on the selected target domain data as training sample data, and determine the degree of influence of the target domain data as training sample data on the training of the task model by setting the weights. How to select and set weights will be described in detail below, and will not be repeated here.
  • the other units in FIG. 10 include the feature extraction unit 111, the task unit 112, the domain invariant feature unit 113, the domain distinguishing feature unit 114, and the I / O interface 115 in FIG. 5, or the feature extraction unit 111. , Task unit 112, domain distinguishing feature unit 113 ', gradient inversion unit 114', and I / O interface 115.
  • An embodiment of the present invention provides a cloud-end system architecture 200.
  • the execution device 210 is implemented by one or more servers, and optionally, cooperates with other computing devices, such as data storage, routers, and loads. Equipment such as an equalizer; the execution device 210 may be arranged on one physical site, or distributed on multiple physical sites.
  • the execution device 210 may use data in the data storage system 220 or call program code in the data storage system 220 to implement all functions of the training device 110.
  • the execution device 210 may train according to the training data in the database 120 A task model, and a task prediction of a target domain is completed according to a request from a local device 231 (232).
  • the execution device 210 does not have the training function of the training device 110, but can complete prediction based on the task model trained by the training device 110.
  • the execution device 210 is configured with the training device 110 to train the task model, and then receives After the request from the local device 231 (232), the prediction is completed and the result is fed back to the local device 231 (232).
  • the user can operate respective user devices (for example, the local device 231 and the local device 232) to interact with the execution device 210.
  • Each local device can represent any computing device, such as a personal computer, computer workstation, smartphone, tablet, smart camera, smart car or other type of cell phone, media consumer device, wearable device, set-top box, game console, and so on.
  • the local device of each user can interact with the execution device 210 through a communication network of any communication mechanism / communication standard.
  • the communication network may be a wide area network, a local area network, a point-to-point connection, or any combination thereof.
  • one or more aspects of the execution device 210 may be implemented by each local device.
  • the local device 301 may provide local data or feedback calculation results to the execution device 210.
  • the execution device 210 may also be implemented by a local device.
  • the local device 231 implements functions (eg, training or prediction) of the device 210 and provides services to its own users, or provides services to users of the local devices 232.
  • the embodiment of the present application provides a training method of a target deep neural network.
  • the target deep neural network is a collective name of a system architecture, and specifically includes a feature extraction module (corresponding to the feature extraction unit 111) and a task module (corresponding to the task unit 112). ), Domain invariant feature module (corresponding to domain invariant feature unit 113) and domain distinguishing feature module (corresponding to domain distinguishing feature unit 114 or domain distinguishing feature unit 113 '), the feature extraction module includes at least one low-level feature network layer (Corresponding to the low-level feature extraction sub-unit 1111) and high-level feature network layer (corresponding to the high-level feature extraction sub-unit 1112).
  • any one of the at least one low-level feature network layer can be used to extract low-level features, and the high-level feature network layer is used to The high-level features are extracted.
  • the domain-invariant feature module is used to enhance the domain-invariance of the high-level features extracted by the feature extraction module, and the domain-distinctive feature module is used to enhance the domain-disturbance of the low-level features extracted by the feature extraction module.
  • the specific steps of this training method are:
  • the low-level feature network layer is used to extract low-level features corresponding to each sample data in the source domain data and the target domain data
  • the high-level feature network layer is used to extract and extract high-level features corresponding to each sample data in the source domain data and the target domain data.
  • the first loss corresponding to each sample data is calculated through the first loss function; specifically, the source domain data and the target domain data are calculated.
  • the high-level feature input domain invariance feature module of each sample data in the sample data obtains the first result corresponding to each sample data; according to the first result corresponding to each sample data in the source domain data and the target domain data and the corresponding domain label, A loss function calculates the first loss corresponding to each sample data.
  • the domain invariance feature module further includes: a gradient inversion module (corresponding to the gradient inversion subunit); the training method further includes: performing gradient inversion processing on the gradient of the first loss through the gradient inversion module, and gradient inversion Orientation can use any existing technology, such as Gradient Reversal Layer (GRL).
  • GRL Gradient Reversal Layer
  • the low-level features of each sample data in the source domain data and the target domain data are input into the domain distinguishing feature module to obtain a second result corresponding to each sample data; according to each sample data in the source domain data and the target domain data, The second result and the corresponding field label are used to calculate the second loss corresponding to each sample data through the second loss function.
  • the high-level features of the sample data in the source domain data are input into the task module to obtain a third result corresponding to the sample data in the source domain data; based on the third result corresponding to the sample data in the source domain data and the corresponding sample label, A third loss corresponding to the sample data in the source domain data is calculated by a third loss function.
  • the total loss is calculated according to the first loss, the second loss, and the third loss;
  • the parameters of the feature extraction module, the parameters of the task module, the parameters of the domain invariant feature module, and the parameters of the domain distinguishing feature module are updated according to the total loss.
  • the trained feature extraction module and task module are used as task models for prediction tasks in the target domain.
  • prediction tasks in the source domain can also be used.
  • the training method further includes the following steps:
  • the high-level features of the sample data in the target domain data are input into the task module, and corresponding prediction sample labels and corresponding confidence degrees are obtained.
  • the target domain training sample data refers to sample data whose corresponding confidence in the target domain data satisfies a preset condition
  • the adaptive threshold is set according to the accuracy of the task model.
  • the task model includes a feature extraction module and a task module.
  • the adaptive threshold is positively related to the accuracy of the task model.
  • the preset condition means that the confidence is greater than or equal to the adaptive threshold.
  • the adaptive threshold is calculated by the following logical function:
  • T c is an adaptive threshold
  • A is the accuracy of the task model
  • ⁇ c is a hyperparameter used to control the inclination of the logic function.
  • the similarity between the source domain data or the target domain data distribution is determined, and the target domain sample weight is set according to the similarity.
  • the similarity can be expressed by the difference between the predicted value and the domain label.
  • a value is set for each of the source domain label and the target domain label in advance, for example, the source domain label (may be referred to as the source domain label) is set to a, and the target domain label (may be referred to as the target domain label) is set.
  • Is b then the range of the predicted value x is between a and b.
  • the degree of similarity can be determined according to the size of
  • weight setting There are two options for weight setting: (1) When the predicted value is closer to the value of the source domain, set a smaller weight; if the predicted value is between the value of the source domain label and the value of the target domain label, set a larger weight . (2) When the predicted value is closer to the value of the source domain label, set a smaller weight; if the output value is closer to the value of the target domain label, set a larger weight.
  • the foregoing smaller weights and larger weights are relative, and specific values can be determined according to actual settings.
  • the relationship between the weight and the similarity can be simply summarized as: the predicted value is more inclined to the source field label value, and the corresponding weight is inclined to a smaller value. That is, it is more likely that the corresponding target domain training sample data is data of the source domain according to the predicted value, then set the target domain training sample data weight to a smaller value, otherwise a larger value can be set.
  • the predicted value is more inclined to the source field label value
  • the corresponding weight is inclined to a smaller value. That is, it is more likely that the corresponding target domain training sample data is data of the source domain according to the predicted value, then set the target domain training sample data weight to a smaller value, otherwise a larger value can be set.
  • the training sample data of the target domain selected according to steps S106-S108 also includes prediction sample labels and weights.
  • the selected target domain training sample data can be used for training, that is, equivalent to the source domain data, and then go through step S101- S105.
  • the training method further includes the steps of training sample data for the target domain, as follows:
  • the first loss corresponding to the target domain training sample data is calculated by the first loss function; specifically, the high-level features of the target domain training sample data are input to the domain invariance
  • the feature module obtains a first result corresponding to the training sample data in the target domain; and calculates a first loss corresponding to the training sample data in the target domain through a first loss function according to the first result corresponding to the training sample data in the target domain and the corresponding domain label.
  • a third loss function is used to calculate the third loss corresponding to the target domain training sample data; specifically, the high-level features of the target domain training sample data are input into the task module.
  • a third result corresponding to the training data of the target domain is obtained; based on the third result corresponding to the training data of the target domain and the corresponding prediction sample label, a third loss corresponding to the training sample data of the target domain is calculated by a third loss function.
  • All steps described in the embodiment corresponding to FIG. 12 may be performed by the training device 110 or the execution device 210 individually, or may be performed by multiple devices or devices, and each device or device performs some of the steps described in the embodiment corresponding to FIG. 12 .
  • all the steps described in the embodiment corresponding to FIG. 12 are performed by the training device 110.
  • the selected target field training sample data is used as labeled training data (including the sample label and the field label), and the training device 110 is input again.
  • the parameters of each unit in the training device 110 at this time are not exactly the same as the parameters when the prediction labels of the training sample data of the target domain are obtained.
  • the parameters of each unit in the training device 110 may be updated at least once.
  • the training method provided in the embodiment of the present application actually trains a task model and a domain discrimination model at the same time.
  • the task model includes a feature extraction module and a task module, a model for a specific task.
  • the domain distinguishing model includes a feature extraction module and a domain distinguishing feature module, which are used to distinguish the domains, that is, the domain (source domain or target domain) to which the data belongs is given for the input data.
  • the label used for training the domain discrimination model is the domain label For example, set the field label of the source field data to 0 and set the field label of the target field data to 1.
  • the domain distinguishing feature module in the domain distinguishing model may be the domain distinguishing feature unit 114 or the domain distinguishing feature unit 113 '.
  • step numbers are not intended to execute the steps in the order of the numbers.
  • the steps have a logical order and can be determined according to the technical solution. Therefore, the numbers do not limit the method flow.
  • the numbers in FIG. 12 are not limitations on the method flow.
  • the training method provided in the embodiment of the present application is implemented based on the enhanced cooperative adversarial network, as shown in FIG. 13, based on the enhanced cooperative adversarial network constructed by CNN.
  • the cooperative adversarial network refers to a network formed by establishing a domain discriminative loss function and a domain invariance loss function based on low-level features and high-level features.
  • the domain discriminative loss function is configured in the domain distinguishing feature unit 114, and the domain invariance loss is established.
  • the function is arranged in the domain invariance feature unit 113.
  • Enhanced collaborative adversarial network is based on the collaborative adversarial network, which adds the process of selecting training data from the target domain data and setting weights for training.
  • the image classifier is taken as an example to describe the training method provided in the embodiment of the present application.
  • source area image data 301 and target area image data 302 are input.
  • the source domain image data 301 is image data labeled with a category label
  • the target domain image data 302 is image data not labeled with a category label.
  • the category label is used to indicate the category of the image data.
  • the trained image classifier is used to predict the image data. category.
  • Image data can be pictures or video streams, or other image data formats.
  • the source domain image data 301 and the target domain image data 302 correspond to respective domain labels, and the domain labels are used to indicate the domain to which the image data belongs. There is a difference between the source domain image data 301 and the target domain image data 302 (for example, the example given in the above application scenario embodiment), and the mathematical distribution is different in the data distribution.
  • the low-level feature extraction 303 corresponds to the low-level feature extraction subunit 1111, and CNN can be used to perform a convolution budget to extract low-level features in the image data.
  • the input data of the low-level feature extraction 303 includes the source domain image data 301, which can be expressed as among them Is the ith one in the source domain image data, For its category label, N s is the number of samples in the source domain image data. Accordingly, the target domain image data 301 can be expressed as No category tags.
  • Low-level feature extraction 303 can be implemented using a series of convolutional layers, normalization layers, and down-sampling layers, represented by F k (x i ; ⁇ k ), where k is the number of low-level feature extraction 303 and ⁇ k is low-level feature extraction 303 Parameters.
  • High-level feature extraction 304 is a further processing of low-level features based on low-level feature extraction 303.
  • high-level feature extraction 304 corresponds to high-level feature extraction subunit 1112.
  • CNN can be used for convolution budget extraction to extract high-level features in image data.
  • low-level feature extraction 303 it can be implemented using a series of convolutional layers, specification layers, and down-sampling layers, which can be represented by F m (x i ; ⁇ m ), where m is the total number of feature extraction layers.
  • the image classification 305 extracts 304 the high-level features input by the layer feature, and outputs the predicted category information, which can be expressed as C: f ⁇ y i or an image classifier C (F (x i ; ⁇ F ), c) Where c is the parameter of the image classifier.
  • Image classification can be extended to a variety of computer vision tasks, including detection, recognition, segmentation, and more.
  • a classification loss function (corresponding to a third loss function) is defined according to an output of the image classification 305 and a category label of the image data (corresponding to the source data category label in FIG. 13) to optimize parameters in the image classification 305.
  • This classification loss function can be defined as Image classification 305 outputs cross-entropy with corresponding class labels.
  • the classification loss function of the source domain image data 301 can be defined as By iteratively optimizing the slave parameters of image classification 305 to minimize the classification loss function, an image classifier is obtained. It should be noted that the image classifier here does not include the feature extraction part. In practice, the image classifier needs to be used in conjunction with feature extraction (low-level feature extraction 303 and high-level feature extraction 304). The training process actually classifies the image. The parameters of 305 (image classifier), low-level feature extraction 303, and high-level feature extraction 304 are updated and optimized.
  • the high-level features of the image used by the image classifier should have domain invariance.
  • domain invariance 306 can make high-level features indistinguishable from domains, thereby having domain invariance.
  • the domain invariance 306 includes a domain classifier set for high-level feature extraction 304, which can be expressed as D (F (x i ; ⁇ F ), w), where w is a parameter of the domain classifier.
  • a domain invariance loss function L D (D (F (x i ; ⁇ F ), w), d i ) (corresponding to the first loss) can also be defined according to the output and domain labels of the domain invariance 306 function).
  • the domain invariance 306 makes the domain invariance loss function not tend to be minimized through the gradient inversion method. Change, but the loss becomes larger.
  • the gradient inversion method can be implemented using any existing technology, and no specific limitation is imposed on the specific method of gradient inversion here.
  • the domain classifier does not include feature extraction.
  • the domain classifier needs to be used in conjunction with feature extraction (low-level feature extraction 303 and high-level feature extraction 304).
  • feature extraction low-level feature extraction 303 and high-level feature extraction 304.
  • the parameters of the domain discriminator, low-level feature extraction 303, and high-level feature extraction 304 in domain invariance 305 are actually updated and optimized.
  • the low-level features of an image include the edges and corners of the image. These features often have a greater relationship with the domain and can be used for domain discrimination. If only the domain-invariant features are emphasized in training, the high-level feature distribution between the source domain image data 301 and the target domain image data 302 is similar, so that the image classification model trained on the source domain image data is in the target domain image The data also has a good effect, so the low-level features also have domain invariance, and a lot of domain distinguishing features are lost.
  • a low-level feature extraction 303 can be performed, and a domain-distortion loss function (corresponding to the second loss function) is defined according to the output of the domain discrimination 307 and the domain label, so that the extracted low-level features have domain discrimination.
  • a domain-distortion loss function (corresponding to the second loss function) is defined according to the output of the domain discrimination 307 and the domain label, so that the extracted low-level features have domain discrimination.
  • L D D (F (x i; ⁇ k), w k), d i) loss function layers, wherein k is added.
  • the domain discriminative loss function is combined with the domain invariant loss function to form a cooperative adversarial network.
  • the overall loss function can be expressed as:
  • ⁇ k is the weight of the k-layer loss function
  • ⁇ m is the weight of the m-layer loss function
  • ⁇ m is negative.
  • the domain discrimination and domain invariance of the features are balanced by weights, and the parameters are optimized during the network training process using the gradient-based method to improve the network performance.
  • the image data of the target domain can be used for the training of the image classification model. Since the target area image data 302 originally does not have a category label, the high-level features obtained by the low-level feature extraction 303 and the high-level feature extraction 304 of the target area image data 302 can be used as the labels of the target area image data 302. That is, the output of the image classification model trained on the target domain using the method described above on the target domain image data 302 is used as its category label, and then the target domain image data with the category label is added as new training data after the iterative training process. Specifically, Refer to FIG. 12 corresponding to 1) -6) in the embodiment.
  • the output of the image classification model for sample data includes category information and confidence.
  • the output confidence is high, it is more likely that the category information is correct. Therefore, you can choose target domain image data with high confidence as the target domain training sample. data. Specifically, a threshold is set first; then, image data of which the confidence level is greater than the threshold is selected from the target field image data 302 as target field training sample data. In addition, it is considered that the accuracy of the image classification model is low during the training process.
  • the setting of this threshold is related to the accuracy of the model, that is, the adaptive threshold is set according to the accuracy of the image classification model currently obtained.
  • the adaptive threshold is set according to the accuracy of the image classification model currently obtained.
  • a weight is set on the selected target domain training sample data.
  • the distribution of the target domain training sample data is relatively close to the source domain image data and the target domain image data, which is more helpful for the training of image classification models. Big weight. If the training data of the target domain can be easily distinguished by the domain classifier, the training sample data of the target domain is less valuable for the training of the image classification model, and its weight in the loss function can be reduced.
  • the sample with the domain discriminator output of 0.5 has the largest weight, and the weights on both sides decrease in order. When the value reaches a certain value, the weight is 0.
  • the weight can be expressed using the following formula:
  • a larger value is used for the weight of the target field training sample data near the target field image data.
  • weights There are many ways to set such weights, such as Then set the weight to Corresponding weight value:
  • a classification loss function can be established for the training sample data in the target domain, which can be expressed as
  • the overall loss function based on the enhanced cooperative adversarial network is composed of three parts, that is, the classification loss function on the image data in the source domain, the cooperative adversarial loss function on the low-level features and high-level features, and the classification on the training data on the target domain
  • the loss function can be expressed as:
  • the overall loss function can be optimized using a back-propagation method based on a stochastic gradient to update the parameters of each part of the enhanced cooperative adversarial network, train an image classification model, and use the image classification model for class prediction of image data in the target domain.
  • the low-level feature extraction 303, high-level feature extraction 304, image classification 305, domain invariance 306, domain discrimination 307, sample data selection 308, and weight setting 309 in FIG. 13 can be regarded as the composition of the enhanced cooperative adversarial network. Modules can also be seen as operating steps in a training method based on enhanced cooperative adversarial networks.
  • the embodiment of the present application provides a chip hardware structure. As shown in FIG. 15, the algorithm / method based on the convolutional neural network described in the embodiment of the present application (the embodiment corresponding to FIG. 12 and the embodiment corresponding to FIG. 13 The algorithms / methods involved may be implemented in whole or in part in the NPU chip shown in FIG. 15.
  • the neural network processor NPU 50 NPU is mounted as a coprocessor on the main CPU (Host CPU), and the Host CPU distributes tasks.
  • the core part of the NPU is an arithmetic circuit 50.
  • the controller 504 controls the arithmetic circuit 503 to extract matrix data in the memory and perform multiplication operations.
  • the arithmetic circuit 503 includes a plurality of processing units (Process Engines, PEs). In some implementations, the arithmetic circuit 503 is a two-dimensional pulsating array. The arithmetic circuit 503 may also be a one-dimensional pulsation array or other electronic circuits capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit 503 is a general-purpose matrix processor.
  • PEs Processing Units
  • the arithmetic circuit 503 is a two-dimensional pulsating array.
  • the arithmetic circuit 503 may also be a one-dimensional pulsation array or other electronic circuits capable of performing mathematical operations such as multiplication and addition.
  • the arithmetic circuit 503 is a general-purpose matrix processor.
  • the operation circuit takes the data corresponding to the matrix B from the weight memory 502 and buffers the data on each PE in the operation circuit.
  • the arithmetic circuit takes matrix A data from the input memory 501 and performs matrix operations on the matrix B. Partial or final results of the obtained matrix are stored in the accumulator 508 accumulator.
  • the unified memory 506 is used to store input data and output data.
  • the weight data is directly accessed to the controller 505 through the memory unit, and the DMAC is transferred to the weight memory 502.
  • the input data is also transferred to the unified memory 506 through the DMAC.
  • BIU is a Bus Interface Unit, that is, a bus interface unit 510, which is used for the interaction of the AXI bus with the DMAC and the instruction fetch buffer 509Instruction and FetchBuffer.
  • the bus interface unit 510 (Bus Interface Unit) is used to fetch the memory 509 to obtain instructions from external memory, and is also used to store the unit access controller 505 to obtain the original data of the input matrix A or weight matrix B from the external memory.
  • the DMAC is mainly used to transfer input data in the external memory DDR to the unified memory 506 or weight data to the weight memory 502 or input data to the input memory 501.
  • the vector calculation unit 507 has a plurality of operation processing units. If necessary, the output of the operation circuit is further processed, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, and so on. It is mainly used for non-convolutional / FC layer network calculation in neural networks, such as Pooling, Batch Normalization, Local Normalization, and so on.
  • the vector calculation unit can 507 store the processed output vector into the unified buffer 506.
  • the vector calculation unit 507 may apply a non-linear function to the output of the arithmetic circuit 503, such as a vector of accumulated values, to generate an activation value.
  • the vector calculation unit 507 generates a normalized value, a merged value, or both.
  • a vector of the processed output can be used as an activation input to the arithmetic circuit 503, for example for use in subsequent layers in a neural network.
  • An instruction fetch memory 509 connected to the controller 504 is used to store instructions used by the controller 504;
  • the unified memory 506, the input memory 501, the weight memory 502, and the fetch memory 509 are all On-Chip memories. External memory is private to the NPU hardware architecture.
  • each layer in the convolutional neural network may be performed by the matrix calculation unit 212 or the vector calculation unit 507.
  • the training device 410 includes a processor 412, a communication interface 413, and a memory 411.
  • the training device 410 may further include a bus 414.
  • the communication interface 413, the processor 412, and the memory 411 may be connected to each other through a bus 414.
  • the bus 414 may be a peripheral component interconnect standard (English: Peripheral Component Interconnect (PCI) bus) or an extended industry standard structure (English: Extended Industry). Standard Architecture (EISA) bus and so on.
  • PCI Peripheral Component Interconnect
  • EISA Standard Architecture
  • the above-mentioned bus 414 can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only a thick line is used in FIG. 16, but it does not mean that there is only one bus or one type of bus.
  • the training device shown in FIG. 16 may be used instead of the training device 110 to execute the method described in the above method embodiment, and the specific implementation may also refer to the corresponding description of the above method embodiment, which is not repeated here.
  • the steps of the method or algorithm described in connection with the disclosure of the embodiments of the present invention may be implemented in a hardware manner, or may be implemented in a manner that a processor executes software instructions.
  • Software instructions can be composed of corresponding software modules.
  • Software modules can be stored in random access memory (English: Random Access Memory, RAM), flash memory, read-only memory (English: Read Only Memory, ROM), erasable and programmable Read-only memory (English: Erasable Programmable ROM, EPROM), electrically erasable programmable read-only memory (English: Electrically EPROM, EEPROM), registers, hard disk, mobile hard disk, read-only optical disk (CD-ROM), or well-known in the art Any other form of storage medium.
  • An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium.
  • the storage medium may also be an integral part of the processor.
  • the processor and the storage medium may reside in an ASIC.
  • the ASIC can reside in a network device.
  • the processor and the storage medium may also exist in the terminal device as discrete components.
  • Office-31 is a standard data set for object recognition. It contains 4110 pictures, of which there are 31 categories of objects. It contains data for four fields Amazon (A), Webcam (W), and Dlsr (D).
  • A Amazon
  • W Webcam
  • D Dlsr
  • ImageCLEF-DA is the CLEF 2014 challenge data set, which contains data from three areas, namely ImageNet ILSVRC2012 (I), Bing (B), and Pascal VOC 2012 (P).
  • the data for each domain contains data for 12 categories, each category has 50 pictures.
  • FIG. 17A and FIG. 17B show the test accuracy based on the method provided by the embodiment of the present application and several other methods, such as the method of ResNet50, DANN, JAN, etc., and mean transfer learning accuracy is also given at the same time.
  • the algorithm based on cooperative adversarial network obtains the best effect except JAN
  • the enhanced cooperative adversarial network obtains the optimal effect
  • the average migration accuracy is higher than the current best method JAN by 2 ⁇ 3 percentage points.
  • the training method based on the enhanced cooperative adversarial network establishes a domain invariance loss function and a domain discriminative loss function based on high-level feature extraction and low-level feature extraction, respectively, while ensuring the domain-invariant features of high-level features, The domain distinguishing features in the low-level features are retained, which can improve the accuracy of image classification prediction applied by the image classifier to the target domain.
  • a person of ordinary skill in the art can understand that all or part of the processes in the method of the foregoing embodiment can be implemented by using a computer program to instruct related hardware.
  • the above program can be stored in a computer-readable storage medium, and the program is being executed. In this case, the processes of the embodiments of the methods described above may be included.
  • the foregoing storage medium includes various media that can store program codes, such as a ROM, a RAM, a magnetic disk, or an optical disc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

本发明涉及人工智能，提出了一种协同对抗网络，在协同对抗网络的低层设置损失函数，用于学习域区分性特征，并且与协同对抗网络的最后一层(即高层)设置的域不变性损失函数形成协同对抗目标函数，实现同时学习域区分性特征与域不变性特征。进一步地，提出了一种增强协同对抗网络，在协同对抗网络的基础上，将目标领域的数据加入协同对抗网络的训练中，以及根据任务模型的精度，设置自适应阈值对目标领域的训练样本进行选择，并根据域区分网络的置信度，对目标领域的训练样本设置权重。通过该协同对抗网络训练的任务模型能够提高应用在目标领域中的预测精度。

Description

深度神经网络的训练方法和装置 技术领域
本发明涉及机器学习领域，特别涉及迁移学习领域中基于对抗网络的训练方法和装置。
背景技术
人工智能(Artificial Intelligence,AI)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用系统。换句话说，人工智能是计算机科学的一个分支，它企图了解智能的实质，并生产出一种新的能以人类智能相似的方式作出反应的智能机器。人工智能也就是研究各种智能机器的设计原理与实现方法，使机器具有感知、推理与决策的功能。人工智能领域的研究包括机器人，自然语言处理，计算机视觉，决策与推理，人机交互，推荐与搜索，AI基础理论等。
深度学习是近年来人工智能领域发展的一个关键推动力，尤其是在计算机视觉的多种任务方面，如目标分类/检测/识别/分割中，取得了令人瞩目的效果；但是，深度学习的成功需要依赖于大量的已标注的数据。然而，标注大量的数据，是一项极其耗时耗力的工作。目前针对相同或相似的任务，可以将依据源领域中公开的数据集或已标注的数据训练好的任务模型直接应用到目标领域的任务预测，目标领域是相对于源领域而言的，目标领域一般没有已标注的数据或者没有足够的已标注的数据，源领域中公开的数据集和已标注的数据可以称作源领域数据，相应的，目标领域中未标注的数据可以称作目标领域数据。由于目标领域数据与源领域数数据的分布不相同，直接使用依据源领域数据训练好的模型的效果不佳。
非监督域适应(unsupervised domain adaption)是一种典型的迁移学习方法，可用于解决上述问题。与直接将依据源领域数据训练好的模型用于目标领域的任务预测不同，非监督域适应方法不仅利用源领域数据进行训练，同时将未标注的目标领域数据融合到训练当中，使训练的模型在目标领域数据上有较好的预测效果。目前，现有技术中性能比较好的非监督域适应方法是基于领域对抗的非监督域适应方法，如图1所示的一种基于领域对抗的非监督域适应训练图像分类器的方法，其特点是在学习图像分类任务的同时使用领域区分器(英文全称：Domain Discriminator)和梯度方向(Gradient Reversal)方法学习域不变性特征。主要步骤是：(1)使用卷积神经网络特征提取器(Convolutional Neural Network Feature Extractor，CNN Feature Extractor)提取的特征除了输入到图像分类器中，还用于建立一个领域区分器，领域区分器可以对输入的特征可以输出领域类别；(2)使用梯度反向方法，在反向传播过程中修改梯度方向，从而使得卷积神经网络特征提取器学习的特征具有域不变性；(3)将以上得到卷积神经网络特征提取器和得到的分类器，用于目标领域的图像分类预测。
发明内容
为了解决基于领域对抗的非监督域适应方法存在的丢失具有域区分性的低层特征的问题。本申请提供了一种基于协同对抗网络的训练方法，能够保留具有域区分性的低层特征，从而提高任务模型的精度。进一步提供了一种增加协同领域对抗的方法，将目标领域中的数据用于训练任务模型，提高训练出的任务模型在目标领域的适配性。
第一方面，本申请提供了一种深度神经网络的训练方法，该训练方法应用于迁移学习领域，具体是将根据源领域数据训练的任务模型应用到目标领域数据的预测，该训练方法包括：提取输入该深度神经网络的源领域数据和目标领域数据中各样本数据所对应的低层特征和高层特征，其中，目标领域数据与源领域数据存在差异，也就是说两者的数据分布不一致；基于源领域数据和目标领域数据中各样本数据的高层特征和对应的领域标签，通过第一损失函数分别计算各样本数据对应的第一损失；基于源领域数据和目标领域数据中各样本数据的低层特征和对应的领域标签，通过第二损失函数分别计算各样本数据对应的第二损失；基于源领域数据中的样本数据的高层特征和对应的样本标签，通过第三损失函数计算源领域数据中的样本数据对应的第三损失；根据上述得到的第一损失、第二损失和第三损失更新目标深度神经网络中各模块的参数。更新是通过损失反向传播对参数进行更新，在反向传播中，第一损失的梯度需要经过梯度反向操作，梯度反向操作的目的实现反向传导梯度使损失变大。通过在高层特征和低层特征分别设置第一损失函数和第二损失函数，可以使得高层特征具有不变性的同时使得低层特征具有域区分性，提高训练得到的模型应用于到目标领域的预测的精度。
第一方面的一种可能的实现方式，该目标深度神经网络包括特征提取模块、任务模块、域不变性特征模块和域区分性特征模块，特征提取模块包括至少一个低层特征网络层和高层特征网络层，至少一个低层特征网络层中的任一个低层特征网络层可用于提取低层特征，高层特征网络层用于提取高层特征，域不变性特征模块用于增强特征提取模块提取的高层特征的领域不变性，域区分性特征模块用于增强特征提取模块提取的低层特征的领域区分性；
其中，上述根据第一损失、第二损失和第三损失更新目标深度神经网络的参数包括：首先根据第一损失、第二损失和第三损失计算总损失；再根据总损失更新特征提取模块的参数、任务模块的参数、域不变性特征模块的参数和域区分性特征模块的参数，需要注意的是，总损失可以是一个样本数据的第一损失、第二损失和第三损失的总和，也可以是多个样本数据的多个第一损失、多个第二损失和多个第三损失的总和。各损失具体在反向传播过程中作于目标神经网络中相应的模块的参数，具体的是第一损失通过反向传播对对域不变性特征模块和特征提取模块的参数进行更新，第二损失通过反向传播对域区分性特征模块和特征提取模块的参数进行更新。第三损失通过反向传播对任务模块和特征提取模块的参数进行更新。损失一般是进一步得到相应的梯度在进行反向传播进行更新相对模块的参数。
第一方面的另一种可能的实现方式，上述基于源领域数据和目标领域数据中各样本数据的高层特征和对应的领域标签，通过第一损失函数分别计算各样本数据对应的第一损失，包括：将源领域数据和目标领域数据中的各样本数据的高层特征输入域不变性特征模块得到各样本数据对应的第一结果；根据源领域数据和目标领域数据中的各样本数据对应的第 一结果和对应的领域标签，通过第一损失函数分别计算各样本数据对应的第一损失。
上述基于源领域数据和目标领域数据中各样本数据的低层特征和对应的领域标签，通过第二损失函数分别计算各样本数据对应的第二损失，包括：将源领域数据和目标领域数据中的各样本数据的低层特征输入域区分性特征模块得到各样本数据对应的第二结果；根据源领域数据和目标领域数据中的各样本数据对应的第二结果和对应的领域标签，通过第二损失函数分别计算各样本数据对应的第二损失。
上述基于源领域数据中的样本数据的高层特征和对应的样本标签，通过第三损失函数计算源领域数据中的样本数据对应的第三损失，包括：将源领域数据中的样本数据的高层特征输入任务模块得到源领域数据中的样本数据对应的第三结果；基于源领域数据中的样本数据对应的第三结果和对应的样本标签，通过第三损失函数计算源领域数据中的样本数据对应的第三损失。
第一方面的另一种可能的实现方式，域不变性特征模块还包括：梯度反向模块；该训练方法还包括：通过该梯度反向模块对第一损失的梯度进行梯度反向。梯度方向可以实现反向传导第一损失的梯度使得第一损失函数的计算的损失变大，使得高层特征具有域不变性特征，
第一方面的另一种可能的实现方式，该训练方法还包括：将目标领域数据中样本数据的高层特征输入任务模块，得到对应的预测样本标签和对应的置信度；根据目标领域数据中样本数据对应的置信度从目标领域数据中选定目标领域训练样本数据，目标领域训练样本数据为目标领域数据中对应的置信度满足预设条件的样本数据。使用目标领域数据用于训练任务模型，能够进一步提高任务模型在目标领域的数据上的分类精度。
第一方面的另一种可能的实现方式，该训练方法还包括：根据目标领域训练样本数据对应的第一结果设置目标领域训练样本数据的权重。当目标领域训练样本数据不易被领域区分器区分时，则目标领域训练样本数据的分布比较接近于源领域图像数据与目标领域图像数据之间，对图像分类模型的训练更有帮助，因此根据第一结果设置权重能将上述描述的不易被领域区分的目标领域训练样本数据在训练中占较大的权重。
第一方面的另一种可能的实现方式，根据目标领域训练样本数据对应的第一结果设置目标领域训练样本数据的权重包括:根据目标领域训练样本数据对应的第一结果与领域标签的相似度，设置目标领域训练样本数据的权重，相似度表示第一结果与领域标签的差值大小。
第一方面的另一种可能的实现方式，上述根据目标领域训练样本数据对应的第一结果与领域标签的相似度，设置目标领域训练样本数据的权重包括：计算目标领域训练样本数据对应的第一结果与源领域的领域标签的第一差值，以及目标领域训练样本数据对应的第一结果与目标领域的领域标签的第二差值；若第一差值的绝对值大于第二差值的绝对值，则设置目标领域训练样本数据的权重为较小的值，例如小于0.5的值；否则，设置目标领域训练样本数据的权重为较大的值，例如大于0.5的值。
第一方面的另一种可能的实现方式，若目标领域训练样本数据对应的第一结果为第一领域标签值至第二领域标签值取值范围中的中间值，则设置目标领域训练样本数据的权重为最大值(例如1)。关于中间值的示例，例如第一领域标签值为0，第二领域标签值为1，中间值是指0.5或者为0.5上下浮动区间中的值。其中第一领域标签值为源领域的领 域标签对应的值，第二领域标签值为目标领域的领域标签对应的值。
第一方面的另一种可能的实现方式，在上述根据目标领域数据中样本数据对应的置信度从目标领域数据中选定目标领域训练样本数据之前，该训练方法还包括：根据任务模型的精度设置自适应阈值，任务模型包括特征提取模块和任务模块，自适应阈值与任务模型的精度正相关；其中，预设条件为置信度大于或等于自适应阈值。
第一方面的另一种可能的实现方式，自适应阈值通过下面逻辑函数计算：
Figure PCTCN2019088846-appb-000001
其中，T c为自适应阈值，A为任务模型的精度，λ c为用于控制逻辑函数的倾斜度的超参数。
第一方面的另一种可能的实现方式，训练方法还包括：通过特征提取模块提取目标领域训练样本数据的低层特征和高层特征；基于目标领域训练样本数据的高层特征和对应的领域标签，通过第一损失函数计算目标领域训练样本数据对应的第一损失；基于目标领域训练样本数据的低层特征和对应的领域标签，通过第二损失函数计算目标领域训练样本数据对应的第二损失；基于目标领域训练样本数据的高层特征和对应的预测样本标签，通过第三损失函数计算目标领域训练样本数据对应的第三损失；根据目标领域训练样本数据对应的第一损失、第二损失和第三损失计算目标领域训练样本数据对应的总损失，其中，目标领域训练样本数据对应的第一损失的梯度经过梯度反向；根据目标领域训练样本数据对应的总损失和目标领域训练样本数据的权重，更新特征提取模块的参数、任务模块的参数、域不变性特征模块的参数和域区分性特征模块的参数。
第一方面的另一种可能的实现方式，上述基于目标领域训练样本数据的高层特征和对应的领域标签，通过第一损失函数计算目标领域训练样本数据对应的第一损失包括：将目标领域训练样本数据的高层特征输入域不变性特征模块得到目标领域训练样本数据对应的第一结果；根据目标领域训练样本数据对应的第一结果和对应的领域标签，通过第一损失函数计算目标领域训练样本数据对应的第一损失；
上述基于目标领域训练样本数据的低层特征和对应的领域标签，通过第二损失函数计算目标领域训练样本数据对应的第二损失包括：将目标领域训练样本数据的低层特征输入域区分性特征模块得到目标领域训练样本数据对应的第二结果；根据目标领域训练样本数据对应的第二结果和对应的领域标签，通过第二损失函数计算目标领域训练样本数据对应的第二损失；
基于目标领域训练样本数据的高层特征和对应的预测样本标签，通过第三损失函数计算目标领域训练样本数据对应的第三损失，包括：将目标领域训练样本数据的高层特征输入任务模块得到目标领域训练样本数据对应的第三结果；基于目标领域训练样本数据对应的第三结果和对应的预测样本标签，通过第三损失函数计算目标领域训练样本数据对应的第三损失。
第二方面，本申请提供了一种训练设备，该训练设备包括存储器及与存储器耦合的处理器；存储器用于存储指令，处理器用于执行指令；其中，处理器执行指令时执行上述第一方面和第一方面的可能的实现方式中描述的方法。
第三方面，本申请提供了一种计算机可读存储介质，该计算机可读存储有计算机程序， 该计算机程序被处理器执行时实现上述第一方面和第一方面的可能的实现方式中描述的方法。
第四方面，本申请提供了一种计算机程序产品，该计算机程序产品包括用于执行上述第一方面和第一方面的可能的实现方式中描述的方法的代码。
第五方面，本申请提供了一种训练装置，该训练装置包括用于执行上述第一方面和第一方面的可能的实现方式中描述的方法的功能单元。
第六方面，本申请提供了一种基于卷积神经网络CNN构建的增强协同对抗网络，该增强协同对抗网络包括：用于提取源领域数据和目标领域数据中各样本数据的低层特征和高层特征的特征提取模块，目标领域数据与源领域数据的数据分布不同；用于接收特征提取模块输出的高层特征且通过第三损失函数分别计算各样本数据对应的第三损失的任务模块，第三损失用于更新特征提取模块和任务模块的参数；用于接收特征提取模块输出的高层特征且通过第一损失函数分别计算各样本数据对应的第一损失的域不变性模块，第一损失用于更新特征提取模块和域不变性模块的参数，使得特征提取模块输出的高层特征具有域不变性；用于接收特征提取模块输出的低层特征且通过第二损失函数分别计算各样本数据对应的第二损失的域区分性模块，第二损失用于更新特征提取模块和域区分性模块的参数，使得特征提取模块输出的低层特征具有域区分性。
第六方面的一种可能的实现方式，该增强协同对抗网络还包括：用于根据目标领域数据中样本数据对应的置信度从目标领域数据中选定目标领域训练样本数据的样本数据选择模块，目标领域数据中样本数据对应的置信度通过将目标领域数据中样本数据的高层特征输入任务模块得到，目标领域训练样本数据为目标领域数据中对应的置信度满足预设条件的样本数据。
第六方面的另一种可能的实现方式，上述样本数据选择模块还用于根据任务模型的精度设置自适应阈值，任务模型包括特征提取模块和任务模块，自适应阈值与任务模型的精度正相关；其中，预设条件为置信度大于或等于自适应阈值。
第六方面的另一种可能的实现方式，该增强协同对抗网络还包括用于根据目标领域训练样本数据对应的第一结果设置目标领域训练样本数据的权重的权重设置模块。
第六方面的另一种可能的实现方式，上述权重设置模块具体用于根据目标领域训练样本数据对应的第一结果与领域标签的相似度，设置目标领域训练样本数据的权重；相似度表示第一结果与领域标签的差值大小。
第六方面的另一种可能的实现方式，上述权重设置模块具体用于计算目标领域训练样本数据对应的第一结果与源领域的领域标签的第一差值，以及目标领域训练样本数据对应的第一结果与目标领域的领域标签的第二差值；若第一差值的绝对值大于第二差值的绝对值，则设置目标领域训练样本数据的权重为较小的值，否则，设置目标领域训练样本数据的权重为较大的值。
第六方面的另一种可能的实现方式，上述权重设置模块具体用于：若目标领域训练样本数据对应的第一结果为第一领域标签值至第二领域标签值取值范围中的中间值，则设置目标领域训练样本数据的权重为最大值，例如1，第一领域标签值为源领域的领域标签对应的值，第二领域标签值为目标领域的领域标签对应的值。中间值的说明可以参见第一方面的相关描述，此处不再赘述。
第七方面，本申请提供了一种基于协同对抗网络的训练数据权重设置方法，该协同对抗网络至少包括特征提取模块、任务模块、域不变性模块，还可以包括域区分性模块，关于各模块可以参考上面第六方面的相关描述，此处不再赘述。该权重设置方法包括：将目标领域数据中样本数据的高层特征输入任务模块得到对应的预测样本标签和对应的置信度；根据目标领域数据中样本数据对应的置信度从目标领域数据中选定目标领域训练样本数据，目标领域训练样本数据为目标领域数据中对应的置信度满足预设条件的样本数据；将将目标领域数据中样本数据的高层特征输入域不变性模块得到目标领域训练样本数据对应的第一结果；根据目标领域训练样本数据对应的第一结果设置目标领域训练样本数据的权重。
第七方面的一种可能的实现方式，上述根据目标领域训练样本数据对应的第一结果设置目标领域训练样本数据的权重具体包括:根据目标领域训练样本数据对应的第一结果与领域标签的相似度，设置目标领域训练样本数据的权重，相似度表示第一结果与领域标签的差值大小。
第七方面的另一种可能的实现方式，上述根据目标领域训练样本数据对应的第一结果与领域标签的相似度，设置目标领域训练样本数据的权重包括：计算目标领域训练样本数据对应的第一结果与源领域的领域标签的第一差值，以及目标领域训练样本数据对应的第一结果与目标领域的领域标签的第二差值；若第一差值的绝对值大于第二差值的绝对值，则设置目标领域训练样本数据的权重为较小的值，例如小于0.5的值；否则，设置目标领域训练样本数据的权重为较大的值，例如大于0.5的值。
第七方面的另一种可能的实现方式，若目标领域训练样本数据对应的第一结果为第一领域标签值至第二领域标签值取值范围中的中间值，则设置目标领域训练样本数据的权重为最大值(例如1)。关于中间值的示例，例如第一领域标签值为0，第二领域标签值为1，中间值是指0.5或者为0.5上下浮动区间中的值。其中第一领域标签值为源领域的领域标签对应的值，第二领域标签值为目标领域的领域标签对应的值。
第七方面的另一种可能的实现方式，在上述根据目标领域数据中样本数据对应的置信度从目标领域数据中选定目标领域训练样本数据之前，该权重设置方法还包括：根据任务模型的精度设置自适应阈值，任务模型包括特征提取模块和任务模块，自适应阈值与任务模型的精度正相关；其中，预设条件为置信度大于或等于自适应阈值。
上述自适应阈值通过下面逻辑函数计算：
Figure PCTCN2019088846-appb-000002
其中，T c为自适应阈值，A为任务模型的精度，λ c为用于控制逻辑函数的倾斜度的超参数。
第八方面，本申请提供了一种设备，该设备包括存储器及与存储器耦合的处理器；存储器用于存储指令，处理器用于执行指令；其中，处理器执行指令时执行上述第七方面和第七方面的可能的实现方式中描述的方法。
第九方面，本申请提供了一种计算机可读存储介质，该计算机可读存储有计算机程序，该计算机程序被处理器执行时实现上述第七方面和第七方面的可能的实现方式中描述的方法。
第十方面，本申请提供了一种计算机程序产品，该计算机程序产品包括用于执行上述第七方面和第七方面的可能的实现方式中描述的方法的代码。
第十一方面，本申请提供了一种权重设置装置，该权重设置装置包括用于执行上述第七方面和第七方面的可能的实现方式中描述的方法的功能单元。
本申请实施例提供的训练方法基于高层特征和低层特征分别建立了域不变性损失函数和域区分性损失函数，在保证高层特征的域不变性特征的同时保留了低层特征中的域区分性特征，能够提高训练得到的任务模型应用到目标领域中进行预测的精度。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例提供的一种基于非监督域适应训练图像分类器的方法示意图；
图2为本发明实施例提供的一种人工智能主体框架示意图；
图3为本发明实施例提供的不同城市的人车图像数据对照示意图；
图4为本发明实施例提供的不同地域的人脸图像数据对照示意图；
图5为本发明实施例提供的一种训练系统架构示意图；
图6为本发明实施例提供的一种特征提取单元的示意图；
图7为本发明实施例提供的一种特征提取CNN的示意图；
图8为本发明实施例提供的一种域不变性特征单元的示意图；
图9为本发明实施例提供的一种训练装置的结构示意图
图10为本发明实施例提供的另一种训练装置的结构示意图；
图11为本发明实施例提供的一种云-端系统架构示意图；
图12为本发明实施例提供的一种训练方法的流程图；
图13为本发明实施例提供的一种基于协同对抗网络的训练方法示意图；
图14为本发明实施例提供的权重设置曲线示意图；
图15为本发明实施例提供的一种芯片硬件结构示意图；
图16为本发明实施例提供的一种训练设备结构示意图；
图17A为本发明实施例提供的在Office-31上的测试结果；
图17B为本发明实施例提供的在ImageCLEF-DA上的测试结果。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
图2示出一种人工智能主体框架示意图，该主体框架描述了人工智能系统总体工作流程，适用于通用的人工智能领域需求。
下面从“智能信息链”(水平轴)和“IT价值链”(垂直轴)两个维度对上述人工智能主题框架进行阐述。
“智能信息链”反映从数据的获取到处理的一列过程。举例来说，可以是智能信息感知、智能信息表示与形成、智能推理、智能决策、智能执行与输出的一般过程。在这个过程中，数据经历了“数据—信息—知识—智慧”的凝练过程。
“IT价值链”从人智能的低层基础设施、信息(提供和处理技术实现)到系统的产业生态过程，反映人工智能为信息技术产业带来的价值。
(1)基础设施：
基础设施为人工智能系统提供计算能力支持，实现与外部世界的沟通，并通过基础平台实现支撑。通过传感器与外部沟通；计算能力由智能芯片(CPU、NPU、GPU、ASIC、FPGA等硬件加速芯片)提供；基础平台包括分布式计算框架及网络等相关的平台保障和支持，可以包括云存储和计算、互联互通网络等。举例来说，传感器和外部沟通获取数据，这些数据提供给基础平台提供的分布式计算系统中的智能芯片进行计算。
(2)数据
基础设施的上一层的数据用于表示人工智能领域的数据来源。数据涉及到图形、图像、语音、文本，还涉及到传统设备的物联网数据，包括已有系统的业务数据以及力、位移、液位、温度、湿度等感知数据。
(3)数据处理
数据处理通常包括数据训练，机器学习，深度学习，搜索，推理，决策等方式。
其中，机器学习和深度学习可以对数据进行符号化和形式化的智能信息建模、抽取、预处理、训练等。
推理是指在计算机或智能系统中，模拟人类的智能推理方式，依据推理控制策略，利用形式化的信息进行机器思维和求解问题的过程，典型的功能是搜索与匹配。
决策是指智能信息经过推理后进行决策的过程，通常提供分类、排序、预测等功能。
(4)通用能力
对数据经过上面提到的数据处理后，进一步基于数据处理的结果可以形成一些通用的能力，比如可以是算法或者一个通用系统，例如，翻译，文本的分析，计算机视觉的处理，语音识别，图像的识别等等。
(5)智能产品及行业应用
智能产品及行业应用指人工智能系统在各领域的产品和应用，是对人工智能整体解决方案的封装，将智能信息决策产品化、实现落地应用，其应用领域主要包括：智能制造、智能交通、智能家居、智能医疗、智能安防、自动驾驶，平安城市，智能终端等。
本申请中涉及的重要概念的相关说明
非监督域适应，是迁移学习的一种典型方法，依据源领域与目标领域的数据进行任务模型的训练，通过训练好的任务模型来实现对目标领域中物体的识别/分类/分割/检测等，其中源领域的数据有标签，而目标领域的数据无标签，并且两种领域数据的分布不相同。需要注意的，在本申请中“源领域的数据”与“源领域数据”，“目标领域的数据”与“目标领域数据”通常上具有相同的含义。
域不变性特征：是指不同领域数据的通用特征，从不同领域数据中提取的特征具有一 致的分布。
域区分性特征：指对于特定领域数据中的特征，对于不同领域数据中提取的特征具有不相同的分布。
本申请描述了一种神经网络的训练方法，该训练方法应用于迁移学习领域的任务/预测模型(以下简称为任务模型)的训练。具体地，可应用于训练基于深度神经网络构建的各种任务模型，包括但不限于分类模型、识别模型、分割模型、检测模型。通过本申请描述的训练方法得到的任务模型可广泛应用到AI拍照、自动驾驶、平安城市等多种具体应用场景，以实现应用场景的智能化。
以自动驾驶应用场景中的人车检测为例，人车检测是自动驾驶感知系统里面的一个基本单元。人车检测的准确程度关系到自动驾驶车辆的安全，能否准确地检测出车辆周围的行人和行车，关键是用于人车检测的检测模型是否具有高精度，然而高精度的检测模型依赖于大量的已标注的人车图像/视频数据。标注数据又是一项庞大的工程，为了达到自动驾驶的精度要求，几乎需要针对不同的城市标注不同数据，这是难以实现的。为了提高训练效率，人车检测模型的迁移是最常用的方法，即直接将依据区域A的已标注的人车图像/视频数据训练的检测模型，应用到没有或没有足够的已标注的人车图像/视频数据的区域B场景中的人车检测，这里的区域A为源领域，区域B为目标领域，区域A的数据为有标签的源领域数据，区域B的数据为无标签的目标领域数据。然而，以城市为例，不同的城市的人种、生活习惯、建筑风格、气候环境、交通设施等以及数据采集设备可能存在很大的差异，即在数据的分布不同，很难保证自动驾驶的精度要求的。如图3所示，左面的四张图像欧洲某一城市的采集设备采集到图像数据，右面的四张图像是亚洲某城市采集设备采集到的图像数据，可以看出，行人皮肤、穿着、姿态存在明显的差异，城市建筑和行车外观也存在很明显的差异。如果将依据图3中一个城市的图像/视频数据训练的检测模型应用到图3中的另一个城市场景，那么检测模型的精度必然会大幅降低。本申请描述的训练方法利用已标注的数据和未标注的数据共同训练任务模型，即利用区域A的已标注的人车图像/视频数据和区域B的为标注的人车图像/视频数据共同训练用于人车检测的检测模型，能够大幅提高依据区域A的人车图像/视频数据训练的检测模型应用到区域B场景中人车检测的精度。
再以人脸识别应用场景为例，人脸识别往往涉及到不同国家、地域的人的识别，不同国家、地域的人的人脸数据会有较大的分布差异。如图4所示，假如欧洲白种人的人脸数据有训练标签作为源领域数据，即已标注的人脸数据；非洲黑种人的人脸数据无训练标签作为目标领域数据，即未标注的人脸数据。由于白种人和黑种人的肤色、脸部轮廓等存在很大的差异，致使人脸数据分布不同；不过，即使黑种人的人脸数据是未标注数据，通过本申请描述的训练方法得到的人脸识别模型也能够提高黑种人的人脸识别准确率。
本发明实施例提供了一种深度神经网络训练系统架构100。如图5所示，系统架构100至少包括训练装置110、数据库120，还包括数据采集设备130、客户设备140和数据存储系统150。
数据采集设备130用于采集数据并将采集到的数据(例如：图片/视频/音频等)存入数据库120作为训练数据。数据库120用于维护和存储训练数据，数据库120存储的训练数据包括源领域数据和目标领域数据，源领域数据可以理解为已标注的数据，目标领域数 据可以理解为未标注的数据，源领域与目标领域是迁移学习领域的相对概念，具体的，可参见图3和图4对应描述理解源领域、目标领域、源领域数据和目标领域数据，上述概念是本技术领域人员能够理解的。训练装置110与数据库120交互，从数据库120中获取需要的训练数据，用来训练任务模型，任务模型包括特征提取模块和任务模块，特征提取模块可以是特征提取单元111，也可以是利用训练后的特征提取单元111的参数构建的深度神经网络；同样地，任务模块可以是任务单元112，也可以是利用训练后的任务单元112的参数构建的模型，例如函数模型、神经网络模型等。训练装置110通过训练得到的任务模型可以应用在客户设备140中，也可以响应客户设备140请求输出预测结果。例如，客户设备140是自动驾驶车辆，训练装置110根据数据库120中的训练数据训练好人车检测模型，当自动驾驶车辆需要执行人车检测时，可以由训练装置110得到的人车检测模型完成人车检车并反馈给自动驾驶车辆，训练好的人车检测模型可以布置在自动驾驶车辆上，也可以是布置在云端，具体形式不做限制。客户设备140在需要的情况下，也可以作为数据库120的数据采集设备，以扩充数据库。
训练装置110包括特征提取单元111、任务单元112、域不变性特征单元113、域区分性特征单元114和I/O接口115，/O接口115用于训练设备110与外界设备进行交互。
特征提取单元111用于提取输入数据的低层特征和高层特征，如图6所示，特征提取单曲单元111包括低层特征提取子单元1111和高层特征提取子单元1112，低层特征提取子单元1111用于提输入数据的低层特征，高层特征提取子单元1112用于提取输入数据的高层特征。具体的，数据输入低层特征提取子单元1111后得到表示低层特征的数据，表示低层特征的数据再输入高层特征提取子单元1112后得到表示高层特征的数据，也就是说高层特征是基于低层特征进一步处理得到的特征。
特征提取单元111可以由软件、硬件(例如电路)或软件与硬件(例如处理器调用代码)结合实现。常用的是通过神经网络实现特征提取单元111的功能，可选的，特征提取单元111的功能由卷积神经网络(Convosutionas Neuras Network，CNN)实现，如图7所示，特征提取CNN包括多个卷积层，通过卷积计算可实现输入数据的特征提取，多个卷积层的最后一层卷积层可以称为高层卷积层，作为高层特征提取子单元1112用于提取高层特征；其他卷积层可称为低层卷积层，作为低层特征提取子单元1111用于提取低层特征。每一个低层卷积层可以均可以输出一个低层特征，即一个数据输入作为特征提取单元111的CNN后，可以输出一个高层特征和至少一个低层特征，低层特征的数量可根据实际训练需求设置，制定具体的输出用于作为低层特征提取子单元1111输出低层特征的低层卷积层。
卷积神经网络(Convosutionas Neuras Network，CNN)是一种带有卷积结构的深度神经网络。卷积神经网络包含了由卷积层和子采样层构成的特征抽取器。该特征抽取器可以看作是滤波器，卷积过程可以看作是使用一个可训练的滤波器与一个输入的图像或者卷积特征平面(feature map)做卷积。卷积层是指卷积神经网络中对输入信号进行卷积处理的神经元层。在卷积神经网络的卷积层中，一个神经元可以只与部分邻层神经元连接。一个卷积层中，通常包含若干个特征平面，每个特征平面可以由一些矩形排列的神经单元组成。同一特征平面的神经单元共享权重，这里共享的权重就是卷积核。共享权重可以理解为提取图像信息的方式与位置无关。这其中隐含的原理是：图像的某一部分的统计信息与其他部分是一样的。即意味着在某一部分学习的图像信息也能用在另一部分上。所以对于 图像上的所有位置，我们都能使用同样的学习得到的图像信息。在同一卷积层中，可以使用多个卷积核来提取不同的图像信息，一般地，卷积核数量越多，卷积操作反映的图像信息越丰富。
卷积核可以以随机大小的矩阵的形式初始化，在卷积神经网络的训练过程中卷积核可以通过学习得到合理的权重。另外，共享权重带来的直接好处是减少卷积神经网络各层之间的连接，同时又降低了过拟合的风险。
卷积神经网络可以采用误差反向传播(back propagation，BP)算法在训练过程中修正初始的超分辨率模型中参数的大小，使得超分辨率模型的重建误差损失越来越小。具体地，前向传递输入信号直至输出会产生误差损失，通过反向传播误差损失信息来更新初始的超分辨率模型中参数，从而使误差损失收敛。反向传播算法是以误差损失为主导的反向传播运动，旨在得到最优的超分辨率模型的参数，例如权重矩阵。
任务单元112的输入是高层特征提取子单元1112输出的高层特征，具体是已标注的源领域数据经过特征提取单元111输出的高层特征，输出是标签。训练后的任务单元112和特征提取单元111可以作为任务模型，任务模型可以用于目标领域的预测任务。
域不变性特征单元113的输入是高层特征提取子单元1112输出的高层特征，输出是对应数据所属的领域(源领域或目标领域)标签。如图8所示，域不变性特征单元113包括域区分特征子单元1131和梯度反向子单元1132，梯度反向子单元1132可以对反向传播的梯度进行梯度反向，使得域区分特征子单元1131输出的领域标签与真实领域标签的误差(即损失)变大。域不变性特征单元113能够实现特征提取单元111输出的高层特征具有领域不变性，也就是降低通过特征提取单元111输出的高层特征较难或无法对领域进行区分。
域区分性特征单元114的输入是低层特征提取子单元1111输出的低层特征，输出是对应数据所属的领域标签。域区分性特征单元114能够使得特征提取单元111输出的低层特征容易对领域进行区分，从而具有域区分性。
需要注意的，域区分性特征单元114与域区分特征子单元1131都可以针对输入特征输出所属的领域，域不变性特征单元113和域区分性特征单元114的主要区别在于域不变性特征单元113还包括梯度反向子单元1132。域区分性特征单元114和特征提取单元111可以构成一个领域区分模型，同样地，忽略梯度反向子单元1132，域不变性特征单元113中的域区分特征子单元1131和特征提取单元111也可以构成一个领域区分模型。
可选的，训练装置110为图9所示的结构，训练装置110包括特征提取单元111、任务单元112、域区分性特征单元113'、梯度反向单元114'和I/O接口115。域区分性特征单元113'和梯度反向单元114'相当于图5中训练装置110的域不变性特征单元113和域区分性特征单元114。
任务单元112、域不变性特征单元113和域区分性特征单元114以及域区分性特征单元113'、梯度反向单元114'可以由软件、硬件(例如电路)或软件与硬件(例如处理器调用代码)结合实现，可以由向量矩阵、函数、神经网络等具体实现，不做限定。任务单元112、域不变性特征单元113和域区分性特征单元114均包括损失函数用于计算输出值与真实值的损失，损失用于更新各单元中的参数，具体更新细节是本技术领域的技术人员所能理解的，不做赘述。
训练装置110包括域不变性特征单元113和域区分性特征单元114，通过源领域数据和 目标领域数据的训练，能够得到的特征提取单元111输出的低层特征具有域区分性，而输出的高层特征具有域不变性，高层特征是基于低层特征进一步得到的，使得高层特征仍能很好的保留具有域区分性的特征，进一步地用于任务模型可以提高预测精度。
如图10所示，训练装置110还包括样本数据选择单元116，样本数据选择单元116用于从目标领域数据中选择满足条件的数据作为训练样本数据用于训练装置110进行的训练。样本数据选择单元116具体包括选择子单元1161和权重设置子单元1162。选择子单元1161用于根据任务模型的精度从目标领域数据中选择出满足条件的数据并添加相应的标签作为训练样本数据。权重设置子单元1162用于给选定的作为训练样本数据的目标领域数据设置权重，通过权重设置以明确作为训练样本数据的目标领域数据对任务模型训练的影响程度。具体如何选择和设置权重，将在下面进行详细描述，此处不再赘述。需要说明的，图10中的其他单元包括图5中的特征提取单元111、任务单元112、域不变性特征单元113、域区分性特征单元114和I/O接口115，或者，特征提取单元111、任务单元112、域区分性特征单元113'、梯度反向单元114'和I/O接口115。
本发明实施例提供了一种云-端系统架构200，如图11所示，执行设备210由一个或多个服务器实现，可选的，与其它计算设备配合，例如：数据存储、路由器、负载均衡器等设备；执行设备210可以布置在一个物理站点上，或者分布在多个物理站点上。可选的，执行设备210可以使用数据存储系统220中的数据，或者调用数据存储系统220中的程序代码实现训练装置110的所有功能；具体地，执行设备210可以根据数据库120中的训练数据训练任务模型，以及根据本地设备231(232)的请求完成目标领域的任务预测。可选的，执行设备210不具备训练装置110的训练功能，但是可以根据训练装置110训练好的任务模型完成预测；具体的，执行设备210配置有训练装置110训练好任务模型后，在接收到本地设备231(232)的请求后完成预测并反馈结果给本地设备231(232)。
用户可以操作各自的用户设备(例如本地设备231和本地设备232)与执行设备210进行交互。每个本地设备可以表示任何计算设备，例如个人计算机、计算机工作站、智能手机、平板电脑、智能摄像头、智能汽车或其他类型蜂窝电话、媒体消费设备、可穿戴设备、机顶盒、游戏机等。
每个用户的本地设备可以通过任何通信机制/通信标准的通信网络与执行设备210进行交互，通信网络可以是广域网、局域网、点对点连接等方式，或它们的任意组合。
在另一种实现中，执行设备210的一个方面或多个方面可以由每个本地设备实现，例如，本地设备301可以为执行设备210提供本地数据或反馈计算结果。
需要注意的，执行设备210的所有功能也可以由本地设备实现。例如，本地设备231实现执行设备210的的功能(例如：训练或预测)并为自己的用户提供服务，或者为本地设备232的用户提供服务。
本申请实施例提供了一种目标深度神经网络的训练方法，该目标深度神经网络是一个系统架构的统称，具体地，包括特征提取模块(对应特征提取单元111)、任务模块(对应任务单元112)、域不变性特征模块(对应域不变性特征单元113)和域区分性特征模块(对应域区分性特征单元114或者域区分性特征单元113')，特征提取模块包括至少一个低层特征网络层(对应低层特征提取子单元1111)和高层特征网络层(对应高层特征提取子单元1112)，至少一个低层特征网络层中的任一个低层特征网络层可用于提取低层特 征，高层特征网络层用于提取高层特征，域不变性特征模块用于增强特征提取模块提取的高层特征的领域不变性，域区分性特征模块用于增强特征提取模块提取的低层特征的领域区分性。如图12所示，该训练方法的具体步骤为：
S101，提取源领域数据和目标领域数据中各样本数据的低层特征和高层特征，目标领域数据与源领域数据在数据分布上不同；
具体地，利用低层特征网络层提取源领域数据和目标领域数据中各样本数据对应的低层特征，利用高层特征网络层提取提取源领域数据和目标领域数据中各样本数据对应的高层特征。
S102，基于源领域数据和目标领域数据中各样本数据的高层特征和对应的领域标签，通过第一损失函数分别计算各样本数据对应的第一损失；具体地，将源领域数据和目标领域数据中的各样本数据的高层特征输入域不变性特征模块得到各样本数据对应的第一结果；根据源领域数据和目标领域数据中的各样本数据对应的第一结果和对应的领域标签，通过第一损失函数分别计算各样本数据对应的第一损失。
进一步地，上述域不变性特征模块还包括：梯度反向模块(对应梯度反向子单元)；该训练方法还包括：通过梯度反向模块对第一损失的梯度进行梯度反向处理，梯度反向的可以使用任一的现有技术，例如Gradient Reversal Layer(GRL)。
S103，基于源领域数据和目标领域数据中各样本数据的低层特征和对应的领域标签，通过第二损失函数分别计算各样本数据对应的第二损失；
具体地，将源领域数据和目标领域数据中的各样本数据的低层特征输入域区分性特征模块得到各样本数据对应的第二结果；根据源领域数据和目标领域数据中的各样本数据对应的第二结果和对应的领域标签，通过第二损失函数分别计算各样本数据对应的第二损失。
S104，基于源领域数据中的样本数据的高层特征和对应的样本标签，通过第三损失函数计算源领域数据中的样本数据对应的第三损失；
具体地，将源领域数据中的样本数据的高层特征输入任务模块得到源领域数据中的样本数据对应的第三结果；基于源领域数据中的样本数据对应的第三结果和对应的样本标签，通过第三损失函数计算源领域数据中的样本数据对应的第三损失。
S105，根据第一损失、第二损失和第三损失更新目标深度神经网络的参数，其中第一损失的梯度经过梯度反向，梯度反向可实现反向传导梯度使损失变大；
具体地，根据第一损失、第二损失和第三损失计算总损失；
根据总损失更新特征提取模块的参数、任务模块的参数、域不变性特征模块的参数和域区分性特征模块的参数。
训练后的特征提取模块和任务模块作为任务模型，用于目标领域的预测任务，当然也可以用源领域的预测任务。
进一步地，该训练方法还包括以下步骤：
S106，将目标领域数据中样本数据的高层特征输入任务模块，得到对应的预测样本标签和对应的置信度。
S107，根据目标领域数据中样本数据对应的置信度从目标领域数据中选定目标领域训练样本数据，目标领域训练样本数据是指目标领域数据中对应的置信度满足预设条件的样本数据；
具体的，根据任务模型的精度设置自适应阈值，任务模型包括特征提取模块和任务模块，自适应阈值与任务模型的精度正相关；其中，预设条件是指置信度大于或等于自适应阈值。
可选的，自适应阈值通过下面逻辑函数计算：
Figure PCTCN2019088846-appb-000003
其中，T c为自适应阈值，A为任务模型的精度，λ c为用于控制逻辑函数的倾斜度的超参数。
S108，根据目标领域训练样本数据对应的第一结果设置目标领域训练样本数据的权重。
具体地，根据域区分特征子单元1131输出的预测值(对应第一结果)，判断其与源领域数据或者目标领域数据分布的相似度，并根据相似度设置目标域样本的权重。相似度可以用预测值与领域标签的差值表示。具体地，预先给源领域标签和目标领域标签各设定一个值，例如，设定源领域的领域标签(可简称源领域标签)为a，设定目标领域的领域标签(可简称目标领域标签)为b，则预测值x的取值范围在a和b之间，可以根据|x-a|与|x-b|的大小来判断相似程度，差值的绝对值越小说明相似程度越大(即更接近)。权重设置可以有两种方案：(1)当预测值更接近源领域领域签的值时，设置较小权重；若预测值在源领域标签的值与目标领域标签的值中间，设置较大权重。(2)当预测值更接近源领域标签的值时，设置较小权重；若输出值与目标领域标签的值更接近时，设置较大权重。上述较小权重和较大权重是相对而言的，可以根据实际设定确定具体数值。权重大小与相似度的关系，可以简单概括为：预测值更倾向于源领域标签值，则相应权重倾向于较小值。也就是，根据预测值判定对应的目标领域训练样本数据是源领域的数据的可能性更大，则设置该目标领域训练样本数据权重较小值，反之可以设置较大值。关于取值设置还可以可参见图14对应实施例的相关描述。
根据步骤S106-S108选定的目标领域训练样本数据除了具有领域标签，还包含预测样本标签和权重，选定的目标领域训练样本数据可用于训练，即相当于源领域数据，重新经过步骤S101-S105，该训练方法还包括针对目标领域训练样本数据的步骤，如下：
1)通过特征提取模块提取目标领域训练样本数据的低层特征和高层特征。
2)基于目标领域训练样本数据的高层特征和对应的领域标签，通过第一损失函数计算目标领域训练样本数据对应的第一损失；具体地，将目标领域训练样本数据的高层特征输入域不变性特征模块得到目标领域训练样本数据对应的第一结果；根据目标领域训练样本数据对应的第一结果和对应的领域标签，通过第一损失函数计算目标领域训练样本数据对应的第一损失。
3)基于目标领域训练样本数据的低层特征和对应的领域标签，通过第二损失函数计算目标领域训练样本数据对应的第二损失；具体地，将目标领域训练样本数据的低层特征输入域区分性特征模块得到目标领域训练样本数据对应的第二结果；根据目标领域训练样本数据对应的第二结果和对应的领域标签，通过第二损失函数计算目标领域训练样本数据对应的第二损失
4)基于目标领域训练样本数据的高层特征和对应的预测样本标签，通过第三损失函数计算目标领域训练样本数据对应的第三损失；具体地，将目标领域训练样本数据的高层 特征输入任务模块得到目标领域训练样本数据对应的第三结果；基于目标领域训练样本数据对应的第三结果和对应的预测样本标签，通过第三损失函数计算目标领域训练样本数据对应的第三损失。
5)根据目标领域训练样本数据对应的第一损失、第二损失和第三损失计算目标领域训练样本数据对应的总损失，其中，目标领域训练样本数据对应的第一损失的梯度经过梯度反向；
6)根据目标领域训练样本数据对应的总损失和目标领域训练样本数据的权重，更新特征提取模块的参数、任务模块的参数、域不变性特征模块的参数和域区分性特征模块的参数。
图12对应的实施例中描述的所有步骤可以由训练装置110或执行设备210单独执行，也可以由多个装置或设备执行，每个装置或设备执行图12对应的实施例中描述的部分步骤。例如图12对应的实施例中描述的所有步骤有训练装置110执行，可理解地，选定的目标领域训练样本数据作为已标注的训练数据(包含样本标签和领域标签)，再次输入训练装置110时的训练装置110中各单元的参数已经与得到目标领域训练样本数据的预测标签时的参数是不完全相同的，此时的训练装置110中各单元的参数可能经过至少一次的更新。
本申请实施例提供的训练方法实际上同时训练了任务模型和领域区分模型。任务模型包括特征提取模块和任务模块，针对特定任务的模型。领域区分模型包括特征提取模块和域区分性特征模块，用于区分所属领域，即针对输入的数据给出该数据所属的领域(源领域或目标领域)，领域区分模型训练使用的标签是领域标签，例如设置源领域数据的领域标签为0，设置目标领域数据的领域标签为1。需要注意的是，领域区分模型中的域区分性特征模块可以是域区分性特征单元114或者域区分性特征单元113'。
需要注意的，上述步骤编号并不是指定按照编号顺序执行各步骤，编号为了方便阅读，各步骤之间具有逻辑顺序，可根据技术方案具体确定，因此，编号并不是对方法流程的限定。同样地，图12中的编号也不是对方法流程的限定。
本申请实施例提供的训练方法是基于增强协同对抗网络实现的，如图13所示的基于CNN构建的增强协同对抗网络。协同对抗网络是指基于低层特征和高层特征分别建立域区分性损失函数和域不变性损失函数形成的网络，可选的，域区分性损失函数配置在域区分性特征单元114，域不变性损失函数配置在域不变性特征单元113。增强协同对抗网络是在协同对抗网络的基础上增加了从目标领域数据中选择训练数据并设置权重用于训练的过程。下面以图像分类器为例描述本申请实施例提供的训练方法。
如图13所示，输入源领域图像数据301和目标领域图像数据302。源领域图像数据301是标注有类别标签的图像数据，目标领域图像数据302是未标注有类别标签的图像数据，类别标签用于指示图像数据的类别，训练后的图像分类器用于预测图像数据的类别。图像数据可以是图片或视频流，也可以是其他图像数据形式。源领域图像数据301和目标领域图像数据302分别对应各自的领域标签，领域标签用于指示图像数据所属的领域。源领域图像数据301与目标领域图像数据302存在差异(例如上面应用场景实施例给出的示例)，体现在数学表达上则是数据分布不同。
低层特征提取303部分
源领域图像数据301和目标领域图像数据302均经过低层特征提取303得到各数据对应的低层特征。低层特征提取303对应低层特征提取子单元1111，可利用CNN进行卷积预算提取图像数据中的低层特征。
具体地，低层特征提取303的输入数据包括源领域图像数据301，可以表示为
Figure PCTCN2019088846-appb-000004
其中
Figure PCTCN2019088846-appb-000005
为源领域图像数据中的第i个，
Figure PCTCN2019088846-appb-000006
为其类别标签，N s为源领域图像数据中样本的数量。相应地，目标领域图像数据301可以表示为
Figure PCTCN2019088846-appb-000007
没有类别标签。低层特征提取303可以使用一系列卷积层、规范层、下采样层实现，用F k(x i；θ k)表示，其中k为低层特征提取303的层数，θ k为低层特征提取303的参数。
高层特征提取304部分
高层特征提取304是在低层特征提取303的基础上对低层特征的进一步的处理，可选的，高层特征提取304对应高层特征提取子单元1112，可以利用CNN进行卷积预算提取图像数据中高层特征，与低层特征提取303一样具体地可以使用一系列卷积层、规范层、下采样层实现，可以用F m(x i；θ m)表示，其中m即为特征提取层的总层数。
图像分类305针对层特征提取304输入的高层特征，输出预测的类别信息，可以表示为C:f→y i，也可以表示为一个图像分类器C(F(x i；Θ F),c)，其中c为图像分类器的参数。图像分类可以扩展到多种计算机视觉任务，包括检测、识别、分割等。另外，根据图像分类305的输出与图像数据的类别标签(对应图13中的源数据类别标签)定义分类损失函数(对应第三损失函数)，以对图像分类305中的参数进行优化。这个分类损失函数可以定义为
Figure PCTCN2019088846-appb-000008
图像分类305输出与对应类别标签的交叉熵。由于源领域图像数据301已有类别标签，可以定义源领域图像数据301的分类损失函数为
Figure PCTCN2019088846-appb-000009
通过迭代优化图像分类305的从参数使得该分类损失函数最小化，得到图像分类器。需要注意是：这里的图像分类器不包含特征提取部分的，在实际中，该图像分类器需要配合特征提取(低层特征提取303和高层特征提取304)使用，训练的过程中实际是对图像分类305(图像分类器)、低层特征提取303和高层特征提取304三者的参数进行更新优化。
域不变性306部分
为使得在源领域图像数据301上训练的图像分类器/模型能够在目标领域图像数据302上同样有较好的分类精度，图像分类器所利用的图像的高层特征应当具有域不变性。为了实现这样的目的，域不变性306能够使得高层特征无法对领域进行区分，从而具有域不变性。具体地，域不变性306包括针对高层特征提取304设置的领域区分器，可以表示为D(F(x i；Θ F),w)，其中w为领域区分器的参数。类似于图像分类器，也可以根据域不变性306的输出与领域标签定义一个域不变性损失函数L D(D(F(x i；Θ F),w),d i)(对应第一损失函数)。与分类损失函数不同的是，为了使源领域图像数据301与目标领域图像数据302之间的高层特征不具有区分性，域不变性306通过梯度反向方法使得域不变性损失函数不是趋于最小化，而是的损失变大。梯度反向方法可以使用任一现有技术实现，此处不对梯度反向的具体方法做任何限制。与图像分类器一样，需要注意的是：这里的领域区分器不包含特征提取部分的，在实际中，该领域区分器需要配合特征提取(低层特征提取303和高层特征提取304)使用，训练的过程中实际是对域不变性305中的领域区分器、低层特征提取303和高层特征提取304三者的参数进行更新优化。
值得注意的，上面的域不变性损失函数与分类损失函数需要同时优化，在训练的过程中组成一个对抗网络，并使用多任务优化方法来解。
域区分性307部分
一般而言，图像的低层特征包括图像的边缘、角点等，这些特征往往是跟领域有较大关系，可以用于领域区分。若在训练中只强调域不变性特征，使得在源领域图像数据301与目标领域图像数据302之间的高层特征分布类似，从而在源领域图像数据上训练得到的图像分类模型在目标领域的图像数据也有较好的效果，则同样使得低层特征也具有了域不变性，丢失了大量域区分性特征。为此可以针对低层特征提取303，根据域区分性307的输出与领域标签定义一个领域区分性损失函数(对应第二损失函数)，使得提取到的低层特征具有域区分性。具体而言，域区分性损失函数可以表示为L D(D(F(x i；θ k),w k),d i)，其中k为所加损失函数的层数。
该域区分性损失函数与域不变性损失函数组在一起，则构成协同对抗网络，总体损失函数可以表示为：
Figure PCTCN2019088846-appb-000010
Figure PCTCN2019088846-appb-000011
其中
Figure PCTCN2019088846-appb-000012
为对于某一层的领域区分目标，λ k为对k层损失函数的权重，为λ m为对m层损失函数的权重，且λ m取负值。在目标函数中，通过权重对特征的域区分性与域不变性进行平衡，并且使用基于梯度的方法在网络训练过程中对参数进行优化，从而提高网络的性能。
样本数据选择308部分
为进一步提高训练的图像分类模型在目标领域的图像数据上的分类精度，可以使用目标领域的图像数据用于图像分类模型的训练。由于目标领域图像数据302原本没有类别标签，可以将目标领域图像数据302通过低层特征提取303、高层特征提取304得到的高层特征，输入图像分类305的输出作为目标领域图像数据302的标签。也就是使用上面描述的方法训练后的图像分类模型在目标领域图像数据302上的输出作为其类别标签，再将拥有类别标签的目标领域图像数据作为新的训练数据加入之后的迭代训练过程，具体的参见图12对应实施例中1)-6)。但是并不是所有的通过图像分类模型获得类别标签的目标领域图像数据都可以作为目标领域训练样本数据。图像分类模型对于样本数据的输出包括类别信息和置信度，当输出的置信度高时，输出类别信息正确的可能性更大，因此，可以选择置信度高的目标领域图像数据作为目标领域训练样本数据。具体地，首先设置一个阈值；再从目标领域图像数据302中选择根据置信度大于该阈值的图像数据作为目标领域训练样本数据。另外，考虑到在训练的过程中，图像分类模型的精度较低。随着训练次数的增加，分类精度会上升，故该阈值的设置与模型的精度有关，即根据当前得到图像分类模型的精度设置自适应的阈值。具体阈值设置可以参见图12对应实施例的相关描述，在此不再赘述。
权重设置309部分
根据域不变性306中域领域区分器的输出，对已选择的目标领域训练样本数据设置权重。当目标领域训练样本数据不易被领域区分器区分时，则目标领域训练样本数据的分布比较接近于源领域图像数据与目标领域图像数据之间，对图像分类模型的训练更有帮助，可以给较大权重。若目标领域训练样本数据很容易被领域区分器区分开，则该目标领域训练样本数据对于图像分类模型的训练价值较小，可以减小它在损失函数的权重。如图14所示，其中领域区分器输出为0.5的样本权重最大，两边的权重依次减小，当达到一定值时，权重为0。该权重可以使用如下式表示：
Figure PCTCN2019088846-appb-000013
其中z为一个可以学习的参数，α是一个常数。基于这个公式，对样本的权重可以表示为
Figure PCTCN2019088846-appb-000014
可选的，对靠近目标领域图像数据的目标领域训练样本数据的权重取较大值。可以采用多种方法设置此类权重，例如对上式中若
Figure PCTCN2019088846-appb-000015
则将权重设置为
Figure PCTCN2019088846-appb-000016
所对应的权重值:
Figure PCTCN2019088846-appb-000017
通过目标领域训练样本数据选择与权重设置之后，可以针对目标领域训练样本数据建立分类损失函数，可以表示为
Figure PCTCN2019088846-appb-000018
其中
Figure PCTCN2019088846-appb-000019
为经过之前训练后的图像分类器在目标领域训练样本数据上的输出。从而，基于增强协同对抗网络的总体损失函数由三部分构成，即在源领域图像数据上的分类损失函数、在低层特征与高层特征上的协同对抗损失函数以及在目标领域训练样本数据上的分类损失函数，可以表示为：
Figure PCTCN2019088846-appb-000020
该总体损失函数可以使用基于随机梯度的反向传播方法进行优化，从而更新增强协同对抗网络中各部分的参数，训练图像分类模型，利用该图像分类模型用于目标领域图像数据的类别预测。在训练过程中，可以先使用源领域图像数据及类别标签，训练一个初始的协同对抗网络，在通过自适应目标领域训练样本数据选择308和权重设置309选择样本和设置权重后，与源领域图像数据共同再训练该初始的协同对抗网络。
需要注意的，图13中的低层特征提取303、高层特征提取304、图像分类305、域不变性306、域区分性307、样本数据选择308和权重设置309可以看作是增强协同对抗网络的组成模块，也可以看作是基于增强协同对抗网络的训练方法中操作步骤。
本申请实施例提供了一种芯片硬件结构，如图15所示，上面本申请实施例中描述的基于卷积神经网络的算法/方法(图12对应的实施例和图13对应的实施例中涉及的算法/方法)可以全部或部分在图15所示的NPU芯片中实现。
神经网络处理器NPU 50NPU作为协处理器挂载到主CPU(Host CPU)上，由Host CPU分配任务。NPU的核心部分为运算电路50，通过控制器504控制运算电路503提取存储器 中的矩阵数据并进行乘法运算。
在一些实现中，运算电路503内部包括多个处理单元(Process Engine,PE)。在一些实现中，运算电路503是二维脉动阵列。运算电路503还可以是一维脉动阵列或者能够执行例如乘法和加法这样的数学运算的其它电子线路。在一些实现中，运算电路503是通用的矩阵处理器。
举例来说，假设有输入矩阵A，权重矩阵B，输出矩阵C。运算电路从权重存储器502中取矩阵B相应的数据，并缓存在运算电路中每一个PE上。运算电路从输入存储器501中取矩阵A数据与矩阵B进行矩阵运算，得到的矩阵的部分结果或最终结果，保存在累加器508accumulator中。
统一存储器506用于存放输入数据以及输出数据。权重数据直接通过存储单元访问控制器505Direct Memory Access Controller，DMAC被搬运到权重存储器502中。输入数据也通过DMAC被搬运到统一存储器506中。
BIU为Bus Interface Unit即，总线接口单元510，用于AXI总线与DMAC和取指存储器509Instruction Fetch Buffer的交互。
总线接口单元510(Bus Interface Unit，简称BIU)，用于取指存储器509从外部存储器获取指令，还用于存储单元访问控制器505从外部存储器获取输入矩阵A或者权重矩阵B的原数据。
DMAC主要用于将外部存储器DDR中的输入数据搬运到统一存储器506或将权重数据搬运到权重存储器502中或将输入数据数据搬运到输入存储器501中。
向量计算单元507多个运算处理单元，在需要的情况下，对运算电路的输出做进一步处理，如向量乘，向量加，指数运算，对数运算，大小比较等等。主要用于神经网络中非卷积/FC层网络计算，如Pooling(池化)，Batch Normalization(批归一化)，Local Response Normalization(局部响应归一化)等。
在一些实现种，向量计算单元能507将经处理的输出的向量存储到统一缓存器506。例如，向量计算单元507可以将非线性函数应用到运算电路503的输出，例如累加值的向量，用以生成激活值。在一些实现中，向量计算单元507生成归一化的值、合并值，或二者均有。在一些实现中，处理过的输出的向量能够用作到运算电路503的激活输入，例如用于在神经网络中的后续层中的使用。
控制器504连接的取指存储器(instruction fetch buffer)509，用于存储控制器504使用的指令；
统一存储器506，输入存储器501，权重存储器502以及取指存储器509均为On-Chip存储器。外部存储器私有于该NPU硬件架构。
其中，卷积神经网络中各层的运算可以由矩阵计算单元212或向量计算单元507执行。
本申请实施例提供了一种训练设备410，如图16所示包括：处理器412、通信接口413、存储器411。可选地，训练设备410还可以包括总线414。其中，通信接口413、处理器412以及存储器411可以通过总线414相互连接；总线414可以是外设部件互连标准(英文：Peripheral Component Interconnect，简称PCI)总线或扩展工业标准结构(英文：Extended Industry Standard Architecture，简称EISA)总线等。上述总线414可以分为地址总线、数据总线、控制总线等。为便于表示，图16中仅用一条粗线表示，但并不表示仅有一根总 线或一种类型的总线。
上述图16所示的训练设备可以用于替代训练装置110以执行上面方法实施例中描述的方法，具体实现还可以对应参照上面方法实施例的相应描述，此处不再赘述。
结合本发明实施例公开内容所描述的方法或者算法的步骤可以硬件的方式来实现，也可以是由处理器执行软件指令的方式来实现。软件指令可以由相应的软件模块组成，软件模块可以被存放于随机存取存储器(英文：Random Access Memory，RAM)、闪存、只读存储器(英文：Read Only Memory，ROM)、可擦除可编程只读存储器(英文：Erasable Programmable ROM，EPROM)、电可擦可编程只读存储器(英文：Electrically EPROM，EEPROM)、寄存器、硬盘、移动硬盘、只读光盘(CD-ROM)或者本领域熟知的任何其它形式的存储介质中。一种示例性的存储介质耦合至处理器，从而使处理器能够从该存储介质读取信息，且可向该存储介质写入信息。当然，存储介质也可以是处理器的组成部分。处理器和存储介质可以位于ASIC中。另外，该ASIC可以位于网络设备中。当然，处理器和存储介质也可以作为分立组件存在于终端设备中。
按照本申请实施例提供的训练方法，在公开的标准数据集Office-31与ImageCLEF-DA上做迁移学习的测试。Office-31是物体识别的一个标准数据集，共包含4110张图片，其中有31个类别的物体。它包含四个领域的数据Amazon(A),Webcam(W),和Dlsr(D)。这里测试从其中任一领域迁移到另外一个领域的学习过程，评估迁移学习的精度。
ImageCLEF-DA是CLEF 2014年挑战赛的数据集，其中包含了三个领域的数据，即ImageNet ILSVRC2012(I),Bing(B),与Pascal VOC 2012(P)。每一个领域的数据都包含12个类别的数据，每个类别有50张图片。同样，这里测试从一个领域迁移到另外一个领域的识别精度，共6种迁移方式。
图17A和图17B给出了基于本申请实施例提供的方法与另外几种方法，如ResNet50、DANN、JAN的方法等的测试精度，并同时给出了平均迁移学习精度。可以看到，基于协同对抗网络的算法(CAN)获得了除JAN外最好的效果，而增强协同对抗网络(本发明)获得了最优效果，平均迁移精度比当前最好方法JAN高2～3个百分点。
因此，本申请实施例提供的基于增强协同对抗网络的训练方法基于高层特征提取和低层特征提取分别建立了域不变性损失函数和域区分性损失函数，在保证高层特征的域不变性特征的同时保留了低层特征中的域区分性特征，能够提高图像分类器应用到目标领域的图像分类预测的精度。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，上述的程序可存储于计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。而前述的存储介质包括：ROM、RAM、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述仅为本发明的几个实施例，本领域的技术人员依据申请文件公开的可以对本发明进行各种改动或变型而不脱离本发明的精神和范围。

Claims (22)

  1. 一种深度神经网络的训练方法，其特征在于，包括：
    提取源领域数据和目标领域数据中各样本数据的低层特征和高层特征，所述目标领域数据与所述源领域数据的数据分布不同；
    基于所述源领域数据和所述目标领域数据中各样本数据的高层特征和对应的领域标签，通过第一损失函数分别计算各样本数据对应的第一损失；
    基于所述源领域数据和所述目标领域数据中各样本数据的低层特征和对应的领域标签，通过第二损失函数分别计算各样本数据对应的第二损失；
    基于所述源领域数据中的样本数据的高层特征和对应的样本标签，通过第三损失函数计算所述源领域数据中的样本数据对应的第三损失；
    根据所述第一损失、所述第二损失和所述第三损失更新目标深度神经网络的参数，其中所述第一损失的梯度经过梯度反向，所述梯度反向可实现反向传导梯度使损失变大。
  2. 根据权利要求1所述的训练方法，其特征在于，所述目标深度神经网络包括特征提取模块、任务模块、域不变性特征模块和域区分性特征模块，所述特征提取模块包括至少一个低层特征网络层和高层特征网络层，所述至少一个低层特征网络层中的任一个低层特征网络层可用于提取低层特征，所述高层特征网络层用于提取高层特征，所述域不变性特征模块用于增强所述特征提取模块提取的高层特征的领域不变性，所述域区分性特征模块用于增强所述特征提取模块提取的低层特征的领域区分性；
    其中，所述根据所述第一损失、所述第二损失和所述第三损失更新目标深度神经网络的参数包括：
    根据所述第一损失、所述第二损失和所述第三损失计算总损失；
    根据所述总损失更新所述特征提取模块的参数、所述任务模块的参数、所述域不变性特征模块的参数和所述域区分性特征模块的参数。
  3. 根据权利要求2所述的训练方法，其特征在于，所述基于所述源领域数据和所述目标领域数据中各样本数据的高层特征和对应的领域标签，通过第一损失函数分别计算各样本数据对应的第一损失，包括：将所述源领域数据和所述目标领域数据中的各样本数据的高层特征输入所述域不变性特征模块得到各样本数据对应的第一结果；根据所述源领域数据和所述目标领域数据中的各样本数据对应的第一结果和对应的领域标签，通过所述第一损失函数分别计算各样本数据对应的第一损失；
    所述基于所述源领域数据和所述目标领域数据中各样本数据的低层特征和对应的领域标签，通过第二损失函数分别计算各样本数据对应的第二损失，包括：将所述源领域数据和所述目标领域数据中的各样本数据的低层特征输入所述域区分性特征模块得到各样本数据对应的第二结果；根据所述源领域数据和所述目标领域数据中的各样本数据对应的第二结果和对应的领域标签，通过所述第二损失函数分别计算各样本数据对应的第二损失；
    所述基于所述源领域数据中的样本数据的高层特征和对应的样本标签，通过第三损失函数计算所述源领域数据中的样本数据对应的第三损失，包括：将所述源领域数据中的样本数据的高层特征输入所述任务模块得到所述源领域数据中的样本数据对应的第三结果； 基于所述源领域数据中的样本数据对应的第三结果和对应的样本标签，通过第三损失函数计算所述源领域数据中的样本数据对应的第三损失。
  4. 根据权利要求2或3所述的训练方法，其特征在于，所述域不变性特征模块还包括：梯度反向模块；
    所述训练方法还包括：
    通过所述梯度反向模块对所述第一损失的梯度进行所述梯度反向。
  5. 根据权利要求3或4所述的训练方法，其特征在于，还包括：
    将所述目标领域数据中样本数据的高层特征输入所述任务模块，得到对应的预测样本标签和对应的置信度；
    根据所述目标领域数据中样本数据对应的置信度从所述目标领域数据中选定目标领域训练样本数据，所述目标领域训练样本数据为所述目标领域数据中对应的置信度满足预设条件的样本数据。
  6. 根据权利要求5所述的训练方法，其特征在于，还包括：
    根据所述目标领域训练样本数据对应的第一结果设置所述目标领域训练样本数据的权重。
  7. 根据权利要求6所述的训练方法，其特征在于，所述根据所述目标领域训练样本数据对应的第一结果设置所述目标领域训练样本数据的权重包括：
    根据所述目标领域训练样本数据对应的第一结果与领域标签的相似度，设置所述目标领域训练样本数据的权重，所述相似度表示第一结果与领域标签的差值大小。
  8. 根据权利要求7所述的训练方法，其特征在于，所述根据所述目标领域训练样本数据对应的第一结果与领域标签的相似度，设置所述目标领域训练样本数据的权重包括：
    计算所述目标领域训练样本数据对应的第一结果与源领域的领域标签的第一差值，以及所述目标领域训练样本数据对应的第一结果与目标领域的领域标签的第二差值；
    若所述第一差值的绝对值大于所述第二差值的绝对值，则设置所述目标领域训练样本数据的权重为较小的值，否则，设置所述目标领域训练样本数据的权重为较大的值。
  9. 根据权利要求7所述的训练方法，其特征在于，若所述目标领域训练样本数据对应的第一结果为第一领域标签值至第二领域标签值取值范围中的中间值，则设置所述目标领域训练样本数据的权重为最大值，所述第一领域标签值为源领域的领域标签对应的值，所述第二领域标签值为目标领域的领域标签对应的值。
  10. 根据权利要求5-9任选一所述的训练方法，其特征在于，在所述根据所述目标领域数据中样本数据对应的置信度从所述目标领域数据中选定目标领域训练样本数据之前，还包括：
    根据任务模型的精度设置自适应阈值，所述任务模型包括所述特征提取模块和所述任务模块，所述自适应阈值与所述任务模型的精度正相关；
    其中，所述预设条件为置信度大于或等于所述自适应阈值。
  11. 根据权利要求10所述的训练方法，其特征在于，所述自适应阈值通过下面逻辑函数计算：
    Figure PCTCN2019088846-appb-100001
    其中，所述T c为所述自适应阈值，所述A为所述任务模型的精度，λ c为用于控制所述逻辑函数的倾斜度的超参数。
  12. 根据权利要求5-11任选一所述的训练方法，其特征在于，所述训练方法还包括：
    通过所述特征提取模块提取所述目标领域训练样本数据的低层特征和高层特征；
    基于所述目标领域训练样本数据的高层特征和对应的领域标签，通过所述第一损失函数计算所述目标领域训练样本数据对应的第一损失；
    基于所述目标领域训练样本数据的低层特征和对应的领域标签，通过所述第二损失函数计算所述目标领域训练样本数据对应的第二损失；
    基于所述目标领域训练样本数据的高层特征和对应的预测样本标签，通过所述第三损失函数计算所述目标领域训练样本数据对应的第三损失；
    根据所述目标领域训练样本数据对应的第一损失、第二损失和第三损失计算所述目标领域训练样本数据对应的总损失，其中，所述目标领域训练样本数据对应的第一损失的梯度经过梯度反向；
    根据所述目标领域训练样本数据对应的总损失和所述目标领域训练样本数据的权重，更新所述特征提取模块的参数、所述任务模块的参数、所述域不变性特征模块的参数和所述域区分性特征模块的参数。
  13. 根据权利要求12所述的训练方法，其特征在于，所述基于所述目标领域训练样本数据的高层特征和对应的领域标签，通过所述第一损失函数计算所述目标领域训练样本数据对应的第一损失包括：将所述目标领域训练样本数据的高层特征输入所述域不变性特征模块得到所述目标领域训练样本数据对应的第一结果；根据所述目标领域训练样本数据对应的第一结果和对应的领域标签，通过所述第一损失函数计算所述目标领域训练样本数据对应的第一损失；
    所述基于所述目标领域训练样本数据的低层特征和对应的领域标签，通过所述第二损失函数计算所述目标领域训练样本数据对应的第二损失包括：将所述目标领域训练样本数据的低层特征输入所述域区分性特征模块得到所述目标领域训练样本数据对应的第二结果；根据所述目标领域训练样本数据对应的第二结果和对应的领域标签，通过所述第二损失函数计算所述目标领域训练样本数据对应的第二损失；
    所述基于所述目标领域训练样本数据的高层特征和对应的预测样本标签，通过第三损失函数计算所述目标领域训练样本数据对应的第三损失，包括：将所述目标领域训练样本数据的高层特征输入所述任务模块得到所述目标领域训练样本数据对应的第三结果；基于所述目标领域训练样本数据对应的第三结果和对应的预测样本标签，通过所述第三损失函数计算所述目标领域训练样本数据对应的第三损失。
  14. 一种训练设备，其特征在于，包括存储器及与所述存储器耦合的处理器；所述存储器用于存储指令，所述处理器用于执行所述指令；其中，所述处理器执行所述指令时执行如上权利要求1至13中任一项所述的方法。
  15. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1至13任一项所述方法。
  16. 一种增强协同对抗网络，其特征在于，所述增强协同对抗网络基于卷积神经网络CNN构建，包括：
    特征提取模块，用于提取源领域数据和目标领域数据中各样本数据的低层特征和高层特征，所述目标领域数据与所述源领域数据的数据分布不同；
    任务模块，用于接收所述特征提取模块输出的高层特征且通过第三损失函数分别计算各样本数据对应的第三损失，所述第三损失用于更新所述特征提取模块和所述任务模块的参数；
    域不变性模块，用于接收所述特征提取模块输出的高层特征且通过第一损失函数分别计算各样本数据对应的第一损失，所述第一损失用于更新所述特征提取模块和所述域不变性模块的参数，使得所述特征提取模块输出的高层特征具有域不变性；
    域区分性模块，用于接收所述特征提取模块输出的低层特征且通过第二损失函数分别计算各样本数据对应的第二损失，所述第二损失用于更新所述特征提取模块和所述域区分性模块的参数，使得所述特征提取模块输出的低层特征具有域区分性。
  17. 根据权利要求16所述的增强协同对抗网络，其特征在于，还包括：样本数据选择模块，用于根据所述目标领域数据中样本数据对应的置信度从所述目标领域数据中选定目标领域训练样本数据，所述目标领域数据中样本数据对应的置信度通过将所述目标领域数据中样本数据的高层特征输入所述任务模块得到，所述目标领域训练样本数据为所述目标领域数据中对应的置信度满足预设条件的样本数据。
  18. 根据权利要求17所述的增强协同对抗网络，其特征在于，所述样本数据选择模块还用于：根据任务模型的精度设置自适应阈值，所述任务模型包括所述特征提取模块和所述任务模块，所述自适应阈值与所述任务模型的精度正相关；其中，所述预设条件为置信度大于或等于所述自适应阈值。
  19. 根据权利要求17或18所述的增强协同对抗网络，其特征在于，还包括权重设置模块，用于根据所述目标领域训练样本数据对应的第一结果设置所述目标领域训练样本数据的权重。
  20. 根据权利要求19所述的增强协同对抗网络，其特征在于，所述权重设置模块具体用于：根据所述目标领域训练样本数据对应的第一结果与领域标签的相似度，设置所述目标领域训练样本数据的权重，所述相似度表示第一结果与领域标签的差值大小。
  21. 根据权利要求20所述的增强协同对抗网络，其特征在于，所述权重设置模块具体用于：计算所述目标领域训练样本数据对应的第一结果与源领域的领域标签的第一差值，以及所述目标领域训练样本数据对应的第一结果与目标领域的领域标签的第二差值；若第一差值的绝对值大于所述第二差值的绝对值，则设置所述目标领域训练样本数据的权重为较小的值，否则，设置所述目标领域训练样本数据的权重为较大的值。
  22. 根据权利要求20所述的增强协同对抗网络，其特征在于，所述权重设置模块具体用于：若所述目标领域训练样本数据对应的第一结果为第一领域标签值至第二领域标签值取值范围中的中间值，则设置所述目标领域训练样本数据的权重为最大值，所述第一领域标签值为源领域的领域标签对应的值，所述第二领域标签值为目标领域的领域标签对应的值。
PCT/CN2019/088846 2018-05-31 2019-05-28 深度神经网络的训练方法和装置 Ceased WO2019228358A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP19812148.5A EP3757905A4 (en) 2018-05-31 2019-05-28 METHOD AND DEVICE FOR TRAINING A DEEP NEURAL NETWORK
US17/033,316 US20210012198A1 (en) 2018-05-31 2020-09-25 Method for training deep neural network and apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810554459.4A CN109902798A (zh) 2018-05-31 2018-05-31 深度神经网络的训练方法和装置
CN201810554459.4 2018-05-31

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/033,316 Continuation US20210012198A1 (en) 2018-05-31 2020-09-25 Method for training deep neural network and apparatus

Publications (1)

Publication Number Publication Date
WO2019228358A1 true WO2019228358A1 (zh) 2019-12-05

Family

ID=66943222

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/088846 Ceased WO2019228358A1 (zh) 2018-05-31 2019-05-28 深度神经网络的训练方法和装置

Country Status (4)

Country Link
US (1) US20210012198A1 (zh)
EP (1) EP3757905A4 (zh)
CN (1) CN109902798A (zh)
WO (1) WO2019228358A1 (zh)

Cited By (33)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110852450A (zh) * 2020-01-15 2020-02-28 支付宝(杭州)信息技术有限公司 识别对抗样本以保护模型安全的方法及装置
CN111461191A (zh) * 2020-03-25 2020-07-28 杭州跨视科技有限公司 为模型训练确定图像样本集的方法、装置和电子设备
CN111680754A (zh) * 2020-06-11 2020-09-18 北京字节跳动网络技术有限公司 图像分类方法、装置、电子设备及计算机可读存储介质
CN111832605A (zh) * 2020-05-22 2020-10-27 北京嘀嘀无限科技发展有限公司 无监督图像分类模型的训练方法、装置和电子设备
CN111914912A (zh) * 2020-07-16 2020-11-10 天津大学 一种基于孪生条件对抗网络的跨域多视目标识别方法
CN112001398A (zh) * 2020-08-26 2020-11-27 科大讯飞股份有限公司 域适应方法、装置、设备、图像处理方法及存储介质
CN112115976A (zh) * 2020-08-20 2020-12-22 北京嘀嘀无限科技发展有限公司 模型训练方法、模型训练装置、存储介质和电子设备
CN112241452A (zh) * 2020-10-16 2021-01-19 百度(中国)有限公司 一种模型训练方法、装置、电子设备及存储介质
CN112364860A (zh) * 2020-11-05 2021-02-12 北京字跳网络技术有限公司 字符识别模型的训练方法、装置和电子设备
CN112580733A (zh) * 2020-12-25 2021-03-30 北京百度网讯科技有限公司 分类模型的训练方法、装置、设备以及存储介质
CN112633579A (zh) * 2020-12-24 2021-04-09 中国科学技术大学 一种基于域对抗的交通流迁移预测方法
CN112634048A (zh) * 2020-12-30 2021-04-09 第四范式（北京）技术有限公司 一种反洗钱模型的训练方法及装置
CN112749758A (zh) * 2021-01-21 2021-05-04 北京百度网讯科技有限公司 图像处理方法、神经网络的训练方法、装置、设备和介质
CN112784776A (zh) * 2021-01-26 2021-05-11 山西三友和智慧信息技术股份有限公司 一种基于改进残差网络的bpd面部情绪识别方法
CN112861977A (zh) * 2021-02-19 2021-05-28 中国人民武装警察部队工程大学 迁移学习数据处理方法、系统、介质、设备、终端及应用
CN112884147A (zh) * 2021-02-26 2021-06-01 上海商汤智能科技有限公司 神经网络训练方法、图像处理方法、装置及电子设备
CN112966345A (zh) * 2021-03-03 2021-06-15 北京航空航天大学 基于对抗训练和迁移学习的旋转机械剩余寿命预测混合收缩方法
CN112990298A (zh) * 2021-03-11 2021-06-18 北京中科虹霸科技有限公司 关键点检测模型训练方法、关键点检测方法及装置
CN113033549A (zh) * 2021-03-09 2021-06-25 北京百度网讯科技有限公司 定位图获取模型的训练方法和装置
CN113111776A (zh) * 2021-04-12 2021-07-13 京东数字科技控股股份有限公司 对抗样本的生成方法、装置、设备及存储介质
CN113222100A (zh) * 2020-02-05 2021-08-06 北京三星通信技术研究有限公司 神经网络模型的训练方法和装置
CN113286311A (zh) * 2021-04-29 2021-08-20 沈阳工业大学 基于多传感器融合的分布式周界安防环境感知系统
CN113792576A (zh) * 2021-07-27 2021-12-14 北京邮电大学 基于有监督域适应的人体行为识别方法、电子设备
CN113869193A (zh) * 2021-09-26 2021-12-31 平安科技（深圳）有限公司 行人再识别模型的训练方法、行人再识别方法及系统
CN113989595A (zh) * 2021-11-05 2022-01-28 西安交通大学 一种基于阴影模型的联邦多源域适应方法及系统
CN114202028A (zh) * 2021-12-13 2022-03-18 四川大学 基于mamtl的滚动轴承寿命阶段识别方法
WO2022069991A1 (en) * 2020-09-30 2022-04-07 International Business Machines Corporation Outlier detection in deep neural network
CN114330665A (zh) * 2021-12-30 2022-04-12 上海高德威智能交通系统有限公司 增量学习方法、装置、电子设备及机器可读存储介质
CN114662675A (zh) * 2022-03-31 2022-06-24 上海易康源医疗健康科技有限公司 神经网络训练方法、图片分类方法及图片分类系统
CN114998602A (zh) * 2022-08-08 2022-09-02 中国科学技术大学 基于低置信度样本对比损失的域适应学习方法及系统
CN115623870A (zh) * 2021-05-14 2023-01-17 华为技术有限公司 参数标定模型、训练方法、参数标定装置和车辆
CN116229208A (zh) * 2023-02-20 2023-06-06 电子科技大学 一种基于自适应加权的深度度量学习对抗防御方法
US12475896B1 (en) * 2025-03-24 2025-11-18 Zhejiang Gongshang University Method for cross-domain speech deepfake detection

Families Citing this family (89)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11087142B2 (en) * 2018-09-13 2021-08-10 Nec Corporation Recognizing fine-grained objects in surveillance camera images
US11222210B2 (en) * 2018-11-13 2022-01-11 Nec Corporation Attention and warping based domain adaptation for videos
GB201819434D0 (en) * 2018-11-29 2019-01-16 Kheiron Medical Tech Ltd Domain adaptation
KR102039138B1 (ko) * 2019-04-02 2019-10-31 주식회사 루닛 적대적 학습에 기반한 도메인 어댑테이션 방법 및 그 장치
CN112633459B (zh) * 2019-09-24 2024-09-20 华为技术有限公司 训练神经网络的方法、数据处理方法和相关装置
CN110674648B (zh) * 2019-09-29 2021-04-27 厦门大学 基于迭代式双向迁移的神经网络机器翻译模型
KR102869896B1 (ko) * 2019-12-12 2025-10-14 삼성전자주식회사 도메인 적응에 기반한 네트워크의 트레이닝 방법, 동작 방법 및 동작 장치
CN111178401B (zh) * 2019-12-16 2023-09-12 上海航天控制技术研究所 一种基于多层对抗网络的空间目标分类方法
CN111091127B (zh) * 2019-12-16 2025-03-07 腾讯科技（深圳）有限公司 一种图像检测方法、网络模型训练方法以及相关装置
KR102537207B1 (ko) * 2019-12-30 2023-05-25 포항공과대학교 산학협력단 머신 러닝에 기반한 이미지 처리 방법 및 장치
US11537901B2 (en) * 2019-12-31 2022-12-27 Robert Bosch Gmbh System and method for unsupervised domain adaptation with mixup training
WO2021136939A1 (en) 2020-01-03 2021-07-08 Tractable Ltd Method of determining repair operations for a damaged vehicle
CN111239137B (zh) * 2020-01-09 2021-09-10 江南大学 基于迁移学习与自适应深度卷积神经网络的谷物质量检测方法
US11200883B2 (en) 2020-01-10 2021-12-14 International Business Machines Corporation Implementing a domain adaptive semantic role labeler
CN111442926B (zh) * 2020-01-11 2021-09-21 哈尔滨理工大学 一种基于深层特征迁移的变负载下不同型号滚动轴承故障诊断方法
CN113191478B (zh) * 2020-01-14 2025-02-14 阿里巴巴集团控股有限公司 神经网络模型的训练方法、装置和系统
CN113379045B (zh) * 2020-02-25 2022-08-09 华为技术有限公司 数据增强方法和装置
CN111444958B (zh) * 2020-03-25 2024-02-13 北京百度网讯科技有限公司 一种模型迁移训练方法、装置、设备及存储介质
CN111598124B (zh) * 2020-04-07 2022-11-11 深圳市商汤科技有限公司 图像处理及装置、处理器、电子设备、存储介质
CN111523649B (zh) * 2020-05-09 2022-06-10 支付宝(杭州)信息技术有限公司 针对业务模型进行数据预处理的方法及装置
CN111723691B (zh) * 2020-06-03 2023-10-17 合肥的卢深视科技有限公司 一种三维人脸识别方法、装置、电子设备及存储介质
CN111783844B (zh) * 2020-06-10 2024-05-28 广东正扬传感科技股份有限公司 基于深度学习的目标检测模型训练方法、设备及存储介质
US11514326B2 (en) * 2020-06-18 2022-11-29 International Business Machines Corporation Drift regularization to counteract variation in drift coefficients for analog accelerators
CN112232506B (zh) * 2020-09-10 2025-01-10 北京迈格威科技有限公司 网络模型训练方法、图像目标识别方法、装置和电子设备
CN112052818B (zh) * 2020-09-15 2024-03-22 浙江智慧视频安防创新中心有限公司 无监督域适应的行人检测方法、系统及存储介质
CN114254790B (zh) * 2020-09-23 2024-12-20 新奥新智科技有限公司 一种设备状态预测方法及装置
CN114529748B (zh) * 2020-11-02 2025-08-05 紫东信息科技(苏州)有限公司 一种图像分类模型训练方法及图像分类方法、系统
CN112426161B (zh) * 2020-11-17 2021-09-07 浙江大学 一种基于领域自适应的时变脑电特征提取方法
CN112528631B (zh) * 2020-12-03 2022-08-09 上海谷均教育科技有限公司 一种基于深度学习算法的智能伴奏系统
CN112632351B (zh) * 2020-12-28 2024-01-16 北京百度网讯科技有限公司 分类模型的训练方法、分类方法、装置及设备
GB2608344A (en) * 2021-01-12 2022-12-28 Zhejiang Lab Domain-invariant feature-based meta-knowledge fine-tuning method and platform
CN114724101B (zh) * 2021-01-12 2024-07-23 北京航空航天大学 基于批标准化的多空间对抗样本防御方法及其装置
CN112364945B (zh) * 2021-01-12 2021-04-16 之江实验室 一种基于域-不变特征的元-知识微调方法及平台
CN112818833B (zh) * 2021-01-29 2024-04-12 中能国际建筑投资集团有限公司 基于深度学习的人脸多任务检测方法、系统、装置及介质
CN113159095B (zh) 2021-01-30 2024-04-30 华为技术有限公司 一种训练模型的方法、图像检索的方法以及装置
CN113031437B (zh) * 2021-02-26 2022-10-25 同济大学 一种基于动态模型强化学习的倒水服务机器人控制方法
CN113065633B (zh) * 2021-02-26 2024-07-09 华为技术有限公司 一种模型训练方法及其相关联设备
CN113052295B (zh) * 2021-02-27 2024-04-12 华为技术有限公司 一种神经网络的训练方法、物体检测方法、装置及设备
CN115049869A (zh) * 2021-03-09 2022-09-13 西交利物浦大学 图像分类方法，参数训练方法和图像分类装置
CN113076834B (zh) * 2021-03-25 2022-05-13 华中科技大学 旋转机械故障信息处理方法、处理系统、处理终端、介质
CN112989702B (zh) * 2021-03-25 2022-08-02 河北工业大学 一种装备性能分析与预测的自学习方法
CN113158364B (zh) * 2021-04-02 2024-03-22 中国农业大学 循环泵轴承故障检测方法及系统
CN113132931B (zh) * 2021-04-16 2022-01-28 电子科技大学 一种基于参数预测的深度迁移室内定位方法
CN113239975B (zh) * 2021-04-21 2022-12-20 国网甘肃省电力公司白银供电公司 一种基于神经网络的目标检测方法和装置
CN113112005B (zh) * 2021-04-27 2024-07-19 南京大学 一种基于注意力机制的领域自适应方法
CN113128478B (zh) * 2021-05-18 2023-07-14 电子科技大学中山学院 模型训练方法、行人分析方法、装置、设备及存储介质
CN113158985B (zh) * 2021-05-18 2024-05-14 深圳市创智链科技有限公司 一种分类识别的方法和设备
CN113269261B (zh) * 2021-05-31 2024-03-12 国网福建省电力有限公司电力科学研究院 一种配网波形智能分类方法
CN113344119B (zh) * 2021-06-28 2024-07-12 南京邮电大学 工业物联网复杂环境下的小样本烟雾监测方法
CN113748430B (zh) * 2021-06-28 2024-05-24 商汤国际私人有限公司 对象检测网络的训练与检测方法、装置、设备和存储介质
WO2023275603A1 (en) * 2021-06-28 2023-01-05 Sensetime International Pte. Ltd. Methods, apparatuses, devices and storage media for training object detection network and for detecting object
CN113505834B (zh) * 2021-07-13 2024-11-05 阿波罗智能技术(北京)有限公司 训练检测模型、确定图像更新信息和更新高精地图的方法
CN113673570A (zh) * 2021-07-21 2021-11-19 南京旭锐软件科技有限公司 电子器件图片分类模型的训练方法、装置及设备
CN113657651B (zh) * 2021-07-27 2024-07-09 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) 基于深度迁移学习的柴油车排放预测方法、介质及设备
CN113591736B (zh) * 2021-08-03 2024-12-03 北京百度网讯科技有限公司 特征提取网络、活体检测模型的训练方法和活体检测方法
CN113610219B (zh) * 2021-08-16 2024-05-14 中国石油大学(华东) 一种基于动态残差的多源域自适应方法
CN113807183B (zh) * 2021-08-17 2024-06-14 华为技术有限公司 模型训练方法及相关设备
CN113792609B (zh) * 2021-08-19 2024-12-06 北京理工大学 自动驾驶中通过对齐多频谱跨域特征检测图像异常的方法
CN113948093B (zh) * 2021-10-19 2024-03-26 南京航空航天大学 一种基于无监督场景适应的说话人识别方法及系统
CN114048568B (zh) * 2021-11-17 2024-04-09 大连理工大学 一种基于多源迁移融合收缩框架的旋转机械故障诊断方法
CN114111942B (zh) * 2021-11-25 2024-07-16 宁夏隆基宁光仪表股份有限公司 基于无磁采样的无磁智能水表计量方法、系统和计量设备
CN114354195A (zh) * 2021-12-31 2022-04-15 南京工业大学 一种深度域自适应卷积网络的滚动轴承故障诊断方法
CN114373100B (zh) * 2022-01-06 2025-02-21 中南大学 一种浮选工况跨域识别方法
CN114443864B (zh) * 2022-01-29 2025-11-21 北京百度网讯科技有限公司 跨模态数据的匹配方法、装置及计算机程序产品
CN114726394B (zh) * 2022-03-01 2022-09-02 深圳前海梵天通信技术有限公司 一种智能通信系统的训练方法及智能通信系统
CN114663725B (zh) * 2022-03-21 2025-07-11 合肥工业大学 多领域自适应模型的训练方法、系统、设备及存储介质
CN114821250B (zh) * 2022-03-23 2024-12-24 支付宝(杭州)信息技术有限公司 一种跨域的模型训练方法、装置以及设备
CN114841246A (zh) * 2022-04-07 2022-08-02 深圳无域科技技术有限公司 一种模型构建方法、装置、电子设备及存储介质
CN114741732B (zh) * 2022-04-28 2025-02-07 重庆长安汽车股份有限公司 一种基于隐私数据保护的智能网联汽车数据训练方法、电子设备及计算机可读存储介质
CN115115060A (zh) * 2022-05-30 2022-09-27 腾讯科技（武汉）有限公司 一种模型的确定方法和相关装置
CN115049627B (zh) * 2022-06-21 2023-06-20 江南大学 基于域自适应深度迁移网络的钢表面缺陷检测方法及系统
CN116010840B (zh) * 2022-11-18 2026-02-24 西安电子科技大学 一种多源域样本重加权的eeg信号跨设备解码方法
CN116206345A (zh) * 2022-12-09 2023-06-02 支付宝(杭州)信息技术有限公司 表情识别模型训练方法、表情识别方法、相关装置及介质
CN116109449A (zh) * 2022-12-09 2023-05-12 华为技术有限公司 一种数据处理方法及相关设备
CN116149369B (zh) * 2023-02-23 2025-11-07 北京理工大学 面向多种目标域任务的飞行器制导控制方法
CN116468096B (zh) * 2023-03-30 2024-01-02 之江实验室 一种模型训练方法、装置、设备及可读存储介质
CN117093929B (zh) * 2023-07-06 2024-03-29 珠海市伊特高科技有限公司 基于无监督域自适应网络的截流过电压预测方法及装置
CN116578924A (zh) * 2023-07-12 2023-08-11 太极计算机股份有限公司 一种用于机器学习分类的网络任务优化方法及系统
CN116630630B (zh) * 2023-07-24 2023-12-15 深圳思谋信息科技有限公司 语义分割方法、装置、计算机设备及计算机可读存储介质
CN116737607B (zh) * 2023-08-16 2023-11-21 之江实验室 样本数据缓存方法、系统、计算机设备和存储介质
CN116882486B (zh) * 2023-09-05 2023-11-14 浙江大华技术股份有限公司 一种迁移学习权重的构建方法和装置及设备
CN117152563B (zh) * 2023-10-16 2024-05-14 华南师范大学 混合目标域自适应模型的训练方法、装置及计算机设备
CN117435916B (zh) * 2023-12-18 2024-03-12 四川云实信息技术有限公司 航片ai解译中的自适应迁移学习方法
CN117609887B (zh) * 2024-01-19 2024-05-10 腾讯科技（深圳）有限公司 数据增强模型训练及数据处理方法、装置、设备、介质
CN118643735B (zh) * 2024-05-28 2025-04-01 湖南大学 土舱压力预测模型迁移学习方法、装置、设备及存储介质
CN118468724B (zh) * 2024-05-30 2025-04-11 苏州大学 差异性对齐指导下域对抗网络机械设备rul预测方法及系统
CN119249090B (zh) * 2024-08-26 2025-10-10 华中科技大学 对偶域自适应双任务学习的机器人磨削预测方法及设备
CN121071409B (zh) * 2025-11-06 2026-02-27 国网浙江省电力有限公司超高压分公司 一种换流变压器绝缘套管绝缘状态评估方法及装置
CN121327144B (zh) * 2025-11-10 2026-04-21 华南理工大学 基于动态难例感知的智能司法多任务预测方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170220951A1 (en) * 2016-02-02 2017-08-03 Xerox Corporation Adapting multiple source classifiers in a target domain
CN107633242A (zh) * 2017-10-23 2018-01-26 广州视源电子科技股份有限公司 网络模型的训练方法、装置、设备和存储介质
CN107958287A (zh) * 2017-11-23 2018-04-24 清华大学 面向跨界大数据分析的对抗迁移学习方法及系统
CN108009633A (zh) * 2017-12-15 2018-05-08 清华大学 一种面向跨领域智能分析的多网络对抗学习方法和系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170220951A1 (en) * 2016-02-02 2017-08-03 Xerox Corporation Adapting multiple source classifiers in a target domain
CN107633242A (zh) * 2017-10-23 2018-01-26 广州视源电子科技股份有限公司 网络模型的训练方法、装置、设备和存储介质
CN107958287A (zh) * 2017-11-23 2018-04-24 清华大学 面向跨界大数据分析的对抗迁移学习方法及系统
CN108009633A (zh) * 2017-12-15 2018-05-08 清华大学 一种面向跨领域智能分析的多网络对抗学习方法和系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3757905A4 *

Cited By (58)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110852450A (zh) * 2020-01-15 2020-02-28 支付宝(杭州)信息技术有限公司 识别对抗样本以保护模型安全的方法及装置
CN110852450B (zh) * 2020-01-15 2020-04-14 支付宝(杭州)信息技术有限公司 识别对抗样本以保护模型安全的方法及装置
CN113222100A (zh) * 2020-02-05 2021-08-06 北京三星通信技术研究有限公司 神经网络模型的训练方法和装置
CN111461191A (zh) * 2020-03-25 2020-07-28 杭州跨视科技有限公司 为模型训练确定图像样本集的方法、装置和电子设备
CN111461191B (zh) * 2020-03-25 2024-01-23 杭州跨视科技有限公司 为模型训练确定图像样本集的方法、装置和电子设备
CN111832605A (zh) * 2020-05-22 2020-10-27 北京嘀嘀无限科技发展有限公司 无监督图像分类模型的训练方法、装置和电子设备
CN111832605B (zh) * 2020-05-22 2023-12-08 北京嘀嘀无限科技发展有限公司 无监督图像分类模型的训练方法、装置和电子设备
CN111680754B (zh) * 2020-06-11 2023-09-19 抖音视界有限公司 图像分类方法、装置、电子设备及计算机可读存储介质
CN111680754A (zh) * 2020-06-11 2020-09-18 北京字节跳动网络技术有限公司 图像分类方法、装置、电子设备及计算机可读存储介质
CN111914912B (zh) * 2020-07-16 2023-06-13 天津大学 一种基于孪生条件对抗网络的跨域多视目标识别方法
CN111914912A (zh) * 2020-07-16 2020-11-10 天津大学 一种基于孪生条件对抗网络的跨域多视目标识别方法
CN112115976A (zh) * 2020-08-20 2020-12-22 北京嘀嘀无限科技发展有限公司 模型训练方法、模型训练装置、存储介质和电子设备
CN112115976B (zh) * 2020-08-20 2023-12-08 北京嘀嘀无限科技发展有限公司 模型训练方法、模型训练装置、存储介质和电子设备
CN112001398B (zh) * 2020-08-26 2024-04-12 科大讯飞股份有限公司 域适应方法、装置、设备、图像处理方法及存储介质
CN112001398A (zh) * 2020-08-26 2020-11-27 科大讯飞股份有限公司 域适应方法、装置、设备、图像处理方法及存储介质
US12493667B2 (en) 2020-09-30 2025-12-09 International Business Machines Corporation Outlier detection in a deep neural network using t-way feature combinations
GB2617915A (en) * 2020-09-30 2023-10-25 Ibm Outlier detection in deep neural network
WO2022069991A1 (en) * 2020-09-30 2022-04-07 International Business Machines Corporation Outlier detection in deep neural network
CN112241452A (zh) * 2020-10-16 2021-01-19 百度(中国)有限公司 一种模型训练方法、装置、电子设备及存储介质
CN112241452B (zh) * 2020-10-16 2024-01-05 百度(中国)有限公司 一种模型训练方法、装置、电子设备及存储介质
CN112364860A (zh) * 2020-11-05 2021-02-12 北京字跳网络技术有限公司 字符识别模型的训练方法、装置和电子设备
CN112633579A (zh) * 2020-12-24 2021-04-09 中国科学技术大学 一种基于域对抗的交通流迁移预测方法
CN112633579B (zh) * 2020-12-24 2024-01-12 中国科学技术大学 一种基于域对抗的交通流迁移预测方法
CN112580733B (zh) * 2020-12-25 2024-03-05 北京百度网讯科技有限公司 分类模型的训练方法、装置、设备以及存储介质
CN112580733A (zh) * 2020-12-25 2021-03-30 北京百度网讯科技有限公司 分类模型的训练方法、装置、设备以及存储介质
CN112634048A (zh) * 2020-12-30 2021-04-09 第四范式（北京）技术有限公司 一种反洗钱模型的训练方法及装置
CN112634048B (zh) * 2020-12-30 2023-06-13 第四范式（北京）技术有限公司 一种反洗钱模型的训练方法及装置
CN112749758A (zh) * 2021-01-21 2021-05-04 北京百度网讯科技有限公司 图像处理方法、神经网络的训练方法、装置、设备和介质
CN112749758B (zh) * 2021-01-21 2023-08-11 北京百度网讯科技有限公司 图像处理方法、神经网络的训练方法、装置、设备和介质
CN112784776A (zh) * 2021-01-26 2021-05-11 山西三友和智慧信息技术股份有限公司 一种基于改进残差网络的bpd面部情绪识别方法
CN112861977B (zh) * 2021-02-19 2024-01-26 中国人民武装警察部队工程大学 迁移学习数据处理方法、系统、介质、设备、终端及应用
CN112861977A (zh) * 2021-02-19 2021-05-28 中国人民武装警察部队工程大学 迁移学习数据处理方法、系统、介质、设备、终端及应用
CN112884147B (zh) * 2021-02-26 2023-11-28 上海商汤智能科技有限公司 神经网络训练方法、图像处理方法、装置及电子设备
CN112884147A (zh) * 2021-02-26 2021-06-01 上海商汤智能科技有限公司 神经网络训练方法、图像处理方法、装置及电子设备
CN112966345B (zh) * 2021-03-03 2022-06-07 北京航空航天大学 基于对抗训练和迁移学习的旋转机械剩余寿命预测混合收缩方法
CN112966345A (zh) * 2021-03-03 2021-06-15 北京航空航天大学 基于对抗训练和迁移学习的旋转机械剩余寿命预测混合收缩方法
CN113033549A (zh) * 2021-03-09 2021-06-25 北京百度网讯科技有限公司 定位图获取模型的训练方法和装置
CN113033549B (zh) * 2021-03-09 2022-09-20 北京百度网讯科技有限公司 定位图获取模型的训练方法和装置
CN112990298B (zh) * 2021-03-11 2023-11-24 北京中科虹霸科技有限公司 关键点检测模型训练方法、关键点检测方法及装置
CN112990298A (zh) * 2021-03-11 2021-06-18 北京中科虹霸科技有限公司 关键点检测模型训练方法、关键点检测方法及装置
CN113111776B (zh) * 2021-04-12 2024-04-16 京东科技控股股份有限公司 对抗样本的生成方法、装置、设备及存储介质
CN113111776A (zh) * 2021-04-12 2021-07-13 京东数字科技控股股份有限公司 对抗样本的生成方法、装置、设备及存储介质
CN113286311B (zh) * 2021-04-29 2024-04-12 沈阳工业大学 基于多传感器融合的分布式周界安防环境感知系统
CN113286311A (zh) * 2021-04-29 2021-08-20 沈阳工业大学 基于多传感器融合的分布式周界安防环境感知系统
CN115623870A (zh) * 2021-05-14 2023-01-17 华为技术有限公司 参数标定模型、训练方法、参数标定装置和车辆
CN113792576B (zh) * 2021-07-27 2023-07-18 北京邮电大学 基于有监督域适应的人体行为识别方法、电子设备
CN113792576A (zh) * 2021-07-27 2021-12-14 北京邮电大学 基于有监督域适应的人体行为识别方法、电子设备
CN113869193A (zh) * 2021-09-26 2021-12-31 平安科技（深圳）有限公司 行人再识别模型的训练方法、行人再识别方法及系统
CN113989595A (zh) * 2021-11-05 2022-01-28 西安交通大学 一种基于阴影模型的联邦多源域适应方法及系统
CN113989595B (zh) * 2021-11-05 2024-05-07 西安交通大学 一种基于阴影模型的联邦多源域适应方法及系统
CN114202028A (zh) * 2021-12-13 2022-03-18 四川大学 基于mamtl的滚动轴承寿命阶段识别方法
CN114202028B (zh) * 2021-12-13 2023-04-28 四川大学 基于mamtl的滚动轴承寿命阶段识别方法
CN114330665A (zh) * 2021-12-30 2022-04-12 上海高德威智能交通系统有限公司 增量学习方法、装置、电子设备及机器可读存储介质
CN114662675A (zh) * 2022-03-31 2022-06-24 上海易康源医疗健康科技有限公司 神经网络训练方法、图片分类方法及图片分类系统
CN114998602A (zh) * 2022-08-08 2022-09-02 中国科学技术大学 基于低置信度样本对比损失的域适应学习方法及系统
CN116229208A (zh) * 2023-02-20 2023-06-06 电子科技大学 一种基于自适应加权的深度度量学习对抗防御方法
CN116229208B (zh) * 2023-02-20 2025-09-26 电子科技大学 一种基于自适应加权的深度度量学习对抗防御方法
US12475896B1 (en) * 2025-03-24 2025-11-18 Zhejiang Gongshang University Method for cross-domain speech deepfake detection

Also Published As

Publication number Publication date
EP3757905A4 (en) 2021-04-28
US20210012198A1 (en) 2021-01-14
EP3757905A1 (en) 2020-12-30
CN109902798A (zh) 2019-06-18

Similar Documents

Publication Publication Date Title
WO2019228358A1 (zh) 深度神经网络的训练方法和装置
US20220375213A1 (en) Processing Apparatus and Method and Storage Medium
WO2021190451A1 (zh) 训练图像处理模型的方法和装置
CN110222717B (zh) 图像处理方法和装置
CN113705769A (zh) 一种神经网络训练方法以及装置
WO2022042713A1 (zh) 一种用于计算设备的深度学习训练方法和装置
WO2020221200A1 (zh) 神经网络的构建方法、图像处理方法及装置
WO2022179492A1 (zh) 一种卷积神经网络的剪枝处理方法、数据处理方法及设备
WO2022042002A1 (zh) 一种半监督学习模型的训练方法、图像处理方法及设备
CN111797893A (zh) 一种神经网络的训练方法、图像分类系统及相关设备
WO2021043112A1 (zh) 图像分类方法以及装置
WO2022111617A1 (zh) 一种模型训练方法及装置
CN111310604A (zh) 一种物体检测方法、装置以及存储介质
WO2022001805A1 (zh) 一种神经网络蒸馏方法及装置
CN112529146A (zh) 神经网络模型训练的方法和装置
CN112580720A (zh) 一种模型训练方法及装置
WO2021103731A1 (zh) 一种语义分割方法、模型训练方法及装置
CN110222718B (zh) 图像处理的方法及装置
WO2021008206A1 (zh) 神经网络结构的搜索方法、图像处理方法和装置
CN113011568A (zh) 一种模型的训练方法、数据处理方法及设备
CN111738403A (zh) 一种神经网络的优化方法及相关设备
CN114266897A (zh) 痘痘类别的预测方法、装置、电子设备及存储介质
WO2021136058A1 (zh) 一种处理视频的方法及装置
CN114169393B (zh) 一种图像分类方法及其相关设备
CN113536970A (zh) 一种视频分类模型的训练方法及相关装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19812148

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2019812148

Country of ref document: EP

Effective date: 20200922

NENP Non-entry into the national phase

Ref country code: DE