Data screening method, device, equipment and medium

Abstract

The present application discloses a data screening method, device, equipment and medium, which relates to the field of machine learning technology. The method includes: obtaining initial data and a first classification network; determining the cross entropy loss corresponding to the initial data based on the prediction result obtained by classifying and predicting the initial data by the first classification network; performing feature sampling on the initial data to determine the reconstruction error information corresponding to the initial data, the reconstruction error information is used to indicate the correlation between the initial data and the classification task; based on the cross entropy loss and the reconstruction error information, the initial data is screened to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain a second classification network for processing the classification task. By combining the cross entropy loss and the reconstruction error information, the initial data can be jointly screened from the uncertainty dimension and the representative dimension, helping the network to resist the interference of noisy data during the training process.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a data screening method, device, equipment and medium. Background Art

[0002] The selection of training data is crucial during the training of deep learning networks, as its quality directly impacts the network's training effectiveness. Typically, human annotators label the data in a dataset based on the training task for the network being trained. However, manual labeling can lead to errors due to carelessness or incompetence. For example, in classification tasks, data may be labeled with the wrong category label. Incorrectly labeled data, also known as noise data, can significantly impact the training of deep learning networks.

[0003] In related art, decoupling is used during network training to reduce noise in training data. Specifically, two networks are trained simultaneously, where the only difference between the two networks is their initialization parameters. Backpropagation updates are then performed only when the two networks diverge, improving network robustness.

[0004] However, during the network training process implemented in the above manner, if the batch data initially input to the network includes biased training samples, the errors for a single network will be trained again during the second training process, resulting in incorrect training of model parameters by noise data, and the trained network accuracy is low. Summary of the Invention

[0005] The present invention provides a method, apparatus, device, and medium for filtering data, which can reduce the negative impact of noise data in training data on network training. The technical solution is as follows:

[0006] In one aspect, a method for screening data is provided, the method comprising:

[0007] Acquire initial data and a first classification network, where the first classification network is a network to be trained for processing a classification task;

[0008] Determining a cross entropy loss corresponding to the initial data based on a prediction result obtained by performing classification prediction on the initial data by the first classification network;

[0009] Performing feature sampling on the initial data to determine reconstruction error information corresponding to the initial data, where the reconstruction error information is used to indicate a degree of association between the initial data and the classification task;

[0010] Based on the cross entropy loss and the reconstruction error information, the initial data is screened to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain a second classification network for processing the classification task.

[0011] In another aspect, a data screening device is provided, the device comprising:

[0012] An acquisition module, configured to acquire initial data and a first classification network, where the first classification network is a network to be trained for processing a classification task;

[0013] a determination module, configured to determine a cross entropy loss corresponding to the initial data based on a prediction result obtained by performing classification prediction on the initial data by the first classification network;

[0014] The determination module is further configured to perform feature sampling on the initial data to determine reconstruction error information corresponding to the initial data, where the reconstruction error information is used to indicate a degree of association between the initial data and the classification task;

[0015] A screening module is used to screen the initial data based on the cross entropy loss and the reconstruction error information to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain a second classification network for processing the classification task.

[0016] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the data screening method described in any one of the embodiments of the present application.

[0017] On the other hand, a computer-readable storage medium is provided, in which at least one program code is stored. The program code is loaded and executed by a processor to implement the method for screening data of a terminal device described in any one of the embodiments of the present application.

[0018] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the data screening method described in any of the above embodiments.

[0019] The technical solution provided by this application includes at least the following beneficial effects:

[0020] During the training of a neural network for a classification task, initial data is input into the network to be trained. A cross-entropy loss corresponding to the initial data is obtained based on the output prediction results. Simultaneously, feature sampling is performed on the initial data to determine reconstruction error information that indicates the degree of correlation between the initial data and the classification task. Based on the cross-entropy loss and reconstruction error information of the initial data, the initial data is filtered to obtain denoised data. This denoised data is used to iteratively train the first classification network to obtain a second classification network that performs the aforementioned classification task. In other words, by combining the cross-entropy loss and reconstruction error information, the initial data is jointly filtered from both the uncertainty dimension and the representativeness dimension to obtain denoised data for network training, helping the network withstand interference from noisy data during training. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0022] Figure 1 This is a schematic diagram of an implementation environment provided by an exemplary embodiment of the present application;

[0023] Figure 2 is a flow chart of a data screening method provided by an exemplary embodiment of the present application;

[0024] Figure 3 is a schematic diagram of an autoencoder provided by an exemplary embodiment of the present application;

[0025] Figure 4 is a schematic diagram of a training framework provided by an exemplary embodiment of the present application;

[0026] Figure 5 is a flow chart of a data screening method provided by another exemplary embodiment of the present application;

[0027] Figure 6 is a flow chart of a data screening method provided by another exemplary embodiment of the present application;

[0028] Figure 7 is a schematic diagram of a network training framework provided by an exemplary embodiment of the present application;

[0029] Figure 8 is a block diagram of a data screening device provided by an exemplary embodiment of the present application;

[0030] Figure 9 is a block diagram of a data screening device provided by another exemplary embodiment of the present application;

[0031] Figure 10 It is a structural diagram of a server provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0033] First, a brief introduction to the terms involved in the embodiments of this application is given:

[0034] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0035] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and smart transportation.

[0036] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0037] Computer vision (CV) is the science of making machines "see." Specifically, it refers to machine vision techniques such as using cameras and computers to replace the human eye in identifying, tracking, and measuring objects. Further image processing is performed to transform the computer's image into an image more suitable for human observation or transmission to instrumentation. As a scientific discipline, computer vision studies related theories and technologies, aiming to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, optical character recognition (OCR), video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, autonomous driving, and smart transportation. It also includes common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0038] Natural language processing (NLP) is a key area of research in computer science and artificial intelligence. It studies the theories and methods that enable effective communication between humans and computers using natural language. Natural language processing (NLP) integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language we use in everyday life—and is closely linked to the study of linguistics. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robotic question answering, and knowledge graphs.

[0039] The data screening method provided in the embodiments of the present application involves machine learning technology in artificial intelligence and can be applied to computer vision scenarios and natural language processing scenarios. The above-mentioned data screening method is mainly used in the network training process for classification tasks, for example, image classification tasks in computer vision scenarios or text classification tasks in natural language processing scenarios. Schematically, the data screening method provided in the embodiments of the present application is applied to the following scenarios for illustration:

[0040] First, the data screening method described above is applied to image recognition. Taking the application scenario of medical image processing as an example, to achieve automatic recognition of lesion information in medical images through artificial intelligence, a recognition model for medical image recognition needs to be trained. This recognition model corresponds to a recognition task, such as melanoma recognition. During the recognition model training process, sample images for training are obtained and labeled according to "healthy / benign / malignant" to obtain training images. However, due to human experience, some training images may be incorrectly labeled during the labeling process. During the training process, an autoencoder is trained using the training images. The training images are input into the autoencoder, and the reconstruction error of the training images is obtained. A network to be trained is generated based on the melanoma recognition task. The training images are input into the network to be trained, and the cross-entropy loss of the training images relative to the network to be trained is obtained. The network parameters of the network to be trained are trained using this cross-entropy loss. Based on the reconstruction error and the cross-entropy loss, denoised images are determined from the training images for iterative training. The denoised image set contains fewer incorrectly labeled images than the total number of incorrectly labeled images in the training image set, ultimately resulting in a network model capable of melanoma recognition. The medical image to be identified is input into the above network model, which can judge the skin condition indicated by the medical image and output the skin condition result indicated by the image to be identified as one of a healthy state, benign melanoma or malignant melanoma.

[0041] Second, the aforementioned data screening method is applied to text recognition. For example, using the application scenario of automatic spam detection, AI is used to automatically identify received text messages and filter out spam. Specifically, a training model is established to classify text messages based on input text. A certain number of training samples are obtained. These training samples are labeled by human annotators as spam. For example, text containing terms like "discount," "consumption," and "lottery" are labeled as spam. These training samples are then fed into the training model and an autoencoder for data screening and model training. Ultimately, a spam recognition model with high recognition accuracy is obtained. This spam recognition model can be applied to scenarios such as filtering spam text messages and identifying spam on social media.

[0042] Third, the data screening method described above is applied to speech recognition. For example, in the speech-to-text application scenario, AI is used to convert speech information into text. During the speech-to-text process, the speech features identified in the speech data need to be classified according to pronunciation, which corresponds to the speech classification task. Similarly, during model training, manually annotated training samples are input into the model to be trained and the autoencoder for data screening and model training, ultimately resulting in a speech recognition model with high recognition accuracy. This speech recognition model can be applied to scenarios such as voice-to-text conversion in social software or voice input in input method functions.

[0043] The data screening method can also be applied to other application scenarios. Here, only the above three application scenarios are used for illustration, and the specific application scenarios are not limited.

[0044] Combined with the above explanations and application scenarios, the implementation environment of the embodiment of this application is explained. Figure 1 The implementation environment includes a terminal 101, a server 102 and a communication network 103.

[0045] Terminal 101 can be an electronic device such as a mobile phone, tablet computer, e-book reader, multimedia player, wearable device, laptop computer, desktop computer, or image / text / speech recognition all-in-one machine. Schematically, the labeler labels the samples used for model training through terminal 101 to obtain initial data, uploads the initial data to server 102, and server 102 uses the initial data to train the network to be trained. Terminal 101 is also used to input test data and upload the test data to server 102. The classification network recognizes the test data, and server 102 returns the recognition result to terminal 101, wherein the classification network is the network obtained by training the network to be trained after denoising the initial data.

[0046] The server 102 is used to provide a model training function to the terminal 101. During the iterative training of the training model, the noise data in the initial data is filtered so that the data used for iterative training each time contains less noise data, thereby reducing the impact of noise data on model training, improving the model accuracy, and finally obtaining a target model capable of processing classification tasks. The server 102 can transmit the target model to the terminal 101, or it can complete the recognition process of the target model on the data to be recognized in the server 102 by receiving the data to be recognized from the terminal 101, and only return the recognition result to the terminal 101.

[0047] It is worth noting that the above-mentioned server 102 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), as well as big data and artificial intelligence platforms.

[0048] Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to enable data computing, storage, processing, and sharing. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, and application technology, all based on the cloud computing business model. It can form a resource pool that can be used on demand with flexibility and convenience. Cloud computing technology will become a crucial support. Backend services for technical network systems, such as video websites, image websites, and more portals, require extensive computing and storage resources. With the rapid development and application of the internet industry, every item will likely have its own unique identification mark and will need to be transmitted to backend systems for logical processing. Data of varying levels will be processed separately, and data from all industries will require a strong system backend, which can only be achieved through cloud computing.

[0049] In some embodiments, server 102 can also be implemented as a node in a blockchain system. Blockchain is a novel application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Blockchain is essentially a decentralized database, a series of data blocks generated using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of the information (to prevent counterfeiting) and generate the next block. Blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0050] The underlying blockchain platform can include processing modules such as user management, basic services, smart contracts, and operation monitoring. The user management module is responsible for managing the identity information of all blockchain participants, including maintaining public and private key generation (account management), key management, and maintaining the correspondence between users' real identities and blockchain addresses (authority management). It also supervises and audits transactions involving certain real identities, providing risk control rule configuration (risk control auditing), with authorization. The basic service module is deployed on all blockchain node devices to verify the validity of business requests and record valid requests in storage after reaching consensus. For a new business request, the basic service first performs interface adaptation and authentication processing (interface adaptation), then encrypts the business information through a consensus algorithm (consensus management). After encryption, the encrypted information is transmitted completely and consistently to the shared ledger (network communication) and recorded and stored. The smart contract module is responsible for contract registration, issuance, contract triggering, and contract execution. Developers can define contract logic in a programming language and publish it to the blockchain (contract registration). According to the logic of the contract terms, the execution is triggered by calling keys or other events to complete the contract logic. It also provides the function of contract upgrade and cancellation. The operation monitoring module is mainly responsible for deployment, configuration modification, contract setting, cloud adaptation, and real-time status visualization during product release.

[0051] Illustratively, the terminal 101 and the server 102 are connected via a communication network 103 .

[0052] Please refer to Figure 2 , which shows a method for screening data shown in an embodiment of the present application. In the embodiment of the present application, the method is applied to Figure 1 In the server shown, schematically, the method can also be implemented in a terminal as a functional module. The embodiment of the present application is only described by taking implementation in a server as an example and does not limit its specific implementation environment. The method includes:

[0053] Step 201: Acquire initial data and a first classification network.

[0054] The first classification network is a network to be trained for processing classification tasks, and the initial data is data used to train network parameters of the first classification network.

[0055] The initial data is training data that has been labeled according to the classification task. For example, this training data can come from a database on the server, where labeled training data is stored. Alternatively, it can be training data uploaded by a terminal, where a human annotator annotates the training data through the terminal and uploads the labeled training data to the server as the initial data for the first classification network.

[0056] The initial data may be at least one of text data, image data, voice data, video data, and the like, and is not limited in the embodiments of the present application.

[0057] Illustratively, the network structure of the first classification network may be a network structure pre-stored in a database, or a network structure obtained from a terminal. In some embodiments, the first classification network may be at least one of a convolutional neural network (CNN), a VGG (Visual Geometry Group) network, a deep residual network (ResNet), or other neural networks capable of handling classification tasks.

[0058] The classification task corresponding to the first classification network can be a task stored in a database corresponding to the network structure of the first classification network, or can be a task instructed by the terminal. The classification task can be at least one of an image classification task, a text classification task, a speech classification task, a video classification task, and the like. The classification target indicated by the classification task corresponds to the initial data. For example, if the classification target indicated by the classification task is to classify an image, the initial data is image data.

[0059] The initial network parameters of the first classification network can be randomly initialized parameters or preset parameters, which are not limited here. The number of the first classification networks can be one or more.

[0060] Step 202: Determine the cross entropy loss corresponding to the initial data based on the prediction result obtained by the first classification network performing classification prediction on the initial data.

[0061] The cross-entropy loss of the initial data is used to measure the uncertainty of the initial data. After the initial data is input into the first classification network, it is classified and predicted by the first classification network, resulting in a prediction result. The parameters of the first classification network at this time are the initial network parameters. The cross-entropy loss corresponding to the initial data can be obtained from this prediction result.

[0062] In some embodiments, the cross entropy loss L CE It is obtained from Formula 1, where p represents the true value, p is a one-hot vector in Formula 1, and q represents the predicted value.

[0063] Formula 1:

[0064] Illustratively, after the cross entropy loss is determined based on the prediction result obtained by classifying the initial data based on the first classification network, the parameters of the first classification network are updated using the cross entropy loss to achieve the first training of the first classification network.

[0065] Step 203: perform feature sampling on the initial data to determine reconstruction error information corresponding to the initial data.

[0066] The reconstruction error information is used to indicate the correlation between the initial data and the classification task.

[0067] In some embodiments, feature sampling of the initial data is achieved through an autoencoder, and reconstruction error information corresponding to the initial data is determined through the output of the autoencoder. Illustratively, the autoencoder can be at least one of a linear autoencoder, a sparse autoencoder, a stacked autoencoder, a denoising autoencoder, and the like. The autoencoder can achieve compression and reconstruction of the initial data.

[0068] The autoencoder is trained by the initial data set, which includes all the initial data used for the first classification network training. Schematically, the training process of the autoencoder includes: obtaining the initial autoencoder; inputting the initial data into the initial autoencoder to obtain the encoding result; and performing supervised training on the initial autoencoder based on the reconstruction error between the initial data and the encoding result to obtain the autoencoder. In one example, the initial data is I, which is fed into the initial autoencoder M, and I' is predicted as the encoding result. The initial autoencoder M is supervised and trained by the reconstruction error, and finally the autoencoder is obtained. Among them, the reconstruction error L REC Calculated by formula 2.

[0069] Formula 2: L REC =|II′|

[0070] Based on the sampling results obtained by the trained autoencoder, the reconstruction error information corresponding to the initial data is determined. Figure 3 As shown, initial data I is input into the autoencoder 300, and a sampling result I' is output. The autoencoder 300 compresses and reconstructs the initial data I. Taking an image as an example, the initial input image undergoes a predetermined number of downsampling and upsampling processes through the autoencoder, and outputs a predicted image. The difference between the predicted image and the initial input image is used to determine the reconstruction error information of the initial input image.

[0071] Step 204 : Based on the cross entropy loss and the reconstruction error information, the initial data is screened to obtain denoised data. The denoised data is used to iteratively train the first classification network to obtain a second classification network for processing the classification task.

[0072] Methods for filtering initial data to obtain denoised data by using cross entropy loss and reconstruction error information include but are not limited to one of the following methods:

[0073] (1) Sort the initial data in the initial data set according to the cross entropy loss and the reconstruction error information respectively to screen the initial data. That is, sort the initial data in the initial data set by the cross entropy loss to obtain a cross entropy queue, and sort the initial data in the initial data set by the reconstruction error information to obtain a reconstruction error queue, wherein the cross entropy queue and the reconstruction error queue are both arranged in reverse order (i.e., the smaller the value, the higher the ranking). Then obtain a preset number of initial data from the cross entropy queue and the reconstruction error queue respectively, for example, obtain the initial data in the first 25% of the queue from the cross entropy queue, and obtain the initial data in the first 25% of the queue from the reconstruction error queue, and integrate the above initial data to obtain the denoising data for iterative training.

[0074] (2) The value information of the initial data is obtained by calculating the cross entropy loss and the reconstruction error information, and then the initial data in the initial data set are sorted according to the value information to screen the initial data. The value information is used to measure the training value of the initial data relative to the first classification network. In principle, the cross entropy loss and the reconstruction error information corresponding to the initial data are added according to the preset weight relationship to obtain the value information, and the initial data in the initial data set are sorted according to the value information, and the noise reduction data is screened from the sorted initial data according to the preset ratio. Taking the preset weight relationship as an example where the weights of the two are the same, the value information L is calculated by formula three. Joint , where L REC Represents the reconstruction error information, L CE represents the cross entropy loss.

[0075] Formula 3: L Joint =L REC +L CE

[0076] In the embodiment of the present application, taking the number of the first classification network as one as an example, the corresponding training framework is as follows: Figure 4As shown, the initial data 401 is respectively input into the autoencoder 410 and the first classification network 420, and the reconstruction error information corresponding to the initial data 401 is determined by the autoencoder 410, and the cross entropy loss corresponding to the initial data 401 is determined by the first classification network 420. The denoised data 1 is determined by the reconstruction error information and the cross entropy loss corresponding to the initial data 401, and the denoised data 1 is respectively input into the first classification network 420 and the autoencoder 410. At this time, the network parameters of the first classification network 420 have been updated by the cross entropy loss corresponding to the initial data 410. Then, in the same manner as above, the denoised data 1 is screened to obtain the denoised data 2, and the denoised data 2 is input into the first classification network 420. At the same time, the iterative training process of the first classification network and the screening process of the input network data are performed, and finally the second classification network capable of handling the classification task is trained.

[0077] In summary, the data screening method provided by the embodiment of the present application, in the training process of the neural network for the classification task, inputs the initial data into the network to be trained, obtains the cross entropy loss corresponding to the initial data based on the output prediction result, and simultaneously performs feature sampling on the initial data to determine the reconstruction error information that can indicate the correlation between the initial data and the classification task. The initial data is screened based on the cross entropy loss and reconstruction error information of the initial data to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain the second classification network that implements the above-mentioned classification task. That is, by combining the cross entropy loss and the reconstruction error information, the initial data can be jointly screened from the uncertainty dimension and the representative dimension to obtain denoised data for network training, helping the network to resist the interference of noise data during the training process.

[0078] Please refer to Figure 5 , which shows a method for screening data shown in an embodiment of the present application. In the embodiment of the present application, the screening process of the initial data is described, and the method includes:

[0079] Step 501: Acquire initial data and a first classification network.

[0080] The first classification network is a network to be trained for performing a classification task, and the initial data is data used to train the network parameters of the first classification network. The initial data is training data labeled according to the classification task. The initial data can be at least one of text data, image data, voice data, video data, and the like, and is not limited in this embodiment of the present application.

[0081] The initial network parameters of the first classification network can be randomly initialized parameters or preset parameters, which are not limited here. The number of first classification networks can be one or more. In some embodiments, the first classification network can be at least one of a convolutional neural network, a VGG network, a deep residual network, and other neural networks capable of handling classification tasks.

[0082] The classification task can be at least one of an image classification task, a text classification task, a speech classification task, a video classification task, etc., wherein the classification target indicated by the classification task has a corresponding relationship with the initial data. For example, if the classification target indicated by the classification task is to classify images, then the initial data is image data.

[0083] Step 502: Determine the cross entropy loss corresponding to the initial data based on the prediction result obtained by the first classification network performing classification prediction on the initial data.

[0084] The cross-entropy loss of the initial data is used to measure the uncertainty of the initial data. Schematically, the first classification network performs classification prediction on the initial data, outputs a prediction result corresponding to the current network parameters, and determines the cross-entropy loss of the initial data based on the prediction result. The cross-entropy loss is calculated using Formula 1.

[0085] After the cross entropy loss is determined, the first classification network updates its parameters according to the cross entropy loss to obtain a training sub-network.

[0086] Step 503 : Determine reconstruction error information corresponding to the initial data based on a sampling result obtained by performing feature sampling on the initial data by the autoencoder.

[0087] The reconstruction error information is used to indicate the degree of correlation between the initial data and the classification task. The autoencoder can be at least one of a linear autoencoder, a sparse autoencoder, a stacked autoencoder, a denoising autoencoder, and the like. The autoencoder is capable of compressing and reconstructing the initial data. The autoencoder is trained using an initial dataset, which includes all initial data used for training the first classification network.

[0088] Initial data is input into the autoencoder for feature sampling, and the autoencoder outputs the sampling result. The reconstruction error information is determined by the difference between the sampling result and the initial data. In one example, the reconstruction error information is the reconstruction error obtained by Formula 2.

[0089] Step 504: Determine the value information corresponding to the initial data based on the cross entropy loss and the reconstruction error information.

[0090] The value information is used to measure the training value of the initial data relative to the first classification network. Schematically, the cross entropy loss and reconstruction error information corresponding to the initial data are added according to a preset weight relationship to obtain the value information. The preset weight relationship can be stored in the database corresponding to the first classification network, or can be input by the terminal. Among them, the cross entropy loss focuses on the uncertainty dimension to reduce the noise of the initial data, while the reconstruction error information focuses on the representative dimension to reduce the noise of the initial data. In some embodiments, the classification task includes information indicating the above-mentioned preset weight relationship, and the preset weight relationship can also be manually set according to actual emphasis requirements.

[0091] In one example, the weights of the two are the same, and the value information is calculated using Formula 3, wherein the larger the numerical value corresponding to the value information is, the lower the training value of the initial data relative to the first classification network.

[0092] Step 505: sort the initial data according to the value information.

[0093] Arrange the initial data in reverse order according to the value information to obtain the corresponding data queue.

[0094] Step 506 : Filter the sorted initial data according to a preset ratio to obtain noise reduction data.

[0095] In some embodiments, the classification task corresponds to an indication of a preset ratio. For example, if the preset ratio is 25% of the total data, the first 25% of the initial data is obtained from the data queue as the noise reduction data. In an exemplary embodiment, the preset ratio can also be determined based on the proportion of noise data in the initial data set, wherein the proportion of noise data in the initial data set is negatively correlated with the preset ratio. That is, the more noise data in the initial data, the lower the preset ratio is set, and the more noise data is filtered out each time.

[0096] The denoised data obtained after screening is used to iteratively train the first classification network. Illustratively, the denoised data is input into the first classification network, and steps 502 to 506 are repeated. That is, the first classification network is trained using the denoised data. The denoised data is then further screened based on the reconstruction error information and cross-entropy loss corresponding to the denoised data to obtain denoised data with a lower noise ratio. The above steps are repeated continuously until the first classification network is trained to converge, resulting in a second classification network capable of handling the classification task.

[0097] In summary, the data screening method provided by the embodiment of the present application, in the training process of the neural network for the classification task, inputs the initial data into the network to be trained, obtains the cross entropy loss corresponding to the initial data based on the output prediction result, and simultaneously performs feature sampling on the initial data to determine the reconstruction error information that can indicate the correlation between the initial data and the classification task. The initial data is screened based on the cross entropy loss and reconstruction error information of the initial data to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain the second classification network that implements the above-mentioned classification task. That is, by combining the cross entropy loss and the reconstruction error information, the initial data can be jointly screened from the uncertainty dimension and the representative dimension to obtain denoised data for network training, helping the network to resist the interference of noise data during the training process.

[0098] Please refer to Figure 6 , which shows a data screening method shown in an embodiment of the present application. In the embodiment of the present application, when the number of the first classification networks is 2, the above data screening method is used as an example to improve the generalization of the network. The method includes:

[0099] Step 601: Acquire initial data, a first classification sub-network, and a second classification sub-network.

[0100] In an embodiment of the present application, the first classification network includes a first classification subnetwork and a second classification subnetwork. The first classification subnetwork and the second classification subnetwork have different initialization parameters, and the first classification subnetwork and the second classification subnetwork are targeted at the same classification task.

[0101] Step 602: Determine a first cross entropy loss corresponding to the initial data based on a first prediction result obtained by performing classification prediction on the initial data by the first classification subnetwork.

[0102] After the initial data is input into the first classification sub-network, the first classification sub-network performs classification prediction to obtain a first prediction result, and a first cross entropy loss is obtained based on the cross entropy between the first prediction result and the input initial data.

[0103] Step 603: Determine a second cross entropy loss corresponding to the initial data based on a second prediction result obtained by performing classification prediction on the initial data by the second classification subnetwork.

[0104] After the initial data is input into the second classification sub-network, the second classification sub-network performs classification prediction to obtain a second prediction result. The second cross entropy loss is obtained based on the cross entropy between the second prediction result and the input initial data.

[0105] In some embodiments, the cross entropy loss is determined based on the first cross entropy loss and the second cross entropy loss. That is, the cross entropy loss corresponding to the initial data is determined by jointly determining the first cross entropy loss and the second cross entropy loss. For example, the cross entropy loss is obtained by weighted summing the first cross entropy loss and the second cross entropy loss of the initial data according to preset weights.

[0106] Step 604: perform feature sampling on the initial data to determine reconstruction error information corresponding to the initial data.

[0107] The reconstruction error information is used to indicate the degree of association between the initial data and the classification task. In the embodiment of the present application, feature sampling of the initial data is achieved through an autoencoder, and the reconstruction error information corresponding to the initial data is determined based on the output of the autoencoder. The autoencoder is trained using all the initial data, and the reconstruction error information corresponding to the initial data is determined based on the sampling results obtained by sampling the features of the initial data using the trained autoencoder.

[0108] Step 605 : Combine the first cross entropy loss and the reconstruction error information to obtain first value information of the initial data.

[0109] The first value information is used to measure the training value of the initial data relative to the first classification sub-network. Schematically, the first cross entropy loss and the reconstruction error information are weighted and added according to a preset weight relationship to obtain the first value information.

[0110] Step 606: Combine the second cross entropy loss and the reconstruction error information to obtain second value information of the initial data.

[0111] The second value information is used to measure the training value of the initial data relative to the second classification sub-network. Schematically, the second cross entropy loss and the reconstruction error information are weighted and added according to a preset weight relationship to obtain the second value information.

[0112] Illustratively, the initial data is sorted based on the first value information and the second value information. In some embodiments, the initial data can be sorted based on both the first value information and the second value information, or can be sorted based on each of the first value information and the second value information. In this embodiment of the present application, the sorting and filtering of the initial data is used as an example for illustration.

[0113] Step 607: sort the initial data based on the first value information to obtain a first data queue.

[0114] Indicatively, the first value information is sorted in reverse order, that is, the smaller the numerical value corresponding to the first value information is, the higher the training value of the initial data relative to the first classification sub-network is, and the higher the corresponding ranking is.

[0115] Step 608: Sort the initial data based on the second value information to obtain a second data queue.

[0116] Indicatively, the second value information is sorted in reverse order, that is, the smaller the numerical value corresponding to the second value information, the higher the training value of the initial data relative to the second classification sub-network, and the higher the corresponding ranking.

[0117] Step 609: Acquire first noise reduction data from the first data queue according to a first preset ratio.

[0118] There is a negative correlation between the proportion of noise data in the initial data set and the preset ratio, that is, the more noise data there is in the initial data and the lower the preset ratio is set, the more noise data will be filtered out each time.

[0119] Step 610: Obtain second noise reduction data from the second data queue according to a second preset ratio.

[0120] In some embodiments, the first preset ratio for screening the first data queue and the second preset ratio for screening the second data queue may be the same or different. Illustratively, when the first preset ratio and the second preset ratio are different, the first preset ratio and the second preset ratio are determined based on the weights of the first classification subnetwork and the second classification subnetwork in the classification task. For example, when the classification task indicates that the weight of the output result of the first network when performing the classification task is higher than the weight of the output result of the second network when performing the classification task, then the first preset ratio is higher than the second preset ratio. The first network is a network obtained by training the first classification subnetwork, and the second network is a network obtained by training the second classification subnetwork.

[0121] Illustratively, denoised data is generated based on the first denoised data and the second denoised data. The first denoised data and the second denoised data can be used together to iteratively train the first classification sub-network and the second classification sub-network, i.e., the first denoised data and the second denoised data are input to the first classification sub-network for iterative training, and the first denoised data and the second denoised data are simultaneously input to the second classification sub-network for iterative training; or they can be cross-inputted to the first classification sub-network and the second classification sub-network for iterative training, i.e., the first denoised data is input to the second classification sub-network for iterative training, and the second denoised data is input to the first classification sub-network for iterative training.

[0122] Step 611: Determine a third cross entropy loss corresponding to the first denoised data based on a prediction result obtained by performing classification prediction on the first denoised data by the second update network.

[0123] The second updated network is a network obtained after the second classification sub-network is parameter updated according to the first cross entropy loss. In the embodiment of the present application, the first denoised data is input into the second updated network to determine the third cross entropy loss corresponding to the first denoised data.

[0124] Step 612: Determine a fourth cross entropy loss corresponding to the second denoised data based on a prediction result obtained by performing classification prediction on the second denoised data by the first update network.

[0125] The first updated network is a network obtained by performing parameter update on the first classification subnetwork according to the second cross entropy loss. In an embodiment of the present application, the second denoised data is input into the first updated network to determine the fourth cross entropy loss corresponding to the first denoised data.

[0126] In an embodiment of the present application, the first denoised data and the second denoised data are cross-input into the first classification sub-network and the second classification sub-network for subsequent network parameter training, that is, the first denoised data is used to train the network parameters of the second update network, and the second denoised data is used to train the network parameters of the first update network.

[0127] Step 613: Perform feature sampling on the first denoised data to determine first reconstruction error information corresponding to the first denoised data.

[0128] The first denoised data is input into the above-mentioned autoencoder, and a sampling result corresponding to the first denoised data is output. Based on the relationship between the sampling result and the input first denoised data, first reconstruction error information corresponding to the first denoised data is determined, wherein the first reconstruction error information is calculated using Formula 2.

[0129] Step 614: perform feature sampling on the second denoised data to determine second reconstruction error information corresponding to the second denoised data.

[0130] The second denoised data is input into the above-mentioned autoencoder, and a sampling result corresponding to the second denoised data is output. Based on the relationship between the sampling result and the input second denoised data, the second reconstruction error information corresponding to the second denoised data is determined, wherein the second reconstruction error information is calculated using Formula 2.

[0131] Step 615 : Filter the first denoised data based on the third cross entropy loss and the first reconstruction error information to obtain third denoised data.

[0132] The third denoised data is used to iteratively train the first updated network. In the embodiment of the present application, the method of filtering the third denoised data from the first denoised data is the same as the method of filtering the first denoised data from the initial data, which will not be described in detail here.

[0133] Step 616 : Filter the second denoised data based on the fourth cross entropy loss and the second reconstruction error information to obtain fourth denoised data.

[0134] The fourth denoised data is used to iteratively train the second updated network. In the embodiment of the present application, the method of filtering the fourth denoised data from the second denoised data is the same as the method of filtering the second denoised data from the initial data, which will not be described in detail here.

[0135] Indicative, such as Figure 7 As shown, it shows the network training framework provided by an embodiment of the present application, the initial data 701 is respectively input into the autoencoder 710, the first classification subnetwork 720 and the second classification subnetwork 730, and the first denoised data is determined by the reconstruction error information determined by the autoencoder 710 and the first cross entropy loss determined by the first classification subnetwork 720, and the first denoised data is input into the second update network 731, wherein the second update network 731 is the network obtained after the second classification subnetwork 730 performs parameter update according to the second cross entropy loss; the second denoised data is determined by the reconstruction error information determined by the autoencoder 710 and the second cross entropy loss determined by the second classification subnetwork 730, and the second denoised data is input into the first update network 721, wherein the first update network 721 is the network obtained after the first classification subnetwork 720 performs parameter update according to the first cross entropy loss, and then the aforementioned data screening and network training process are repeated until the two networks converge to obtain the first network 722 and the second network 732 for processing the classification task.

[0136] In an embodiment of the present application, a first network and a second network can be finally trained according to the provided data screening method, wherein the first network and the second network process the same classification task, and the network parameters of the first network and the second network may be the same or different. When the parameters of the two networks are different, the target network for a specific application can be determined based on the classification situation focused on by the network parameters. For example, taking the image classification task as an example, the trained first network is more accurate in classifying "cats", while the second network is more accurate in classifying "dogs". Then, when the target network application is to recognize cats, the first network is selected as the target network.

[0137] Optionally, the first network and the second network can be used together as the target network, that is, the recognition results of the first network are combined with the recognition results of the second network to obtain a target recognition result. For example, the data to be recognized is input into the first network to obtain a first result, and the data to be recognized is input into the second network to obtain a second result. The first and second results are combined with different weights to obtain the target result. The specific application of the first and second networks is not limited here.

[0138] In summary, the data screening method provided in the embodiment of the present application, during the training process of the neural network for the classification task, the initial data is respectively input into the first classification subnetwork and the second classification subnetwork, and the first cross entropy loss and the second cross entropy loss corresponding to the initial data are respectively obtained according to the output first prediction result and the second prediction result. At the same time, feature sampling is performed on the initial data to determine the reconstruction error information that can indicate the correlation between the initial data and the classification task. The initial data is screened based on the first cross entropy loss and the reconstruction error information of the initial data to obtain the first denoised data, and the initial data is screened based on the second cross entropy loss and the reconstruction error information of the initial data to obtain the second denoised data. The first denoised data is used to train the network parameters of the second classification subnetwork, and the second denoised data is used to train the network parameters of the first classification subnetwork. At the same time, the third denoised data is screened from the first denoised data, and the fourth denoised data is screened from the second denoised data. The two networks to be trained are cross-trained by the third denoised data and the fourth denoised data. The above data screening and network parameter cross-training process are repeated until the two networks converge to obtain two networks that can handle the classification task. That is, by combining cross entropy loss and reconstruction error information, the initial data can be jointly screened from the uncertainty dimension and the representative dimension to obtain denoised data for network training, helping the network to resist the interference of noisy data during training, and at the same time improving the generalization of the network through cross training.

[0139] The following are device embodiments of the present application, which can be used to implement the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0140] Figure 8 A block diagram of a data screening device provided by one embodiment of the present application is shown. The device has the function of implementing the above-mentioned method example, and the function can be implemented by hardware or by hardware executing corresponding software. The device may include:

[0141] An acquisition module 810 is configured to acquire initial data and a first classification network, where the first classification network is a network to be trained for processing a classification task;

[0142] A determination module 820 is configured to determine a cross entropy loss corresponding to the initial data based on a prediction result obtained by performing classification prediction on the initial data by the first classification network;

[0143] The determining module 820 is further configured to perform feature sampling on the initial data to determine reconstruction error information corresponding to the initial data, where the reconstruction error information is used to indicate a degree of association between the initial data and the classification task;

[0144] The screening module 830 is used to screen the initial data based on the cross entropy loss and the reconstruction error information to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain a second classification network for processing the classification task.

[0145] In some optional embodiments, such as Figure 9 As shown, the screening module 830 further includes:

[0146] a determining unit 831 configured to determine value information corresponding to the initial data based on the cross entropy loss and the reconstruction error information, wherein the value information is used to measure the training value of the initial data relative to the first classification network;

[0147] a sorting unit 832, configured to sort the initial data according to the value information;

[0148] The screening unit 833 is configured to screen the sorted initial data according to a preset ratio to obtain the noise reduction data.

[0149] In some optional embodiments, the first classification network includes a first classification subnetwork and a second classification subnetwork, and the first classification subnetwork and the second classification subnetwork have different initialization parameters;

[0150] The determining module 820 is further configured to determine a first cross entropy loss corresponding to the initial data based on a first prediction result obtained by performing classification prediction on the initial data by the first classification subnetwork;

[0151] The determining module 820 is further configured to determine a second cross entropy loss corresponding to the initial data based on a second prediction result obtained by performing classification prediction on the initial data by the second classification subnetwork;

[0152] The determination module 820 is further configured to determine the cross entropy loss based on the first cross entropy loss and the second cross entropy loss.

[0153] In some optional embodiments, the determining unit 831 is further configured to combine the first cross entropy loss and the reconstruction error information to obtain first value information of the initial data, where the first value information is used to measure the training value of the initial data relative to the first classification subnetwork;

[0154] The determining unit 831 is further configured to combine the second cross entropy loss and the reconstruction error information to obtain second value information of the initial data, where the second value information is used to measure the training value of the initial data relative to the second classification subnetwork;

[0155] The sorting unit 832 is further configured to sort the initial data based on the first value information and the second value information.

[0156] In some optional embodiments, the determining unit 831 is further configured to sort the initial data based on the first value information to obtain a first data queue;

[0157] The determining unit 831 is further configured to sort the initial data based on the second value information to obtain a second data queue;

[0158] The screening unit 833 is further configured to obtain first noise reduction data from the first data queue according to a first preset ratio;

[0159] The screening unit 833 is further configured to obtain second noise reduction data from the second data queue according to a second preset ratio;

[0160] The screening unit 833 is further configured to generate the noise reduction data according to the first noise reduction data and the second noise reduction data.

[0161] In some optional embodiments, the determination module 820 is further configured to determine a third cross entropy loss corresponding to the first denoised data based on a prediction result obtained by performing classification prediction on the first denoised data by a second update network, where the second update network is a network obtained by performing parameter update on the second classification subnetwork according to the first cross entropy loss;

[0162] The determining module 820 is further configured to determine a fourth cross entropy loss corresponding to the second denoised data based on a prediction result obtained by performing classification prediction on the second denoised data by a first updating network, where the first updating network is a network obtained by performing parameter update on the first classification subnetwork according to the second cross entropy loss;

[0163] The determining module 820 is further configured to perform feature sampling on the first denoised data to determine first reconstruction error information corresponding to the first denoised data;

[0164] The determining module 820 is further configured to perform feature sampling on the second denoised data to determine second reconstruction error information corresponding to the second denoised data;

[0165] The screening module 830 is further configured to screen the first denoised data based on the third cross entropy loss and the first reconstruction error information to obtain third denoised data, where the third denoised data is used to iteratively train the first update network.

[0166] The screening module 830 is further configured to screen the second denoised data based on the fourth cross entropy loss and the second reconstruction error information to obtain fourth denoised data, and the fourth denoised data is used to iteratively train the second update network.

[0167] In some optional embodiments, the acquisition module 810 is further configured to acquire an initial autoencoder;

[0168] The device further comprises:

[0169] A training module 840 is configured to input the initial data into the initial autoencoder to obtain an encoding result;

[0170] The training module 840 is further configured to perform supervised training on the initial autoencoder according to a reconstruction error between the initial data and the encoding result to obtain an autoencoder;

[0171] The determination module 820 is further configured to determine the reconstruction error information corresponding to the initial data based on a sampling result obtained by the autoencoder performing feature sampling on the initial data.

[0172] In summary, the data screening device provided by the embodiment of the present application, during the training process of the neural network for the classification task, inputs the initial data into the network to be trained, obtains the cross entropy loss corresponding to the initial data based on the output prediction result, and simultaneously performs feature sampling on the initial data to determine the reconstruction error information that can indicate the correlation between the initial data and the classification task. The initial data is screened based on the cross entropy loss and reconstruction error information of the initial data to obtain denoised data, which is used to iteratively train the first classification network to obtain a second classification network that implements the above-mentioned classification task. That is, by combining the cross entropy loss and the reconstruction error information, the initial data can be jointly screened from the uncertainty dimension and the representative dimension to obtain denoised data for network training, helping the network to resist the interference of noise data during the training process.

[0173] It should be noted that the data screening device provided in the above embodiment is merely an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the data screening device provided in the above embodiment and the data screening method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0174] Figure 10 The following is a schematic diagram of the structure of a server provided by an exemplary embodiment of the present application. Specifically, it includes the following structure.

[0175] The server 1000 includes a central processing unit (CPU) 1001, a system memory 1004 including a random access memory (RAM) 1002 and a read-only memory (ROM) 1003, and a system bus 1005 connecting the system memory 1004 and the CPU 1001. The server 1000 also includes a mass storage device 1006 for storing an operating system 1013, application programs 1014, and other program modules 1015.

[0176] The mass storage device 1006 is connected to the central processing unit 1001 via a mass storage controller (not shown) connected to the system bus 1005. The mass storage device 1006 and its associated computer-readable media provide non-volatile storage for the server 1000. That is, the mass storage device 1006 may include a computer-readable medium (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0177] Without loss of generality, computer-readable media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other solid-state storage technologies, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, tape cassettes, magnetic tape, disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that computer storage media are not limited to the above-mentioned ones. The above-mentioned system memory 1104 and mass storage device 1006 can be collectively referred to as memory.

[0178] According to various embodiments of the present application, the server 1000 may also be connected to a remote computer on a network such as the Internet for operation. That is, the server 1000 may be connected to the network 1012 via the network interface unit 1011 connected to the system bus 1005, or the network interface unit 1011 may be used to connect to other types of networks or remote computer systems (not shown).

[0179] The memory also includes one or more programs, which are stored in the memory and configured to be executed by the CPU.

[0180] Embodiments of the present application also provide a computer device comprising a processor and memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the target object prediction method provided in each of the above method embodiments. Optionally, the computer device may be a terminal or a server.

[0181] An embodiment of the present application also provides a computer-readable storage medium, which stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement the target object prediction method provided by the above-mentioned method embodiments.

[0182] Embodiments of the present application further provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the target object prediction method described in any of the above embodiments.

[0183] Optionally, the computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a solid-state drive (SSD), or an optical disk. Among them, the random access memory may include a resistance random access memory (ReRAM) and a dynamic random access memory (DRAM). The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0184] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0185] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A data screening method, characterized in that: The method comprises: Acquire initial data and a first classification network, where the first classification network is a network to be trained for processing a classification task; Determining a cross entropy loss corresponding to the initial data based on a prediction result obtained by performing classification prediction on the initial data by the first classification network; Performing feature sampling on the initial data to determine reconstruction error information corresponding to the initial data, where the reconstruction error information is used to indicate a degree of association between the initial data and the classification task; Based on the cross entropy loss and the reconstruction error information, the initial data is screened to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain a second classification network for processing the classification task.

2. The method according to claim 1, characterized in that The screening of the initial data based on the cross entropy loss and the reconstruction error information to obtain denoised data includes: Determining value information corresponding to the initial data based on the cross entropy loss and the reconstruction error information, where the value information is used to measure the training value of the initial data relative to the first classification network; sorting the initial data according to the value information; The noise reduction data is obtained by screening the sorted initial data according to a preset ratio.

3. The method according to claim 2, characterized in that The first classification network includes a first classification subnetwork and a second classification subnetwork, and the first classification subnetwork and the second classification subnetwork have different initialization parameters; Determining the cross entropy loss corresponding to the initial data based on the prediction result obtained by performing classification prediction on the initial data by the first classification network includes: Determining a first cross entropy loss corresponding to the initial data based on a first prediction result obtained by performing classification prediction on the initial data by the first classification subnetwork; Determining a second cross entropy loss corresponding to the initial data based on a second prediction result obtained by performing classification prediction on the initial data by the second classification subnetwork; The cross entropy loss is determined based on the first cross entropy loss and the second cross entropy loss.

4. The method according to claim 3, characterized in that The determining, based on the cross entropy loss and the reconstruction error information, the value information corresponding to the initial data includes: Combining the first cross entropy loss and the reconstruction error information to obtain first value information of the initial data, where the first value information is used to measure the training value of the initial data relative to the first classification subnetwork; Combining the second cross entropy loss and the reconstruction error information to obtain second value information of the initial data, where the second value information is used to measure the training value of the initial data relative to the second classification subnetwork; The sorting of the initial data according to the value information includes: The initial data is sorted based on the first value information and the second value information.

5. The method according to claim 4, characterized in that The sorting of the initial data based on the first value information and the second value information includes: sorting the initial data based on the first value information to obtain a first data queue; sorting the initial data based on the second value information to obtain a second data queue; The step of screening the sorted initial data according to a preset ratio to obtain the noise reduction data includes: Acquiring first noise reduction data from the first data queue according to a first preset ratio; Acquiring second noise reduction data from the second data queue according to a second preset ratio; The noise reduction data is generated based on the first noise reduction data and the second noise reduction data.

6. The method according to claim 5, characterized in that After the initial data is screened based on the cross entropy loss and the reconstruction error information to obtain the denoised data, the method further includes: Determining a third cross entropy loss corresponding to the first denoised data based on a prediction result obtained by performing classification prediction on the first denoised data by a second update network, where the second update network is a network obtained by performing parameter update on the second classification subnetwork according to the first cross entropy loss; Determining a fourth cross entropy loss corresponding to the second denoised data based on a prediction result obtained by performing classification prediction on the second denoised data by a first update network, where the first update network is a network obtained by performing parameter update on the first classification subnetwork according to the second cross entropy loss; performing feature sampling on the first denoised data to determine first reconstruction error information corresponding to the first denoised data; performing feature sampling on the second denoised data to determine second reconstruction error information corresponding to the second denoised data; Based on the third cross entropy loss and the first reconstruction error information, the first denoised data is screened to obtain third denoised data, where the third denoised data is used for iteratively training the first update network; Based on the fourth cross entropy loss and the second reconstruction error information, the second denoised data is screened to obtain fourth denoised data, and the fourth denoised data is used for iterative training of the second update network.

7. The method according to any one of claims 1 to 6, characterized in that: The performing feature sampling on the initial data to determine reconstruction error information corresponding to the initial data includes: Get the initial autoencoder; Inputting the initial data into the initial autoencoder to obtain an encoding result; Performing supervised training on the initial autoencoder according to a reconstruction error between the initial data and the encoding result to obtain an autoencoder; The reconstruction error information corresponding to the initial data is determined based on a sampling result obtained by the autoencoder performing feature sampling on the initial data.

8. A data screening device, characterized in that: The device comprises: An acquisition module, configured to acquire initial data and a first classification network, where the first classification network is a network to be trained for processing a classification task; a determination module, configured to determine a cross entropy loss corresponding to the initial data based on a prediction result obtained by performing classification prediction on the initial data by the first classification network; The determination module is further configured to perform feature sampling on the initial data to determine reconstruction error information corresponding to the initial data, where the reconstruction error information is used to indicate a degree of association between the initial data and the classification task; A screening module is used to screen the initial data based on the cross entropy loss and the reconstruction error information to obtain denoised data, and the denoised data is used to iteratively train the first classification network to obtain a second classification network for processing the classification task.

9. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the data screening method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one program code, and the program code is loaded and executed by a processor to implement the data screening method according to any one of claims 1 to 7.

Images