Data processing method and device, equipment and storage medium

Abstract

The application discloses a data processing method and device, equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: obtaining a noise measurement index of a first candidate task based on an initial neural network model; obtaining a sampling probability of the first candidate task based on the noise measurement index of the first candidate task; selecting a reference task meeting a first selection condition from the first candidate task according to the sampling probability; updating the parameters of the initial neural network model based on the reference task; learning a target task based on the neural network model after the parameter update, obtaining a target neural network model, and using the target neural network model for data processing. The sampling probability obtained by considering the noise measurement index is a sampling probability matched with the interference degree of the candidate task on the parameter update of the model. The reliability of the reference task selected from the candidate task based on the sampling probability is high, which is beneficial to improving the update quality of the parameters of the model, thereby improving the accuracy of data processing.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of artificial intelligence, and particularly relate to a data processing method and device, equipment and a storage medium. BACKGROUND

[0002] With the development of artificial intelligence technology, there are more and more methods for data processing. For example, a model is first obtained based on a meta-learning model training process, and then the model is used for data processing. Meta-learning is a common method for solving small sample problems. By using the meta-learning model training process, a model with good performance can be trained based on less data.

[0003] In related technologies, the meta-learning model training process includes: selecting a task from candidate tasks by a uniform sampling method, updating the parameters of a neural network model based on the selected task, learning a target task based on the model after the parameter update, and obtaining a target neural network model.

[0004] The uniform sampling method gives different candidate tasks the same sampling probability, which is difficult to ensure the reliability of the selected task, and is likely to result in poor quality of the update of the parameters of the model. The model after the parameter update is difficult to accurately learn the target task, and the quality of the obtained target neural network model is poor. It is difficult to accurately process data matching the target task by using the target neural network model. SUMMARY

[0005] Embodiments of the present application provide a data processing method, device, equipment and storage medium, which can be used to improve the accuracy of data processing. The technical solution is as follows:

[0006] In one aspect, the present application provides a data processing method, which comprises:

[0007] obtaining a noise measurement index of a first candidate task based on an initial neural network model, the noise measurement index being used to measure the degree of interference of the first candidate task on updating the parameters of the initial neural network model, the first candidate task being a task matching a target task to be learned;

[0008] obtaining a sampling probability of the first candidate task based on the noise measurement index of the first candidate task; selecting a reference task from the first candidate task, the sampling probability of the reference task satisfying a first selection condition;

[0009] updating the parameters of the initial neural network model based on the reference task; learning the target task based on the neural network model after the parameter update to obtain a target neural network model, the target neural network model being used for data processing.

[0010] Also provided is a data processing method, comprising:

[0011] obtaining to-be-processed data matched with a target task and a target neural network model, the target neural network model being obtained by any of the above data processing methods;

[0012] calling the target neural network model to process the to-be-processed data, to obtain a processing result corresponding to the to-be-processed data.

[0013] In another aspect, a data processing apparatus is provided, comprising:

[0014] a first obtaining unit configured to obtain a noise measurement index of a first candidate task based on an initial neural network model, the noise measurement index being used to measure an interference degree of the first candidate task on updating parameters of the initial neural network model, the first candidate task being a task matched with a target task to be learned;

[0015] a second obtaining unit configured to obtain a sampling probability of the first candidate task based on the noise measurement index of the first candidate task, and select a reference task from the first candidate task, the sampling probability of the reference task satisfying a first selection condition;

[0016] an updating unit configured to update the parameters of the initial neural network model based on the reference task;

[0017] a third obtaining unit configured to learn the target task based on the neural network model after the parameters are updated, to obtain a target neural network model, the target neural network model being used for data processing.

[0018] In a possible implementation, the first candidate task corresponds to training data and test data; the first obtaining unit is configured to obtain a task processing model corresponding to the first candidate task based on the training data corresponding to the first candidate task and the initial neural network model, test the task processing model corresponding to the first candidate task by using the test data corresponding to the first candidate task, obtain a test performance of the task processing model corresponding to the first candidate task, obtain a first update gradient corresponding to the parameters of the initial neural network model based on the training data corresponding to the first candidate task, obtain a second update gradient corresponding to the parameters of the initial neural network model based on the test data corresponding to the first candidate task, and obtain the noise measurement index of the first candidate task based on a similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task.

[0019] In a possible implementation, the first candidate task corresponds to training data and test data; the first obtaining unit is configured to obtain, based on the training data corresponding to the first candidate task and the initial neural network model, a task processing model corresponding to the first candidate task; test the task processing model corresponding to the first candidate task by using test data corresponding to the first candidate task to obtain a test performance of the task processing model corresponding to the first candidate task; and obtain, based on the test performance of the task processing model corresponding to the first candidate task, a second-order update gradient corresponding to the parameters of the initial neural network model, and use the second-order update gradient as a noise measurement index of the first candidate task.

[0020] In a possible implementation, the second obtaining unit is configured to invoke the target scheduling model to obtain, based on the noise measurement index of the first candidate task, a sampling probability of the first candidate task.

[0021] In a possible implementation, the second obtaining unit is configured to invoke the target scheduling model to encode a current parameter update round to obtain first encoding information; encode the noise measurement index of the first candidate task to obtain second encoding information; splice the first encoding information and the second encoding information to obtain target encoding information; and predict the sampling probability of the first candidate task based on the target encoding information.

[0022] In a possible implementation, the second obtaining unit is further configured to invoke an initial scheduling model to obtain, based on a noise measurement index of a second candidate task, an initial sampling probability of the second candidate task; and select, from the second candidate task, an initial task whose initial sampling probability satisfies a second selection condition.

[0023] The updating unit is further configured to obtain, based on the initial task and the initial neural network model, a to-be-verified model; obtain a verification performance of the to-be-verified model based on a verification task; obtain, based on the verification performance, a third update gradient corresponding to parameters of the initial scheduling model; and update the parameters of the initial scheduling model based on the third update gradient to obtain the target scheduling model.

[0024] In a possible implementation, the verification task corresponds to training data and test data; and the updating unit is further configured to train the to-be-verified model based on the training data corresponding to the verification task to obtain a task processing model corresponding to the verification task; test the task processing model corresponding to the verification task by using test data corresponding to the verification task to obtain a test performance of the task processing model corresponding to the verification task; and obtain the verification performance of the to-be-verified model based on the test performance of the task processing model corresponding to the verification task.

[0025] In a possible implementation, the second obtaining unit is further configured to obtain a sample noise measurement index of a sample task and a standard sampling probability of the sample task; and call an initial scheduling model to obtain an initial sampling probability of the sample task based on the sample noise measurement index.

[0026] The updating unit is further configured to obtain a target loss based on the initial sampling probability and the standard sampling probability of the sample task; obtain a fourth update gradient corresponding to a parameter of the initial scheduling model based on the target loss; and update the parameter of the initial scheduling model based on the fourth update gradient to obtain the target scheduling model.

[0027] In a possible implementation, the second obtaining unit is configured to map the noise measurement index of the first candidate task to a sampling probability of the first candidate task based on a mapping relationship between a noise measurement index and a sampling probability.

[0028] In a possible implementation, the number of reference tasks is at least one, and any reference task corresponds to training data and test data; and the updating unit is configured to obtain a task processing model corresponding to the any reference task based on the training data corresponding to the any reference task and the initial neural network model; test the task processing model corresponding to the any reference task by using the test data corresponding to the any reference task to obtain a test performance of the task processing model corresponding to the any reference task; obtain a target update gradient corresponding to a parameter of the initial neural network model based on the test performances of the task processing models corresponding to the at least one reference task; and update the parameter of the initial neural network model based on the target update gradient.

[0029] In a possible implementation, the training data corresponding to the first candidate task includes training sample data and a standard processing result corresponding to the training sample data; the first obtaining unit is configured to call a reference model to process the training sample data to obtain a predicted processing result corresponding to the training sample data, a parameter of the reference model being the same as a parameter of the initial neural network model; obtain a first loss based on the predicted processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data; and update the parameter of the reference model based on the first loss to obtain a task processing model corresponding to the first candidate task.

[0030] In a possible implementation, the first candidate task is a task of predicting an activity of a compound on a first candidate target protein, the training sample data is representation information of a first compound, and the prediction processing result is a predicted activity of the first compound on the first candidate target protein; the first obtaining unit is configured to invoke the reference model to process the representation information of the first compound to obtain the predicted activity of the first compound on the first candidate target protein; and the first loss is obtained based on the predicted activity of the first compound on the first candidate target protein and a standard activity of the first compound on the first candidate target protein.

[0031] The data processing apparatus also includes:

[0032] The fourth obtaining unit is configured to obtain to-be-processed data matched with a target task and a target neural network model, the target neural network model being obtained by the data processing method.

[0033] The processing unit is configured to invoke the target neural network model to process the to-be-processed data to obtain a processing result corresponding to the to-be-processed data.

[0034] In a possible implementation, the target task is a task of predicting an activity of a compound on a target target protein, the to-be-processed data is representation information of a target compound, and the processing result corresponding to the to-be-processed data is a target activity of the target compound on the target target protein; and the processing unit is configured to invoke the target neural network model to process the representation information of the target compound to obtain the target activity of the target compound on the target target protein.

[0035] In a possible implementation, the apparatus also includes:

[0036] The determining unit is configured to determine a lead compound of the target target protein from the target compound based on the target activity of the target compound on the target target protein.

[0037] In another aspect, a computer device is provided, which includes a processor and a memory, and the memory stores at least one computer program, which is loaded and executed by the processor to enable the computer device to implement the data processing method.

[0038] In another aspect, a computer readable storage medium is also provided, which stores at least one computer program, which is loaded and executed by a processor to enable a computer to implement the data processing method.

[0039] In another aspect, a computer program product is also provided, which comprises computer programs or computer instructions loaded and executed by a processor to enable a computer to implement any of the above-mentioned data processing methods.

[0040] The technical solutions provided by the embodiments of the present application at least bring the following beneficial effects:

[0041] The technical solutions provided by the embodiments of the present application consider the noise measurement index of the candidate task to obtain the sampling probability of the candidate task, and the noise measurement index is used to measure the interference degree of the candidate task on the parameters of the updated model. The sampling probability obtained by considering the noise measurement index is a sampling probability matched with the interference degree of the candidate task on the parameters of the updated model. The reliability of the reference task selected from the candidate task based on such a sampling probability is higher, the reliability of the parameters of the model updated based on the reference task with higher reliability is higher, which is conducive to improving the update quality of the parameters of the model. The model based on the parameter update can learn the target task more accurately, thereby being conducive to improving the quality of the obtained target neural network model. The target neural network model can be used to process the data matched with the target task more accurately. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0043] Figure 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application;

[0044] Figure 2 is a schematic diagram of a whole process of drug research and development provided by an embodiment of the present application;

[0045] Figure 3 is a flowchart of a data processing method provided by an embodiment of the present application;

[0046] Figure 4 is a schematic diagram of a process of calling a target scheduling model to obtain a sampling probability of a first candidate task based on a noise measurement index of the first candidate task provided by an embodiment of the present application;

[0047] Figure 5 is a flowchart of an implementation manner of training a target scheduling model provided by an embodiment of the present application;

[0048] Figure 6is a schematic diagram of a process for implementing parameter updating according to a scheduling model provided by an embodiment of the present application.

[0049] Figure 7 is a flowchart of another implementation manner of training a target scheduling model provided by an embodiment of the present application.

[0050] Figure 8 is a flowchart of a data processing method provided by an embodiment of the present application.

[0051] Figure 9 is a schematic diagram of a data processing apparatus provided by an embodiment of the present application.

[0052] Figure 10 is a schematic diagram of a data processing apparatus provided by an embodiment of the present application.

[0053] Figure 11 is a schematic diagram of a server provided by an embodiment of the present application.

[0054] Figure 12 is a schematic diagram of a terminal provided by an embodiment of the present application. DETAILED DESCRIPTION

[0055] To make the objectives, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0056] It should be noted that the terms "first", "second", and the like in the present application are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. The implementation described in the following exemplary embodiments does not represent all implementations consistent with the present application. Rather, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0057] In exemplary embodiments, the data processing method provided by the embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, and assisted driving.

[0058] Artificial Intelligence (AI) is the use of digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, and artificial intelligence attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0059] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, automatic driving, intelligent transportation and several other major directions.

[0060] The scheme provided by the embodiments of the present application relates to machine learning technology in artificial intelligence technology. Machine learning (ML) is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, teaching learning, meta-learning and other technologies.

[0061] With the research and progress of artificial intelligence technology, artificial intelligence technology is researched and applied in many fields, such as common smart home, smart wearable device, virtual assistant, smart speaker, smart marketing, unmanned vehicle, autonomous vehicle, unmanned aerial vehicle, robot, intelligent medical treatment, intelligent customer service, Internet of Vehicles, autonomous driving, intelligent transportation, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0062] Figure 1 A schematic diagram of an implementation environment provided by the embodiments of the present application is shown. The implementation environment can include a terminal 11 and a server 12.

[0063] The data processing method provided in the embodiments of the present application can be executed by the terminal 11, can be executed by the server 12, and can also be executed by the terminal 11 and the server 12 together, which is not limited in the embodiments of the present application. For the case that the data processing method provided in the embodiments of the present application is executed by the terminal 11 and the server 12 together, the server 12 undertakes the main calculation work, and the terminal 11 undertakes the secondary calculation work; or the server 12 undertakes the secondary calculation work, and the terminal 11 undertakes the main calculation work; or the server 12 and the terminal 11 adopt a distributed computing architecture to perform collaborative calculation.

[0064] Optionally, the terminal 11 can be any electronic product that can interact with a user through one or more of a keyboard, a touchpad, a touch screen, a remote controller, voice interaction, a handwriting device, etc., such as a PC (Personal Computer), a mobile phone, a smart phone, a PDA (Personal Digital Assistant), a wearable device, a PPC (Pocket PC), a tablet computer, a smart car machine, a smart television, a smart sound box, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, a VR (Virtual Reality) device, an AR (Augmented Reality) device, etc. The server 12 can be a server or a server cluster composed of multiple servers, or a cloud computing service center. The terminal 11 and the server 12 establish a communication connection through a wired or wireless network.

[0065] Those skilled in the art should understand that the terminal 11 and the server 12 described above are only examples, and other existing or future terminal or server, such as those applicable to the present application, should also be included in the protection scope of the present application and are hereby incorporated by reference.

[0066] The data processing method provided in the embodiments of the present application involves a model training process based on meta-learning. Meta-learning is one of the common methods for solving small sample problems. Meta-learning can also be referred to as learning to learn, that is, using past knowledge and experience to guide the learning of new tasks, so that the model has the ability to learn to learn. Meta-learning hopes to enable the model to acquire the ability to learn to learn, so that it can quickly learn new tasks on the basis of existing knowledge. Meta-learning can effectively alleviate the problem of high computational cost caused by a large number of parameter tuning and task switching model retraining.

[0067] With the development of artificial intelligence technology, there are more and more types of meta-learning algorithms, for example, MAML (Model Agnostic Meta Learning), MetaMix (a meta-learning algorithm), FRML (a meta-learning algorithm), and the like. The core idea of MAML is to learn the initial parameters of a neural network, i.e., meta-parameters, from a large number of candidate tasks, which can enable a new machine learning task to quickly converge to a better solution under the condition of a small sample. MetaMix and FRML are both improved on the basis of MAML.

[0068] In the model training process based on meta-learning, the parameters of the neural network model need to be updated based on the candidate tasks, and then the target task is learned based on the model after the parameters are updated, so as to obtain a target neural network model suitable for the target task by using less data. Due to the memory limitation of the computer device, it is difficult to load all the data of the candidate tasks into the memory to participate in the parameter updating process, therefore, in the actual process, the strategy of sampling part of the tasks (reference tasks) to participate in the parameter updating process each time is adopted, which corresponds to the stochastic gradient optimization. In the related technology, the reference tasks are selected from the candidate tasks by means of uniform sampling, and then the parameters of the model are updated based on the reference tasks. The uniform sampling method gives the same sampling probability to different candidate tasks, and this method cannot handle the problem of a considerable number of tasks with large noise, which is easy to cause the parameter updating process to be biased by the tasks with large noise, and the generalization performance is reduced, thereby causing the updating quality of the parameters of the model to be poor, and the model based on the updated parameters is difficult to accurately learn the target task, and the quality of the obtained target neural network model is poor, and it is difficult to accurately process the data matched with the target task by using the target neural network model.

[0069] The model training process based on meta-learning involved in the data processing method provided by the embodiments of the present application considers the noise measurement index of the candidate task to obtain the sampling probability of the candidate task, and the noise measurement index is used to indicate the interference degree of the candidate task to the updated parameters of the model. The sampling probability obtained by considering the noise measurement index is the sampling probability matched with the interference degree of the candidate task to the updated parameters of the model, the reliability of the reference task selected from the candidate task based on this sampling probability is high, the reliability of the updated parameters of the model based on the reference task with high reliability is high, which is beneficial to improving the updating quality of the parameters of the model, the model after the parameters are updated can accurately learn the target task, thereby being beneficial to improving the quality of the obtained target neural network model, and the target neural network model can accurately process the data matched with the target task.

[0070] In an example embodiment, the meta-learning based model training process can be applied in various scenarios. For example, the meta-learning based model training process can be applied in a scenario of virtually screening lead compounds. The purpose of the scenario of virtually screening lead compounds is to virtually screen lead compounds for a specific target protein, where a lead compound refers to a compound that has preliminary activity for the specific target protein. By virtually screening lead compounds for the specific target protein, the drug exploration space can be effectively reduced, and the drug development can be accelerated. In the scenario of virtually screening lead compounds, the task on which the meta-learning based model training process is based is a task of predicting the activity of a compound for a target protein.

[0071] As one of the main means of virtual screening, a ligand-based virtual screening method (also referred to as a molecule structure-based virtual screening method) uses the activity data of molecules (or compounds) determined through wet experiments to establish a prediction model (activity prediction model) for predicting the activity of a compound for a target protein, so as to predict the activity of other compounds in a large compound library and screen out active lead compounds. Since the cost of wet experiments is huge, and the amount of measured activity data of compounds under a large number of target proteins is very small, the accuracy of the activity prediction model is hindered. The meta-learning based model training can transfer the meta-knowledge of the interaction between the molecular structure and the protein under other similar target proteins, thereby greatly improving the activity prediction performance of the compound under a small sample data.

[0072] Currently, the technologies for implementing the molecule structure-based virtual screening method through the meta-learning based model training mainly include: (1) MAML technology: The compounds with measured activity under each target protein are divided into a training set and a test set. This technology aims to find an initialization of an activity prediction model that is good for generalization for all target proteins. Starting from the initialization, through gradient optimization on the training set of a specific target protein for a number of steps, a neural network model (i.e., an activity prediction model) for predicting the activity of a compound for the specific target protein can be obtained, and the neural network model has good generalization performance on the test set of the specific target protein. (2) MetaMix technology: On the basis of the MAML technology, the training set and the test set in the target protein are linearly combined to expand the test set, so as to better evaluate the generalization performance and learn a better neural network model initialization. (3) FRML technology: On the basis of the MAML technology, the neural network structure is modularized, and different modules are adapted for different target proteins (tasks) to solve the problem of wide distribution and large difference of target proteins.

[0073] The current meta-learning-based model training process randomly samples meta-training tasks (a task is a task of predicting the activity of a compound against a certain target protein) with a uniform distribution probability, that is, all tasks are sampled with the same probability, without considering the presence of noise in the tasks. This sampling strategy is only effective when all tasks are equally important. However, due to the limitations of instrument measurement accuracy, changes in experimental environment, unsatisfied experimental conditions, and various factors, there are a considerable number of tasks with large noise in the virtual screening scenario, for example, all compounds in a certain task are measured to have the same activity value. Based on the meta-learning-based model training process involved in the data processing method provided in the embodiments of the present application, the sampling probability of the task can be obtained by considering the noise measurement index of the task, and then the task used to update the parameters in the current iteration is selected according to the sampling probability. This way can make the meta-learning-based model training process still effective under a considerable number of noisy tasks, and ensure the robustness of the meta-learning-based model training process in various virtual screening business scenarios.

[0074] Exemplarily, the virtual screening method based on molecular structure is a process in the whole process of drug research and development. Applying the meta-learning-based model training process involved in the data processing method provided in the embodiments of the present application in the process of the virtual screening method based on molecular structure can support the drug research and development platform to provide services for pharmaceutical factories and pharmaceutical chemistry fields, and accelerate the research and development of lead compounds. Exemplarily, the whole process of drug research and development can be as shown in Figure 2 The drug research and development process includes target recognition, lead compound discovery, lead compound discovery and optimization, candidate compound confirmation and development, clinical experiment, etc. In the target recognition link, the protein structure prediction process is included; in the lead compound discovery link, the virtual screening based on target structure, the virtual screening based on molecular structure, and the molecule generation process are included; in the lead compound discovery and optimization link, the ADMET (Absorption Distribution Metabolism Excretion Toxicity) property prediction and molecule generation process are included; in the candidate compound confirmation and development link, the synthetic route planning process is included.

[0075] It should be noted that the above-described meta-learning-based model training process applied to the scenario of virtual screening of lead compounds is merely an illustrative example, and this application does not limit it. Exemplarily, the meta-learning-based model training process can also be applied to image classification scenarios, molecular property prediction scenarios, etc. In image classification scenarios, the task based on the meta-learning-based model training process is to classify images. In molecular property prediction scenarios, the task based on the meta-learning-based model training process is to predict the properties of molecules. Exemplarily, the properties of a molecule can refer to its ADMET properties. Of course, the properties of a molecule can also be other properties, which will not be elaborated upon in this application.

[0076] This application provides a data processing method, which is applied to the above-mentioned... Figure 1 Taking the implementation environment shown as an example, this data processing method is executed by a computer device, which can be... Figure 1 Terminal 11 in the implementation environment shown can also be Figure 1 The server 12 in the implementation environment shown is not limited in this embodiment. Figure 3 As shown, the data processing method provided in this application embodiment may include the following steps 301 to 304.

[0077] In step 301, a noise metric for the first candidate task is obtained based on the initial neural network model. The noise metric is used to measure the degree of interference of the first candidate task with updating the parameters of the initial neural network model.

[0078] The first candidate task is the task that matches the target task to be learned.

[0079] The target task refers to the actual task that needs to be learned during the training of a meta-learning model. By learning the target task, the resulting target neural network model can more accurately process the data to be processed that matches the target task. The target task can be flexibly set according to the actual application scenario, and this application embodiment does not limit it.

[0080] Exemplarily, the target task refers to a task of predicting a target attribute of an object of a reference type. The reference type refers to a type of the object to be studied in an actual application scenario. The specific case of the object of the reference type can be flexibly set. For example, the object of the reference type can refer to a compound, or can refer to an image, and the like. Of course, in some embodiments, the object of the reference type can also refer to text, video, audio, and the like. The target attribute is used to describe the property of an aspect of the object of the reference type. The specific case of the target attribute can be flexibly set according to the specific case of the object of the reference type. Exemplarily, if the object of the reference type is a compound, the target attribute can refer to the activity on a target protein. The target protein can be any target protein. The target protein refers to a biological macromolecule that can bind to a drug molecule (which can also be referred to as a drug compound) and produce a pharmacological effect. Exemplarily, if the object of the reference type is an image, the target attribute can refer to whether it belongs to a target category. The target category can be any image category.

[0081] The specific type of the target task is different under different cases of the object of the reference type and the target attribute. Exemplarily, if the object of the reference type is a compound and the target attribute is the activity on a target protein, the target task can be a task of predicting the activity of the compound on the target protein. The activity of the compound on the target protein is used to indicate the possibility that the compound can act on the target protein. The greater the activity of the compound on the target protein, the greater the possibility that the compound can act on the target protein.

[0082] Exemplarily, if the object of the reference type is an image and the target attribute is whether it belongs to a target category, the target task can be a task of predicting whether the image belongs to the target category. Exemplarily, the task of predicting whether the image belongs to the target category can also be referred to as a task of performing binary classification on the image with respect to the target category.

[0083] The type of the target task described above is only exemplarily exemplified, and the embodiments of the present application are not limited thereto. The type of the target task can also be other cases. Exemplarily, the target task can also be a task of predicting the ADME property of a molecule, a task of predicting the language category of text, and the like.

[0084] After the target task is determined, a first candidate task matching the target task can be determined. The first candidate task matching the target task refers to a task that provides prior knowledge for the learning of the target task in the process of the meta-learning-based model training, so as to reduce the data required for learning the target task. Exemplarily, the provision of the prior knowledge for the learning of the target task is embodied in the process of updating the parameters of the initial neural network model according to the first candidate task to obtain the model for learning the target task. That is, the first candidate task is a task that is preset to be available for updating the parameters of the initial neural network model in the process of the meta-learning-based model training. Exemplarily, the number of the first candidate tasks is multiple.

[0085] The type of the first candidate task is related to the type of the target task. Exemplarily, the target task is a task of predicting a target attribute of an object of a reference type, and the first candidate task is a task of predicting a candidate attribute of the object of the reference type. The candidate attribute is an attribute matching the target attribute, and the candidate attribute can be the same as the target attribute or different from the target attribute. For the case that the candidate attribute is different from the target attribute, the candidate attribute and the target attribute are similar attributes to ensure the training effect of the model.

[0086] Exemplarily, for the case that the object of the reference type is a compound, the target attribute is an activity against a target protein (i.e., the target task is a task of predicting the activity of the compound against the target protein), the candidate attribute can be an activity against a first candidate protein, and the first candidate protein can be the same as or different from the target protein. In this case, the first candidate task is a task of predicting the activity of the compound against the first candidate protein. Exemplarily, the types of the first candidate proteins against which different first candidate tasks are directed can be the same or different. Exemplarily, if the types of the first candidate proteins against which different first candidate tasks are directed are different, it is beneficial to further improve the generalization ability of the parameters of the model updated according to the first candidate task.

[0087] Exemplarily, for the case that the object of the reference type is an image, the target attribute is whether it belongs to a target category (i.e., the target task is a task of predicting whether the image belongs to the target category), the candidate attribute can be whether it belongs to a first candidate category, and the first candidate category can be the same as or different from the target category. In this case, the first candidate task is a task of predicting whether the image belongs to the first candidate category. Exemplarily, the first candidate categories against which different first candidate tasks are directed can be the same or different. Exemplarily, if the first candidate categories against which different first candidate tasks are directed are different, it is beneficial to further improve the generalization ability of the parameters of the model updated according to the first candidate task.

[0088] The above-mentioned types of the first candidate task are only exemplary, and embodiments of the present application are not limited thereto. The types of the first candidate task can also be other cases. For example, the first candidate task can also be a task of predicting ADME properties of a molecule, a task of predicting a language category of text, etc.

[0089] For example, the first candidate task corresponds to task data, and the task data corresponding to the first candidate task is data related to the first candidate task. For example, the task data corresponding to different first candidate tasks is different, so as to ensure the generalization ability of the parameters of the model updated according to the first candidate task and improve the training quality of the model training based on meta-learning. For example, the task data corresponding to the first candidate task is labeled data, and the type of the task data corresponding to the first candidate task is related to the type of the first candidate task. For example, the task data corresponding to the first candidate task can include task sample data and standard processing results corresponding to the task sample data, and the standard processing results corresponding to the task sample data refer to the labels of the task sample data.

[0090] For example, the first candidate task is a task of predicting candidate attributes of an object of a reference type, and the task sample data can refer to the representation information of the sample object, and the standard processing results corresponding to the task sample data can refer to the candidate attribute labels of the sample object. The type of the sample object is the reference type. The representation information of the sample object is used to represent the sample object, and the candidate attribute labels of the sample object are used to indicate the real candidate attributes of the sample object.

[0091] Next, the representation information of the sample object and the candidate attribute labels of the sample object will be exemplarily described in combination with the specific type of the first candidate task.

[0092] For example, if the first candidate task is a task of predicting the activity of a compound on a first candidate target protein, the representation information of the sample object can refer to the representation information of the sample compound, and the candidate attribute labels of the sample object can refer to the standard activity of the sample compound on the first candidate target protein. In this case, the task sample data refers to the representation information of the sample compound, and the standard processing results corresponding to the task sample data refer to the standard activity of the sample compound on the first candidate target protein. For example, the sample compound can be any compound with known standard activity on the first candidate target protein, and the standard activity of the sample compound on the first candidate target protein can be determined by wet experiments. The wet experiment refers to research by using molecular, cellular, and physiological experimental methods in a laboratory.

[0093] The characterization information of the sample compound is used to characterize the sample compound. Exemplarily, the characterization information of the sample compound can refer to a molecular formula of the sample compound, a molecular string of the sample compound, or a Morgan fingerprints feature of the sample compound. Exemplarily, the characterization information of the sample compound can be in the form of a vector or a matrix, and the dimension of the vector or the matrix is set empirically or adjusted flexibly according to an application scenario, for example, the dimension is 1024.

[0094] The number of task data included in the first candidate task is at least one. Exemplarily, the more the number of task data included in the first candidate task, the greater the contribution of the first candidate task to the model update can be. Exemplarily, assuming that n (n is an integer not less than 1) sample compounds are determined for activity of a first candidate target protein in an activity determination experiment of the first candidate target protein, the number of task data corresponding to the first candidate task is n, and the task data corresponding to the first candidate task can be represented as wherein x i represents the characterization information of the sample compound in the i th (i is an integer not less than 1 and not greater than n) task data, y i represents the standard activity of the sample compound in the i th task data for the first candidate target protein.

[0095] Exemplarily, if the first candidate task is a task of predicting whether an image belongs to a first candidate category, the characterization information of the sample object can refer to the characterization information of the sample image, and the candidate attribute label of the sample object can refer to the standard classification result corresponding to the sample image. In this case, the task sample data refers to the characterization information of the sample image, and the standard processing result corresponding to the task sample data refers to the standard classification result corresponding to the sample image. Exemplarily, the standard classification result corresponding to the sample image is that the sample image belongs to the first candidate category or that the sample image does not belong to the first candidate category, and the standard classification result corresponding to the sample image can be obtained by being labeled by a technician. The sample image can be any image with a known standard classification result. The characterization information of the sample image is used to characterize the sample image. Exemplarily, the characterization information of the sample image can refer to pixel values of pixel points in the sample image, or refer to image features of the sample image, etc.

[0096] Of course, when the first candidate task is a task of other types, the characterization information of the sample object and the candidate attribute label of the sample object can also be other cases, which are not described one by one here.

[0097] Exemplarily, the task data corresponding to the first candidate task can be divided into training data corresponding to the first candidate task and test data corresponding to the first candidate task. The training data corresponding to the first candidate task includes training sample data and a standard processing result corresponding to the training sample data, and the test data corresponding to the first candidate task includes test sample data and a standard processing result corresponding to the test sample data.

[0098] The standard processing result corresponding to the training sample data refers to a label of the training sample data, and the standard processing result corresponding to the test sample data refers to a label of the test sample data. Exemplarily, the first candidate task is a task of predicting a candidate attribute of a reference type object, the training sample data refers to representation information of a first object, the standard processing result corresponding to the training sample data refers to a candidate attribute label of the first object, the test sample data refers to representation information of a second object, and the standard processing result corresponding to the test sample data refers to a candidate attribute label of the second object. Exemplarily, the first object and the second object can be different objects in sample objects.

[0099] Next, the representation information of the first object, the candidate attribute label of the first object, the representation information of the second object, and the candidate attribute label of the second object are exemplarily described in combination with the specific type of the first candidate task.

[0100] Exemplarily, for the case that the first candidate task is a task of predicting activity of a compound on a first candidate target protein, the representation information of the first object can refer to representation information of a first compound, the candidate attribute label of the first object can refer to standard activity of the first compound on the first candidate target protein, the representation information of the second object can refer to representation information of a second compound, and the candidate attribute label of the second object can refer to standard activity of the second compound on the first candidate target protein. Exemplarily, the first compound and the second compound can be different compounds in sample compounds. In this case, the training sample data refers to the representation information of the first compound, the standard processing result corresponding to the training sample data refers to the standard activity of the first compound on the first candidate target protein, the test sample data refers to the representation information of the second compound, and the standard processing result corresponding to the test sample data refers to the standard activity of the second compound on the first candidate target protein.

[0101] Exemplarily, for the case that the first candidate task is a task of predicting whether an image belongs to a first candidate category, the representation information of the first object can refer to representation information of the first image, the candidate attribute label of the first object can refer to a standard classification result corresponding to the first image, the representation information of the second object can refer to representation information of the second image, and the candidate attribute label of the second object can refer to a standard classification result corresponding to the second image. Exemplarily, the first image and the second image are different images in the sample images. In this case, the training sample data refers to the representation information of the first image, the standard processing result corresponding to the training sample data refers to the standard classification result corresponding to the first image, the test sample data refers to the representation information of the second image, and the standard processing result corresponding to the test sample data refers to the standard classification result corresponding to the second image.

[0102] The initial neural network model is a model to be trained by using the model training process based on meta-learning. The structure of the initial neural network model can be set according to experience, or can be flexibly adjusted according to the type of the target task, which is not limited in the embodiments of the present application. Exemplarily, the structure of the initial neural network model is a multi-layer perceptron with m (m is an integer not less than 1) layers, and the value of m is set according to experience, for example, the value of m is 4, or the value of m is 3, etc. The first multi-layer perceptron in the m-layer multi-layer perceptron is an input layer, and the other multi-layer perceptrons are processing layers, and the number of neurons included in the processing layers can be set according to experience or flexibly adjusted according to the actual application scenario, for example, taking the value of m as 3 as an example, the number of neurons included in the middle three multi-layer perceptrons can be 128, 128, and 128 respectively, and of course, the number of neurons can also be other cases.

[0103] The function of the initial neural network model is related to the type of the target task. Exemplarily, when the target task is a task of predicting the activity of a compound on a target protein, the function of the initial neural network model is to predict the activity of the compound. When the target task is a task of predicting whether an image belongs to a target category, the function of the initial neural network model is to perform binary classification on the image. When the target task is a task of predicting the ADME properties of a molecule, the function of the initial neural network model is to predict the ADME properties of the molecule.

[0104] In the process of training the initial neural network model based on meta-learning, the parameters of the initial neural network model are first updated based on the first candidate task. Since there can be tasks with large noise in the first candidate task, such tasks are not conducive to the updating of the parameters of the model, therefore, after the first candidate task is determined, the noise measurement index of the first candidate task is obtained based on the initial neural network model, so as to select the tasks not conducive to the updating of the parameters according to the noise measurement index.

[0105] The noise measurement index of the first candidate task is used to measure the noise of the first candidate task, and the noise of the first candidate task refers to the interference degree of the first candidate task on updating the parameters of the initial neural network model, that is, the noise measurement index of the first candidate task is used to measure the interference degree of the first candidate task on updating the parameters of the initial neural network model. Exemplarily, the noise measurement index of the first candidate task is positively correlated with the interference degree of the first candidate task on updating the parameters of the initial neural network model, that is, the greater the noise measurement index of the first candidate task, the greater the interference degree of the first candidate task on updating the parameters of the initial neural network model.

[0106] The implementation manner of the embodiment of the present application for obtaining the noise measurement index of the first candidate task based on the initial neural network model is not limited, as long as the noise measurement index can measure the interference degree of the first candidate task on updating the parameters of the initial neural network model, that is, the noise of the first candidate task can be measured.

[0107] In a possible implementation manner, the first candidate task corresponds to training data and test data, and the implementation manner of obtaining the noise measurement index of the first candidate task based on the initial neural network model includes the following steps 3011 to 3024.

[0108] Step 3011: Obtain a task processing model corresponding to the first candidate task based on the training data corresponding to the first candidate task and the initial neural network model.

[0109] The task processing model corresponding to the first candidate task refers to a model for processing the first candidate task which is trained based on the initial neural network model and the training data corresponding to the first candidate task.

[0110] In an exemplary embodiment, the implementation process of obtaining the task processing model corresponding to the first candidate task based on the training data corresponding to the first candidate task and the initial neural network model includes: constructing a reference model which is the same as the initial neural network model in structure and parameters; training the reference model by using the training data corresponding to the first candidate task, and taking the trained model as the task processing model corresponding to the first candidate task.

[0111] Exemplarily, the training data corresponding to the first candidate task includes training sample data corresponding to the first candidate task and a standard processing result corresponding to the training sample data, and the process of training the reference model by using the training data corresponding to the first candidate task includes: calling the reference model to process the training sample data corresponding to the first candidate task to obtain a predicted processing result corresponding to the training sample data; obtaining a first loss based on the predicted processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data; updating parameters of the reference model based on the first loss to obtain a task processing model corresponding to the first candidate task. The parameters of the reference model are the same as the parameters of the initial neural network model.

[0112] The first loss is used to measure the difference between the predicted processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data, and the first loss can be calculated by substituting the predicted processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data into a loss function. The type of the loss function used to calculate the first loss is not limited in the embodiments of the present application, and can be set according to experience or flexibly adjusted according to application scenarios. For example, the loss function can be a cross-entropy loss function, or a mean square error loss function, etc. It should be noted that the number of the training data corresponding to the first candidate task is one or more, and the first loss is obtained based on all the training data corresponding to the first candidate task.

[0113] Exemplarily, taking the number of the training data corresponding to the first candidate task as n (n is an integer not less than 1) and the loss function as a mean square error loss function as an example, the first loss can be calculated based on the following formula 1:

[0114]

[0115] wherein, x i represents the training sample data in the i th (i is an integer not less than 1 and not greater than n) training data corresponding to the first candidate task; y i represents the standard processing result in the i th training data corresponding to the first candidate task; θ represents the parameters of the reference model; f θ (x i ) represents the predicted processing result obtained by calling the reference model; represents the first loss.

[0116] Exemplarily, the process of updating the parameters of the reference model based on the first loss to obtain the task processing model corresponding to the first candidate task includes: obtaining an update gradient of the parameters of the reference model based on the first loss; updating the parameters of the reference model based on the update gradient, and in response to the update process satisfying a first condition, regarding a model having the parameters obtained when the update process satisfies the first condition as the task processing model corresponding to the first candidate task.

[0117] Since the first loss is a loss obtained by forward propagation on the basis of the parameters of the reference model, after the first loss is obtained, a gradient of the first loss with respect to the parameters of the reference model can be calculated by a chain rule, and the gradient is taken as an update gradient of the parameters of the reference model. After the update gradient of the parameters of the reference model is calculated, the parameters of the reference model are updated based on the update gradient. For example, the parameters of the reference model are updated based on the update gradient in the following manner: a first product of a first learning rate and the update gradient is calculated, and a difference between the parameters of the reference model and the first product is taken as updated parameters of the reference model. The first learning rate is set according to experience or is flexibly adjusted according to an application scenario, and the embodiments of the present application do not limit this.

[0118] For example, taking the first candidate task as the kth (k is an integer not less than 1) first candidate task, the process of updating the parameters of the reference model based on the update gradient can be implemented based on formula 2:

[0119]

[0120] wherein θ0 represents the parameters of the reference model; and a represents the first learning rate. represents training sample data in the training data corresponding to the kth first candidate task; represents the standard processing result in the training data corresponding to the kth first candidate task; represents a predicted processing result obtained based on the training sample data in the training data corresponding to the kth first candidate task; represents the first loss; represents the update gradient of the parameters of the reference model; and θ' represents the updated parameters of the reference model.

[0121] After the parameters of the reference model are updated based on the update gradient, it is determined whether the update process satisfies a first condition. If the update process satisfies the first condition, a model having parameters obtained when the update process satisfies the first condition is taken as a task processing model corresponding to the first candidate task. If the update process does not satisfy the first condition, the model having the updated parameters is continuously trained by using the training data corresponding to the first candidate task until the update process satisfies the first condition, and a model having parameters obtained when the update process satisfies the first condition is taken as the task processing model corresponding to the first candidate task.

[0122] The first condition is set empirically or adjusted flexibly according to application scenarios, and embodiments of the present application do not limit the same. For example, the first condition can be that the number of updates reaches a first number threshold, the first loss converges, the first loss is less than a first loss threshold, and the like. The first number threshold and the first loss threshold are set empirically or adjusted flexibly according to application scenarios.

[0123] For example, the first candidate task is a task of predicting a candidate attribute of an object of a reference type, the training sample data corresponding to the first candidate task refers to representation information of the first object, the standard processing result corresponding to the training sample data refers to a candidate attribute label of the first object, and the predicted processing result corresponding to the training sample data refers to a candidate attribute prediction result of the first object. In this case, the process of obtaining the task processing model corresponding to the first candidate task includes: calling the reference model to process the representation information of the first object to obtain the candidate attribute prediction result of the first object; obtaining the first loss based on the candidate attribute prediction result of the first object and the candidate attribute label of the first object; and updating the parameters of the reference model based on the first loss to obtain the task processing model corresponding to the first candidate task.

[0124] Next, the process of obtaining the task processing model corresponding to the first candidate task is introduced in combination with specific types of the representation information of the first object, the candidate attribute prediction result of the first object, and the candidate attribute label of the first object.

[0125] For example, when the first candidate task is a task of predicting the activity of a compound on a first candidate target protein, the representation information of the first object refers to representation information of the first compound, the candidate attribute label of the first object refers to a standard activity of the first compound on the first candidate target protein, and the candidate attribute prediction result of the first object refers to a predicted activity of the first compound on the first candidate target protein. In this case, the process of obtaining the task processing model corresponding to the first candidate task includes: calling the reference model to process the representation information of the first compound to obtain the predicted activity of the first compound on the first candidate target protein; obtaining the first loss based on the predicted activity of the first compound on the first candidate target protein and the standard activity of the first compound on the first candidate target protein; and updating the parameters of the reference model based on the first loss to obtain the task processing model corresponding to the first candidate task.

[0126] Exemplarily, when the first candidate task is a task of predicting whether an image belongs to a first candidate category, the representation information of the first object refers to representation information of the first image, the candidate attribute label of the first object refers to a standard classification result corresponding to the first image, and the candidate attribute prediction result of the first object refers to a predicted classification result corresponding to the first image. In this case, the process of obtaining the task processing model corresponding to the first candidate task includes: calling the reference model to process the representation information of the first image to obtain the predicted classification result corresponding to the first image; obtaining the first loss based on the predicted classification result corresponding to the first image and the standard classification result corresponding to the first image; and obtaining the update of the parameters of the reference model based on the first loss to obtain the task processing model corresponding to the first candidate task.

[0127] Step 3012: testing the task processing model corresponding to the first candidate task by using test data corresponding to the first candidate task to obtain a test performance of the task processing model corresponding to the first candidate task.

[0128] The test performance of the task processing model corresponding to the first candidate task is used to measure the performance of the task processing model corresponding to the first candidate task in processing the first candidate task.

[0129] In an exemplary embodiment, the test data corresponding to the first candidate task includes test sample data corresponding to the first candidate task and a standard processing result corresponding to the test sample data. The process of testing the task processing model corresponding to the first candidate task by using the test data corresponding to the first candidate task to obtain the test performance of the task processing model corresponding to the first candidate task includes: calling the task processing model corresponding to the first candidate task to process the test sample data corresponding to the first candidate task to obtain a test processing result corresponding to the test sample data; obtaining a second loss based on the test processing result corresponding to the test sample data and the standard processing result corresponding to the test sample data; and taking the second loss as the test performance of the task processing model corresponding to the first candidate task.

[0130] The principle of obtaining the second loss is the same as that of obtaining the first loss, which will not be repeated here. In the case of taking the second loss as the test performance of the task processing model corresponding to the first candidate task, the smaller the test performance is, the better the performance of the task processing model corresponding to the first candidate task in processing the first candidate task is; the larger the test performance is, the worse the performance of the task processing model corresponding to the first candidate task in processing the first candidate task is.

[0131] Exemplarily, the first candidate task is a task of predicting a candidate attribute of an object of a reference type, the test sample data corresponding to the first candidate task refers to representation information of a second object, the standard processing result corresponding to the test sample data refers to a candidate attribute label of the second object, and the test processing result corresponding to the test sample data refers to a candidate attribute test result of the second object. In this case, the process of obtaining the test performance of the task processing model corresponding to the first candidate task includes: calling the task processing model corresponding to the first candidate task to process the representation information of the second object to obtain the candidate attribute test result of the second object; obtaining a second loss based on the candidate attribute test result of the second object and the candidate attribute label of the second object; and taking the second loss as the test performance of the task processing model corresponding to the first candidate task.

[0132] Next, the process of obtaining the test performance of the task processing model corresponding to the first candidate task is introduced in combination with specific types of the representation information of the second object, the candidate attribute test result of the second object, and the candidate attribute label of the second object.

[0133] Exemplarily, when the first candidate task is a task of predicting the activity of a compound on a first candidate target protein, the representation information of the second object refers to representation information of a second compound, the candidate attribute label of the second object refers to a standard activity of the second compound on the first candidate target protein, and the candidate attribute test result of the second object refers to a test activity of the second compound on the first candidate target protein. In this case, the process of obtaining the test performance of the task processing model corresponding to the first candidate task includes: calling the task processing model corresponding to the first candidate task to process the representation information of the second compound to obtain the test activity of the second compound on the first candidate target protein; obtaining a second loss based on the test activity of the second compound on the first candidate target protein and the standard activity of the second compound on the first candidate target protein; and taking the second loss as the test performance of the task processing model corresponding to the first candidate task.

[0134] Exemplarily, when the first candidate task is a task of predicting whether an image belongs to a first candidate category, the representation information of the second object refers to representation information of a second image, the candidate attribute label of the second object refers to a standard classification result corresponding to the second image, and the candidate attribute test result of the second object refers to a test classification result corresponding to the second image. In this case, the process of obtaining the test performance of the task processing model corresponding to the first candidate task includes: calling the task processing model corresponding to the first candidate task to process the representation information of the second image to obtain the test classification result corresponding to the second image; obtaining a second loss based on the test classification result corresponding to the second image and the standard classification result corresponding to the second image; and taking the second loss as the test performance of the task processing model corresponding to the first candidate task.

[0135] Step 3013: obtaining a first update gradient corresponding to the parameters of the initial neural network model based on the training data corresponding to the first candidate task; and obtaining a second update gradient corresponding to the parameters of the initial neural network model based on the test data corresponding to the first candidate task.

[0136] The first update gradient is an update gradient of the parameters of the initial neural network model obtained based on the training data corresponding to the first candidate task. In an example embodiment, the training data corresponding to the first candidate task includes training sample data and a standard processing result corresponding to the training sample data. The implementation process of obtaining the first update gradient corresponding to the parameters of the initial neural network model based on the training data corresponding to the first candidate task includes: calling the initial neural network model to process the training sample data to obtain a predicted processing result corresponding to the training sample data; obtaining a third loss based on the predicted processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data; and obtaining the first update gradient corresponding to the parameters of the initial neural network model based on the third loss. The principle of the above implementation process of obtaining the first update gradient is the same as that of the implementation process of obtaining the update gradient of the parameters of the reference model in step 3011, and the difference lies in that the model relied on in step 3011 is the reference model, and the model relied on here is the initial neural network model.

[0137] The second update gradient is an update gradient of the parameters of the initial neural network model obtained based on the test data corresponding to the first candidate task. In an example embodiment, the test data corresponding to the first candidate task includes test sample data and a standard processing result corresponding to the test sample data. The implementation process of obtaining the second update gradient corresponding to the parameters of the initial neural network model based on the test data corresponding to the first candidate task includes: calling the initial neural network model to process the test sample data to obtain a test processing result corresponding to the test sample data; obtaining a fourth loss based on the test processing result corresponding to the test sample data and the standard processing result corresponding to the test sample data; and obtaining the second update gradient corresponding to the parameters of the initial neural network model based on the fourth loss. The principle of the above implementation process of obtaining the second update gradient is the same as that of the implementation process of obtaining the update gradient of the parameters of the reference model in step 3011, and the difference lies in that the model relied on in step 3011 is the reference model, and the model relied on here is the initial neural network model, and the data relied on in step 3011 is the training data corresponding to the first candidate task, and the data relied on here is the test data corresponding to the first candidate task.

[0138] Step 3014: obtaining a noise measurement index of the first candidate task based on the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task.

[0139] After obtaining the first update gradient and the second update gradient, similarity between the first update gradient and the second update gradient is calculated. The manner of calculating the similarity between the first update gradient and the second update gradient can be set according to experience, or can be flexibly adjusted according to application scenarios, and the embodiments of the present application do not limit this. Exemplarily, the forms of the first update gradient and the second update gradient are both vectors, and the inner product between the first update gradient and the second update gradient can be taken as the similarity between the first update gradient and the second update gradient. Of course, the similarity between the first update gradient and the second update gradient can also be calculated by calculating the Euclidean distance between the first update gradient and the second update gradient, etc.

[0140] Exemplarily, since the first update gradient is obtained on the basis of the training data corresponding to the first candidate task, and the second update gradient is obtained on the basis of the test data corresponding to the first candidate task, the similarity between the first update gradient and the second update gradient can be used to measure the generalization distance between the training data and the test data corresponding to the first candidate task, and the greater the similarity between the first update gradient and the second update gradient, the smaller the generalization distance between the training data and the test data corresponding to the first candidate task.

[0141] The similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task respectively represent the result and the process of learning on the first candidate task with the parameters of the initial neural network model as initialization. Exemplarily, the similarity between the first update gradient and the second update gradient can be considered as the inner product of the parameter update gradient of the initial neural network model with respect to the training data and the test data of the first candidate task. The test performance of the task processing model corresponding to the first candidate task can be considered as the objective function value (i.e. loss) of the initial neural network model on the test data corresponding to the first candidate task.

[0142] Exemplarily, when the test performance of the task processing model corresponding to the first candidate task is large, there are two possible reasons: the first reason is that the learning difficulty of the first candidate task is large, and the initial neural network model has not learned how to better process the first candidate task; the second reason is that the noise of the first candidate task is large, and the initial neural network model cannot learn how to better process the first candidate task. In combination with the similarity between the first update gradient and the second update gradient, when the similarity is small, it indicates that the difference between the training data and the test data corresponding to the first candidate task is large, that is, the first candidate task itself has a high learning difficulty; and when the similarity is large, it indicates that the first candidate task itself does not have a generalization challenge, that is, the learning difficulty of the first candidate task is small (in this case, because there is no noise and the task is simple, the test performance cannot be large), or the noise is uniformly distributed on the training data and test data corresponding to the first candidate task, that is, the noise is large.

[0143] In summary, when the test performance of the task processing model corresponding to the first candidate task is large, and the similarity between the first update gradient and the second update gradient is large, it is very likely that the noise of the first candidate task is large, that is, it is very likely that the interference degree of the first candidate task on updating the parameters of the initial neural network model is large. In other cases (such as the test performance of the task processing model corresponding to the first candidate task is large, and the similarity between the first update gradient and the second update gradient is small), it is more likely that the noise of the first candidate task is small, that is, it is more likely that the interference degree of the first candidate task on updating the parameters of the initial neural network model is small, and it can also indicate that the first candidate task is a truly challenging task that can provide a large contribution to parameter updating.

[0144] According to the above analysis, the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task can to some extent reflect the size of the noise of the first candidate task, that is, the size of the interference degree of the first candidate task on updating the parameters of the initial neural network model, so after obtaining the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task, the indicators including the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task can be used as the noise measurement indicator of the first candidate task.

[0145] In another possible implementation manner, the implementation manner of obtaining the noise measurement indicator of the first candidate task based on the initial neural network model includes the following steps 301A to 301C.

[0146] Step 301A: based on the training data corresponding to the first candidate task and the initial neural network model, obtaining a task processing model corresponding to the first candidate task.

[0147] The implementation of this step 301A is described in the above step 3011, which will not be repeated here.

[0148] Step 301B: testing the task processing model corresponding to the first candidate task by using the test data corresponding to the first candidate task, to obtain the test performance of the task processing model corresponding to the first candidate task.

[0149] The implementation of this step 301B is described in the above step 3012, which will not be repeated here.

[0150] Step 301C: based on the test performance of the task processing model corresponding to the first candidate task, obtaining the second-order update gradient corresponding to the parameters of the initial neural network model, and taking the second-order update gradient as the noise measurement index of the first candidate task.

[0151] Since the test performance of the task processing model corresponding to the first candidate task is obtained on the basis of the parameters of the task processing model corresponding to the first candidate task, and the parameters of the task processing model corresponding to the first candidate task are obtained on the basis of the parameters of the initial neural network model, after obtaining the test performance of the processing model corresponding to the first candidate task, the second-order gradient of the test performance of the processing model corresponding to the first candidate task with respect to the parameters of the initial neural network model can be calculated by the chain rule, and the second-order gradient is taken as the second-order update gradient corresponding to the parameters of the initial neural network model. Illustratively, the second-order update gradient can be represented by a Hessian matrix. Illustratively, the second-order update gradient can be considered as the second-order update gradient of the parameters of the initial neural network model with respect to the test data corresponding to the first candidate task. Illustratively, the second-order update gradient can also be referred to as the second-order curvature.

[0152] Exemplarily, the second-order gradient can generally represent a smoothness around the test performance of the processing model corresponding to the first candidate task. When the second-order update gradient is large, the test performance of the processing model corresponding to the first candidate task is less smooth, and the generalization difficulty of the initial neural network model with respect to the first candidate task is greater, that is, it is indicated that the noise of the first candidate task can be greater. Therefore, the second-order update gradient can be used as an index reflecting another dimension of the noise of the first candidate task. If the second-order update gradient is large, it is indicated that the first candidate task has a greater possibility of noise, that is, there is a greater possibility that the first candidate task has a greater interference degree on updating the parameters of the initial neural network model. If the second-order update gradient is small, it is indicated that the first candidate task has a smaller possibility of noise, that is, there is a greater possibility that the first candidate task has a smaller interference degree on updating the parameters of the initial neural network model.

[0153] Exemplarily, the smaller interference degree of the first candidate task on updating the parameters of the initial neural network model can be considered as updating the parameters of the initial neural network model using the first candidate task, which is beneficial to guarantee the generalization performance of the updated parameters. Exemplarily, the smaller second-order update gradient can be considered as a greater local linearity degree of the second-order curvature.

[0154] According to the above analysis, the second-order update gradient can reflect the size of the noise of the first candidate task to a certain extent, that is, reflect the size of the interference degree of the first candidate task on updating the parameters of the initial neural network model. Therefore, after obtaining the second-order update gradient, the second-order update gradient can be used as a noise measurement index of the first candidate task.

[0155] It should be noted that the implementation processes of the steps 3011 to 3014 and the steps 301A to 301C are only exemplary implementation processes of obtaining the noise measurement index of the first candidate task, and the embodiments of the present application are not limited thereto. Exemplarily, the three indexes including the similarity between the first update gradient and the second update gradient, the test performance of the task processing model corresponding to the first candidate task, and the second-order update gradient can be used as the noise measurement index of the first candidate task.

[0156] The embodiments of the present application take one first candidate task as an example to introduce the implementation process of obtaining the noise measurement index of the first candidate task. For the case that the number of first candidate tasks is multiple, the noise measurement index of each first candidate task needs to be obtained respectively. Exemplarily, the noise measurement indexes of different first candidate tasks are obtained in the same principle manner to ensure that the noise measurement indexes of different first candidate tasks have comparability.

[0157] In step 302, based on the noise measurement index of the first candidate task, the sampling probability of the first candidate task is obtained; and a reference task whose sampling probability meets a first selection condition is selected from the first candidate task.

[0158] The sampling probability of the first candidate task is a probability of selecting the first candidate task to update the parameters of the initial neural network model. According to the sampling probability of the first candidate task, it can be determined which task is selected to update the parameters of the initial neural network model. The sampling probability of the first candidate task is obtained by considering the noise measurement index of the first candidate task. The noise measurement index is used to indicate the interference degree of the first candidate task on updating the parameters of the model. The sampling probability obtained by considering the noise measurement index is the sampling probability matched with the interference degree of the first candidate task on updating the parameters of the model. Based on this sampling probability, the reliability of selecting the task to update the parameters of the initial neural network model from the candidate task is high, thereby facilitating to improve the updating quality of the parameters of the model, improve the robustness of the model training process based on meta-learning in the case that there are tasks with larger noise, and improve the training quality of the model training based on meta-learning.

[0159] The implementation manner of obtaining the sampling probability of the first candidate task based on the noise measurement index of the first candidate task is not limited in the embodiments of the present application. It can be set according to experience or flexibly adjusted according to actual application scenarios. Illustratively, the greater the noise measurement index of the first candidate task, the greater the noise of the first candidate task, and the greater the interference degree of the first candidate task on updating the parameters of the initial neural network model, and the less the information contributed to updating the parameters of the initial neural network model. The goal of obtaining the sampling probability of the first candidate task based on the noise measurement index of the first candidate task is to give the first candidate task a greater sampling probability if the noise measurement index of the first candidate task is smaller, and to give the first candidate task a smaller sampling probability if the noise measurement index of the first candidate task is greater. Illustratively, if the noise measurement index of the first candidate task is smaller, it can be considered that the first candidate task is a truly challenging task for updating the parameters.

[0160] In an exemplary embodiment, the implementation manner of obtaining the sampling probability of the first candidate task based on the noise measurement index of the first candidate task includes but is not limited to the following manner 1 and manner 2.

[0161] Manner 1: calling the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task.

[0162] The target scheduling model is a trained model capable of inputting a sampling probability of a task according to a noise measurement index of the input task. The target scheduling model is called to obtain the sampling probability of the first candidate task, which is beneficial to improving the reliability and accuracy of the sampling probability of the first candidate task. The calling of the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task is related to the structure of the target scheduling model, and the embodiments of the present application do not limit the implementation manner.

[0163] In the example embodiment, the implementation process of calling the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task includes the following steps 3021 to 3023.

[0164] Step 3021: The target scheduling model is called to encode the current parameter update round to obtain first encoding information.

[0165] The current parameter update round refers to the current update round of the parameters of the neural network model. The current parameter is updated once based on the selected task, and the current parameter update round is increased by 1. For example, the parameters of the initial neural network model can refer to the parameters that have not been updated after initialization, or can refer to the parameters that have been updated one or more times after initialization, and the embodiments of the present application do not limit this. The current parameter update round can refer to the update round relative to the parameters after initialization. For example, if the parameters of the initial neural network model refer to the parameters that have not been updated after initialization, the reference value can be used as the current parameter update round when the reference task is selected. The reference value is set according to experience, for example, the reference value can be 1, or 0, etc. If the parameters of the initial neural network model refer to the parameters that have been updated t (t is an integer not less than 1) times after initialization, the sum of the reference value and L can be used as the current parameter update round when the reference task is selected.

[0166] The first encoding information is information obtained by encoding the current parameter update round, which can represent the current parameter update round. The representation form of the first encoding information of the embodiments of the present application is not limited, for example, the representation form of the first encoding information can be a vector, or a matrix, etc.

[0167] Exemplarily, the target scheduling model comprises a first encoder configured to encode the current parameter update round, that is, the first encoder in the target scheduling model is invoked to encode the current parameter update round to obtain first encoding information. The first encoder is any model capable of encoding, and the type of the first encoder is not limited in the embodiments of the present application and can be flexibly selected according to experience. Exemplarily, the first encoder can be a convolution model or an LSTM (Long Short-Term Memory) model.

[0168] In step 3022, the noise measurement indicator of the first candidate task is encoded to obtain second encoding information.

[0169] The first encoding information is information obtained by encoding the noise measurement indicator of the first candidate task and can represent the noise measurement indicator of the first candidate task. The representation form of the second encoding information is not limited in the embodiments of the present application. For example, the representation form of the second encoding information can be a vector or a matrix. Exemplarily, the representation form of the second encoding information is the same as that of the first encoding information to facilitate splicing.

[0170] Exemplarily, for the case that the noise measurement indicator of the first candidate task comprises multiple indicators (for example, the noise measurement indicator of the first candidate task comprises the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task), encoding the noise measurement indicator of the first candidate task can mean that each indicator included in the noise measurement indicator of the first candidate task is encoded. Each indicator in the noise measurement indicator is encoded to obtain a second encoding information. That is, the number of second encoding information is the same as the number of indicators included in the noise measurement indicator of the first candidate task.

[0171] Exemplarily, the target scheduling model comprises a second encoder configured to encode the noise measurement indicator of the first candidate task, that is, the second encoder in the target scheduling model is invoked to encode the noise measurement indicator of the first candidate task to obtain the second encoding information. The second encoder is any model capable of encoding, and the type of the second encoder is not limited in the embodiments of the present application and can be flexibly selected according to experience. Exemplarily, the second encoder can be a convolution model or an LSTM model. It should be noted that for the case that the noise measurement indicator of the first candidate task comprises multiple indicators, the number of second encoders can be multiple, and each second encoder is configured to encode one indicator in the noise measurement indicator of the first candidate task.

[0172] Step 3023: splicing the first encoded information and the second encoded information to obtain target encoded information; predicting the sampling probability of the first candidate task based on the target encoded information.

[0173] By splicing the first encoded information and the second encoded information to obtain the target encoded information, the target encoded information integrates the information of the current parameter update round and the noise measurement index of the first candidate task, and predicting the sampling probability of the first candidate task based on the target encoded information comprehensively considers the current parameter update round and the noise measurement index of the first candidate task, which is beneficial to guarantee the reliability of the sampling probability of the first candidate task.

[0174] Exemplarily, the target scheduling model includes a decoder for decrypting the target encoded information to obtain the sampling probability of the first candidate task. That is, the decoder in the target scheduling model is called to predict the sampling probability of the first candidate task based on the target encoded information. The type of the decoder can be set according to experience, and the embodiments of the present application do not limit this. Exemplarily, the decoder can be an MLP (Multilayer Perception).

[0175] Exemplarily, taking the noise measurement index of the first candidate task including the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task as an example, the process of calling the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task can be as shown in Figure 4 The current parameter update round is input into the first encoder in the target scheduling model for encoding to obtain the first encoded information; the test performance of the task processing model corresponding to the first candidate task is input into the first second encoder in the target scheduling model for encoding to obtain the first second encoded information; and the similarity between the first update gradient and the second update gradient is input into the second second encoder in the target scheduling model for encoding to obtain the second second encoded information. The first encoded information, the first second encoded information and the second second encoded information are spliced to obtain the target encoded information; and the target encoded information is input into the decoder in the target scheduling model for prediction to obtain the sampling probability of the first candidate task.

[0176] It should be noted that the process of calling the target scheduling model based on the noise measurement index of the first candidate task to obtain the sampling probability of the first candidate task based on steps 3021 to 3023 is only an example, and the embodiments of the present application are not limited thereto. For example, the process of calling the target scheduling model based on the noise measurement index of the first candidate task to obtain the sampling probability of the first candidate task can also be: calling the target scheduling model to encode the noise measurement index of the first candidate task to obtain second encoding information, and predicting the sampling probability of the first candidate task based on the second encoding information.

[0177] For example, the input of the target scheduling model includes the noise measurement index of the first candidate task, and the output is the sampling probability of the first candidate task, wherein the noise measurement index of the first candidate task is obtained based on the initial neural network model. For example, the processing process of the target scheduling model can be represented by the following formula 3:

[0178]

[0179] wherein w k represents the sampling probability of the kth first candidate task, g() represents the processing logic of the target scheduling model, represents the kth first candidate task, and θ0 represents the parameters of the initial neural network model.

[0180] For example, taking the noise measurement index of the first candidate task including the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task as an example, the processing process of the target scheduling model can be represented by the following formula 4:

[0181]

[0182] wherein, represents the test performance of the task processing model corresponding to the kth first candidate task; represents the first update gradient; represents the second update gradient; represents the similarity between the first update gradient and the second update gradient calculated by inner product; represents the training sample data in the training data corresponding to the kth first candidate task; represents the standard processing result in the training data corresponding to the kth first candidate task; represents the predicted processing result based on the initial neural network model and the training sample data in the training data corresponding to the kth first candidate task; represents the third loss; and θ0 represents the parameters of the initial neural network model.

[0183] denotes the test sample data in the test data corresponding to the kth first candidate task; denotes the standard processing result in the test data corresponding to the kth first candidate task; denotes the test processing result obtained based on the initial neural network model and the test sample data in the test data corresponding to the kth first candidate task; denotes the fourth loss; θ k denotes the parameter of the task processing model corresponding to the kth first candidate task; denotes the test processing result obtained based on the task processing model corresponding to the kth first candidate task and the test sample data in the test data corresponding to the kth first candidate task. For other parameters, refer to formula 3.

[0184] Before the noise measurement index based on the first candidate task is realized based on mode 1 to obtain the sampling probability of the first candidate task, a target scheduling model needs to be trained. The target scheduling model is obtained by training the initial scheduling model. The implementation mode of training the target scheduling model based on the initial scheduling model is not limited in the embodiments of the present application, as long as the trained target scheduling model has the function of predicting the sampling probability of the task according to the noise measurement index of the task. Exemplarily, the implementation mode of training the target scheduling model includes but is not limited to Figure 5 the implementation mode shown in Figure 7 the implementation mode shown in

[0185] Mode 2: mapping the noise measurement index of the first candidate task to the sampling probability of the first candidate task based on the mapping relationship between the noise measurement index and the sampling probability.

[0186] The mapping relationship between the noise measurement index and the sampling probability is used to indicate how to map the noise measurement index of the task to the sampling probability of the task. The mapping relationship between the noise measurement index and the sampling probability can be set by the technician. The form of the mapping relationship between the noise measurement index and the sampling probability is not limited in the embodiments of the present application, as long as it can indicate how to map the noise measurement index of the task to the sampling probability of the task. Exemplarily, the mapping relationship between the noise measurement index and the sampling probability can be a function relationship with the sampling probability as the dependent variable and the noise measurement index as the independent variable. Exemplarily, the mapping relationship between the noise measurement index and the sampling probability can also be a one-to-one correspondence between the noise measurement index and the sampling probability.

[0187] The mapping relationship between the noise measurement index and the sampling probability can indicate how to map the noise measurement index of the task to the sampling probability of the task. Therefore, the noise measurement index of the first candidate task is mapped to the sampling probability of the first candidate task according to the mapping relationship between the noise measurement index and the sampling probability. This way of obtaining the sampling probability of the first candidate task does not need to rely on the scheduling model, has a smaller amount of calculation, and is beneficial to improving the efficiency of obtaining the sampling probability of the first candidate task.

[0188] It should be noted that the number of the first candidate tasks is multiple, and each first candidate task corresponds to a noise measurement index. Based on the noise measurement index of each first candidate task, the sampling probability of each first candidate task can be obtained.

[0189] After obtaining the sampling probabilities of all the first candidate tasks, a task whose sampling probability satisfies a first selection condition can be selected from all the first candidate tasks, and the task whose sampling probability satisfies the first selection condition is taken as a reference task. The reference task is a task used to update the parameters of the initial neural network model. The task whose sampling probability satisfies the first selection condition is a task with a higher sampling probability. The sampling probability satisfying the first selection condition is set according to experience or is flexibly adjusted according to an application scenario, which is not limited in the embodiments of the present application.

[0190] In the example embodiments, the sampling probability satisfying the first selection condition can mean that the sampling probability is not less than a first probability threshold. The first probability threshold is set according to experience or is flexibly adjusted according to an application scenario, for example, the first probability threshold is 0.8, or the first probability threshold is 0.9, etc. In the example embodiments, the sampling probability satisfying the first selection condition can also mean that the sampling probability is the top L (L is an integer not less than 1) sampling probability among the sampling probabilities of all the first candidate tasks. The value of L can constrain the number of tasks for updating the initial neural network model. The value of L is set according to experience or is flexibly adjusted according to an application scenario, for example, the value of L is 4, or the value of L is 6, etc.

[0191] In step 303, the parameters of the initial neural network model are updated based on the reference task.

[0192] After the reference task is selected, the parameters of the initial neural network model are updated based on the reference task. Since the reference task is selected based on the sampling probability of the first candidate task, the sampling probability of the first candidate task is the probability that matches the interference degree of the first candidate task on updating the parameters of the initial neural network model, the selection reliability of the reference task is high, and updating the parameters of the initial neural network model based on the reference task is beneficial to guarantee the reliability of parameter updating, avoid the parameter updating process being biased by a task with large noise, thereby improving the generalization performance of the updated parameters, improving the robustness of the model training process based on meta-learning in the presence of a task with large noise, and improving the quality of model training. In some embodiments, the reference task for updating the parameters of the initial neural network model can also be referred to as a meta-training task.

[0193] In a possible implementation, the number of reference tasks is at least one, and any reference task corresponds to training data and test data; the process of updating the parameters of the initial neural network model based on the reference task includes the following steps 3031 to step 3033.

[0194] Step 3031: Based on the training data corresponding to any reference task and the initial neural network model, a task processing model corresponding to any reference task is obtained.

[0195] The task processing model corresponding to any reference task refers to a model for processing any reference task obtained by training the training data corresponding to any reference task based on the initial neural network model. The implementation principle of this step 3031 is the same as that of step 3011, and the difference lies in that the task in step 3011 is the first candidate task, and the task in step 3031 is any reference task.

[0196] Step 3032: Test the task processing model corresponding to any reference task by using the test data corresponding to any reference task, and obtain the test performance of the task processing model corresponding to any reference task.

[0197] The test performance of the task processing model corresponding to any reference task is used to measure the performance of the task processing model corresponding to any reference task in processing any reference task. The implementation principle of this step 3032 is the same as that of step 3012, and the difference lies in that the task in step 3012 is the first candidate task, and the task in step 3032 is any reference task.

[0198] Step 3033: Based on the test performance of the task processing model corresponding to at least one reference task, a target update gradient corresponding to the parameters of the initial neural network model is obtained, and the parameters of the initial neural network model are updated based on the target update gradient.

[0199] Referring to the processes of step 3031 and step 3032, the test performance of the task processing model corresponding to the at least one reference task can be obtained, and then the sum of the test performance of the task processing model corresponding to the at least one reference task can be taken as the target test performance. For example, since the target test performance is obtained based on the test performance of the task processing model corresponding to the at least one reference task, the test performance of the task processing model corresponding to the at least one reference task is obtained based on the parameters of the task processing model corresponding to the at least one reference task, and the task processing model corresponding to the at least one reference task is obtained based on the parameters of the initial neural network model, after the target test performance is obtained, the gradient of the target test performance with respect to the parameters of the initial neural network model can be calculated by using the chain rule, and the gradient is taken as the target update gradient corresponding to the parameters of the initial neural network model.

[0200] After the target update gradient corresponding to the parameters of the initial neural network model is calculated, the parameters of the initial neural network model are updated based on the target update gradient. For example, the parameters of the initial neural network model are updated based on the target update gradient in the following manner: a second product of a second learning rate and the target update gradient is calculated, and the difference between the parameters of the initial neural network model and the second product is taken as the updated parameters of the initial neural network model. The second learning rate is set according to experience or is flexibly adjusted according to the application scenario, and the embodiments of the present application do not limit this. This parameter updating process can be regarded as updating the parameters by taking the performance of the task processing model corresponding to the at least one reference task on the test data corresponding to the at least one reference task as feedback, and this parameter updating process comprehensively considers the at least one reference task, which is beneficial to guarantee the generalization of the updated parameters.

[0201] For example, taking the number of the at least one reference task as K (K is an integer not less than 1) for example, the process of updating the parameters of the initial neural network model based on the target update gradient can be implemented based on the following formula 5:

[0202]

[0203]

[0204] wherein θ0 represents the parameters of the initial neural network model; θ'0 represents the updated parameters of the initial neural network model; and β represents the second learning rate; represents the test sample data in the test data corresponding to the kth (k is an integer not less than 1 and not greater than K) reference task; represents the standard processing result in the test data corresponding to the kth reference task; and θ k represents the parameters of the task processing model corresponding to the kth reference task; ​a test processing result obtained based on the task processing model corresponding to the kth reference task and test sample data in the test data corresponding to the kth reference task; a test performance of the task processing model corresponding to the kth reference task; a target test performance; a target update gradient corresponding to a parameter of the initial neural network model.

[0205] Exemplarily, the task processing models corresponding to the K reference tasks have the same structure as the initial neural network model, and the difference lies in the parameters of the models. The parameters of the initial neural network model are θ0, the parameters of the task processing model corresponding to the first reference task are θ1, the parameters of the task processing model corresponding to the second reference task are θ2, and so on, and the parameters of the task processing model corresponding to the Kth reference task are θK. K . Wherein, θ1, θ2 and θKare obtained by gradient optimization with θ0as the initial value. The core of the model training based on meta-learning is to learn a parameter that can achieve good generalization performance on multiple tasks, and use the parameter as the initialization parameter of the target task to be learned. The initialization parameter is the parameter that the model needs to learn for the target task. K

[0206] In step 304, the target neural network model is obtained by learning the target task based on the neural network model with updated parameters.

[0207] After updating the parameters of the initial neural network model based on the reference tasks, the neural network model with updated parameters can be obtained. Then, the target neural network model suitable for the target task is obtained by learning the target task based on the neural network model with updated parameters. The target neural network model is used for data processing. Exemplarily, the target neural network model is used for processing data matching the target task.

[0208] Since the neural network model with updated parameters is obtained by updating the parameters of the initial neural network model using the reference tasks, the reference tasks are tasks with high reliability selected from the first candidate tasks according to the sampling probability. Therefore, the reliability of updating the parameters of the model using the reference tasks is high, which is conducive to improving the updating quality of the parameters of the model. The target task can be accurately learned based on the model with updated parameters, thereby improving the quality of the obtained target neural network model, and further improving the accuracy of data processing using the target neural network model.

[0209] ​In an example embodiment, the process of learning the target task based on the parameter-updated neural network model to obtain a target neural network model includes: obtaining a first neural network model based on the parameter-updated neural network model, training the first neural network model based on training data corresponding to the target task to obtain the target neural network model.

[0210] For example, the manner of obtaining the first neural network model based on the parameter-updated neural network model includes: in response to the parameter updating process satisfying a reference condition, taking the parameter-updated neural network model as the first neural network model; in response to the parameter updating process not satisfying the reference condition, continuing the parameter updating in the manner of steps 301 to 303 until the parameter updating process satisfies the reference condition, and taking the neural network model obtained when the parameter updating process satisfies the reference condition as the first neural network model. That is, the structure of the first neural network model is the same as that of the initial neural network model, and the parameters of the first neural network model are obtained by updating the parameters of the initial neural network model one or more times.

[0211] The reference condition satisfied by the parameter updating process is set according to experience or flexibly adjusted according to an application scenario, and embodiments of the present application do not limit this. For example, the reference condition satisfied by the parameter updating process can mean that the number of times of updating the parameters of the neural network model reaches a second number threshold, can also mean that the target test performance converges, and can also mean that the target test performance is less than a second loss threshold. The second number threshold and the second loss threshold are set according to experience or flexibly adjusted according to an application scenario, and embodiments of the present application do not limit this.

[0212] In an example embodiment, after obtaining the first neural network model, the first neural network model can be trained based on training data corresponding to the target task to obtain a target neural network model, so as to learn the target task. The target neural network model is a model suitable for the target task, and the model suitable for the target task can be used to process to-be-processed data matching the target task. For example, since the parameters of the first neural network model have good generalization ability on tasks of the same type as the first candidate task, in the process of training the first neural network model based on training data corresponding to the target task, the target neural network model can be obtained relatively quickly through fewer steps of training (or referred to as gradient optimization, parameter updating, etc.).

[0213] The implementation principle and steps of training the first neural network model based on the training data corresponding to the target task are the same as those of training the reference model based on the training data corresponding to the first candidate task in step 3011, and the difference lies in that the data relied on in step 3011 is the training data of the first candidate task, and the model relied on is the reference model, and the data relied on here is the training data corresponding to the target task, and the model relied on is the first neural network model.

[0214] After obtaining the target neural network model, the target neural network model can also be applied. The process of applying the target neural network model is detailed in Figure 8 The embodiments shown in the figure will not be described here.

[0215] The data processing method provided by the embodiments of the present application obtains the sampling probability of the candidate task by considering the noise measurement index of the candidate task, and the noise measurement index is used to measure the interference degree of the candidate task on the parameters of the updated model. The sampling probability obtained by considering the noise measurement index is a sampling probability matched with the interference degree of the candidate task on the parameters of the updated model. The reliability of the reference task selected from the candidate task based on such a sampling probability is higher, the reliability of updating the parameters of the model based on the reference task with higher reliability is higher, which is conducive to improving the updating quality of the parameters of the model. The model based on the parameter update can accurately learn the target task, thereby facilitating the improvement of the quality of the obtained target neural network model, and the target neural network model can be used to accurately process data matched with the target task.

[0216] Figure 5 A method for training a target scheduling model is shown, which can include the following steps 501 to 503.

[0217] In step 501, an initial scheduling model is called to obtain an initial sampling probability of a second candidate task based on a noise measurement index of the second candidate task; and an initial task whose initial sampling probability satisfies a second selection condition is selected from the second candidate task.

[0218] The initial scheduling model refers to the scheduling model to be trained, and the second candidate task is a task relied on for training the initial scheduling model. The second candidate task is of the same type as the first candidate task to ensure the training reliability of the scheduling model. The second candidate task can be the same as the first candidate task or different from the first candidate task. The implementation principle of calling the initial scheduling model to obtain the initial sampling probability of the second candidate task based on the noise measurement index of the second candidate task is the same as that of calling the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task in the embodiment shown in the figure, which will not be described here. Figure 3 The implementation principle of calling the initial scheduling model to obtain the initial sampling probability of the second candidate task based on the noise measurement index of the second candidate task is the same as that of calling the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task in the embodiment shown in the figure, which will not be described here.

[0219] The implementation principle of selecting an initial task from the second candidate task, whose initial sampling probability satisfies the second selection condition, is the same as that of selecting a reference task from the first candidate task, whose sampling probability satisfies the first selection condition, in the embodiment shown in FIG. 6, and details are not repeated here. It should be noted that the second selection condition can be the same as the first selection condition, or different from the first selection condition. Figure 3 The implementation principle of selecting a reference task from the first candidate task, whose sampling probability satisfies the first selection condition, in the embodiment shown in FIG. 6 is the same as that of selecting a reference task from the first candidate task, whose sampling probability satisfies the first selection condition, in the embodiment shown in FIG. 6, and details are not repeated here. It should be noted that the second selection condition can be the same as the first selection condition, or different from the first selection condition.

[0220] In step 502, based on the initial task and the initial neural network model, a to-be-verified model is obtained.

[0221] The to-be-verified model is a model obtained on the basis of the initial task and the initial neural network model, which is used to measure the verification performance of the verification task. In an exemplary embodiment, based on the initial task and the initial neural network model, the process of obtaining the to-be-verified model includes: constructing a candidate model with the same structure and parameters as the initial neural network model, updating the parameters of the candidate model based on the initial task, and taking the model with the updated parameters as the to-be-verified model.

[0222] Exemplarily, the process of updating the parameters of the candidate model based on the initial task includes: training the candidate model based on the training data corresponding to the initial task to obtain a task processing model corresponding to the initial task; testing the task processing model corresponding to the initial task by using the test data corresponding to the initial task to obtain the test performance of the task processing model corresponding to the initial task; obtaining an update gradient corresponding to the parameters of the candidate model based on the test performance of the task processing model corresponding to the initial task, and updating the parameters of the candidate model based on the update gradient corresponding to the parameters of the candidate model. The implementation principle of this implementation process is the same as that of the process of updating the parameters of the initial neural network model based on the reference task in the embodiment shown in FIG. 6, and details are not repeated here. Figure 3 The implementation principle of updating the parameters of the candidate model based on the initial task in the embodiment shown in FIG. 6 is the same as that of updating the parameters of the initial neural network model based on the reference task in the embodiment shown in FIG. 6, and details are not repeated here.

[0223] In step 503, the verification performance of the to-be-verified model is obtained based on the verification task, the third update gradient corresponding to the parameters of the initial scheduling model is obtained based on the verification performance, the parameters of the initial scheduling model are updated based on the third update gradient to obtain a target scheduling model.

[0224] The verification task is a task used to verify the to-be-verified model, and the type of the verification task is the same as that of the second candidate task. The number of verification tasks can be one or multiple. The verification performance of the to-be-verified model can be used to measure the effect of parameter updating of the initial task selected according to the initial scheduling model, and updating the initial scheduling model based on the verification performance of the to-be-verified model can update the parameters of the scheduling model in the direction of improving the verification performance, thereby indirectly improving the prediction accuracy of the scheduling model on the sampling probability by improving the verification performance of the model obtained according to the selected task.

[0225] In an exemplary embodiment, the verification task corresponds to training data and test data. The process of obtaining the verification performance of the model to be verified based on the verification task includes: training the model to be verified based on the training data corresponding to the verification task to obtain the task processing model corresponding to the verification task; testing the task processing model corresponding to the verification task using the test data corresponding to the verification task to obtain the test performance of the task processing model corresponding to the verification task; and obtaining the verification performance of the model to be verified based on the test performance of the task processing model corresponding to the verification task. The implementation principle of obtaining the test performance of the task processing model corresponding to the verification task is similar to... Figure 3 The implementation principle for obtaining the test performance of the task processing model corresponding to at least one reference task in the illustrated embodiment is the same, and will not be repeated here. After obtaining the test performance of the task processing model corresponding to the verification task, the sum of the test performance of the task processing models corresponding to the verification tasks is taken as the verification performance of the model to be verified.

[0226] For example, since the verification performance of the model to be verified is obtained based on the test performance of the task processing model corresponding to the verification task, the test performance of the task processing model corresponding to the verification task is obtained based on the parameters of the task processing model corresponding to the verification task, and the task processing model corresponding to the verification task is obtained based on the parameters of the initial scheduling model, after obtaining the verification performance of the model to be verified, the gradient of the verification performance of the model to be verified with respect to the parameters of the initial scheduling model can be calculated by using the chain rule, and this gradient can be used as the third update gradient corresponding to the parameters of the initial scheduling model.

[0227] After calculating the third update gradient corresponding to the parameters of the initial scheduling model, the parameters of the initial scheduling model are updated based on the third update gradient. For example, the method for updating the parameters of the initial scheduling model based on the third update gradient is as follows: calculate the third product of the third learning rate and the third update gradient, and use the difference between the parameters of the initial scheduling model and the third product as the updated parameters of the initial scheduling model. The third learning rate is set empirically or flexibly adjusted according to the application scenario; this embodiment does not limit this.

[0228] For example, the number of verification tasks is K. v (K v Taking (a number of integers not less than 1) as an example, the process of updating the parameters of the initial scheduling model based on the third update gradient can be implemented based on Equation 6:

[0229]

[0230] Where φ represents the parameters of the initial scheduling model; φ′ represents the updated parameters of the initial scheduling model; and γ represents the third learning rate. It represents the k-th (k is not less than 1 and not greater than K)v a test sample data in the test data corresponding to the kth verification task; denotes a standard processing result in the test data corresponding to the kth verification task; θ k denotes a parameter of the task processing model corresponding to the kth verification task; denotes a test processing result obtained based on the task processing model corresponding to the kth verification task and the test sample data in the test data corresponding to the kth verification task; denotes a test performance of the task processing model corresponding to the kth verification task; denotes a verification performance of the to-be-verified model; denotes a third update gradient corresponding to the parameter of the initial scheduling model.

[0231] Exemplarily, updating the parameter of the initial scheduling model based on the third update gradient to obtain the target scheduling model can mean that the model obtained after updating the parameter of the initial scheduling model based on the third update gradient is taken as the target scheduling model; or can mean that after updating the parameter of the initial scheduling model based on the third update gradient, the parameter of the scheduling model is continuously updated in the manner of steps 501 to 503 until the updating process of the parameter of the scheduling model satisfies the second condition, and the model with the parameter obtained when the second condition is satisfied is taken as the target scheduling model. Exemplarily, the second condition can mean that the number of times of updating the parameter of the scheduling model reaches a third number threshold, can mean that the verification performance converges, or can mean that the verification performance is not less than a third loss threshold. The third number threshold and the third loss threshold are set according to experience or are flexibly adjusted according to an application scenario.

[0232] Exemplarily, the to-be-verified model can be regarded as a temporary time sequence meta-model, which can avoid introducing an ultra-long optimization link and a large optimization cost and reduce optimization difficulty.

[0233] Exemplarily, the training process of the scheduling model can be cross-performed with the updating process of the parameter of the neural network model, and the cross-performed process includes: training the scheduling model based on the neural network model to obtain a training-once scheduling model, updating the parameter of the neural network model by using a task selected according to the training-once scheduling model to obtain a parameter-updated-once neural network model; training the training-once scheduling model based on the parameter-updated-once neural network model to obtain a training-twice scheduling model, updating the parameter of the parameter-updated-once neural network model by using a task selected according to the training-twice scheduling model to obtain a parameter-updated-twice neural network model. This is iterated until a parameter-updated-G(G is an integer not less than 1) times neural network model is obtained. The value of G is set according to experience or is flexibly adjusted according to an application scenario, which is not limited in the embodiments of the present application.

[0234] In the case that the training process of the scheduling model can be cross-performed with the updating process of the parameters of the neural network model, Figure 3 and Figure 5 The initial neural network model in the illustrated embodiment can refer to the neural network model after the parameter updating l (l is an integer not less than 0 and not greater than (G-1)) times, the initial scheduling model can refer to the scheduling model trained l times, and the target scheduling model can refer to the scheduling model trained (l+1) times.

[0235] Exemplarily, the process of implementing the parameter updating according to the scheduling model can be as shown in Figure 6 The second candidate task is obtained from the task pool, the noise measurement index of the second candidate task is obtained based on the initial neural network model, the initial scheduling model is called to obtain the sampling probability of the second candidate task based on the noise measurement index of the second candidate task, the initial task is selected through task sampling according to the sampling probability of the second candidate task, the test performance of the task processing model corresponding to the selected initial task is calculated based on the initial task and the initial neural network model, the to-be-verified model is obtained based on the feedback of the test performance of the task processing model corresponding to the selected initial task, and the target scheduling model is obtained based on the feedback of the verification performance of the to-be-verified model.

[0236] The first candidate task is obtained from the task pool, the noise measurement index of the first candidate task is obtained based on the initial neural network model, the target scheduling model is called to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task, the reference task is selected through task sampling according to the sampling probability of the first candidate task, the test performance of the task processing model corresponding to the selected reference task is calculated based on the reference task and the initial neural network model, and the parameters of the initial neural network model are updated based on the test performance of the task processing model corresponding to the selected reference task.

[0237] Figure 7 Another method for training the target scheduling model is shown, which can include the following steps 701 to 703.

[0238] In step 701, the sample noise measurement index of the sample task and the standard sampling probability of the sample task are obtained.

[0239] The sample noise measurement index of the sample task is used to measure the interference degree of the sample task on updating the parameters of the initial neural network model, and the principle of obtaining the sample noise measurement index of the sample task is the same as that of obtaining the noise measurement index of the first candidate task, which will not be described here. Exemplarily, the sample task and the first candidate task are of the same type to ensure the training reliability of the scheduling model.

[0240] The standard sampling probability of the sample task can be considered as the real sampling probability of the sample task, and the standard sampling probability of the sample task is used to provide a supervision signal for the process of predicting the sampling probability by the scheduling model. For example, the standard sampling probability of the sample task can be obtained by analyzing the sample noise measurement index of the sample task by the technical personnel. For example, the number of sample tasks is multiple, so as to ensure the training quality of the scheduling model. Each sample task corresponds to a sample noise measurement index and a standard sampling probability.

[0241] In step 702, the initial scheduling model is called to obtain an initial sampling probability of the sample task based on the sample noise measurement index; and a target loss is obtained based on the initial sampling probability and the standard sampling probability of the sample task.

[0242] The implementation principle of calling the initial scheduling model to obtain the initial sampling probability of the sample task based on the sample noise measurement index is the same as that of calling the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task, which will not be described here. Figure 3 The implementation principle of calling the initial scheduling model to obtain the initial sampling probability of the sample task based on the sample noise measurement index is the same as that of calling the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task, which will not be described here.

[0243] The target loss is used to measure the difference between the initial sampling probability and the standard sampling probability of the sample task. The target loss can be calculated by substituting the initial sampling probability and the standard sampling probability of the sample task into a target loss function. The type of the target loss function used in the embodiments of the present application is not limited, and can be set according to experience or flexibly adjusted according to the application scenario. For example, the target loss function can be a cross-entropy loss function, or a mean square error loss function, etc. It should be noted that the number of sample tasks is one or more, and the target loss is obtained based on all sample tasks.

[0244] In step 703, a fourth update gradient corresponding to the parameters of the initial scheduling model is obtained based on the target loss; and the parameters of the initial scheduling model are updated based on the fourth update gradient to obtain the target scheduling model.

[0245] Since the target loss is a loss obtained by forward propagation based on the parameters of the initial scheduling model, after obtaining the target loss, the gradient of the target loss with respect to the parameters of the initial scheduling model can be calculated by the chain rule, and the gradient is taken as the fourth update gradient corresponding to the parameters of the initial scheduling model.

[0246] After the fourth update gradient corresponding to the parameters of the initial scheduling model is calculated, the parameters of the initial scheduling model are updated based on the fourth update gradient. For example, the parameters of the initial scheduling model are updated based on the fourth update gradient in the following manner: a fourth product of a fourth learning rate and the fourth update gradient is calculated, and a difference between the parameters of the initial scheduling model and the fourth product is taken as the updated parameters of the initial scheduling model. The fourth learning rate is set according to experience or is flexibly adjusted according to an application scenario, which is not limited in the embodiments of the present application.

[0247] For example, updating the parameters of the initial scheduling model based on the fourth update gradient to obtain the target scheduling model can mean that the model obtained after updating the parameters of the initial scheduling model based on the fourth update gradient is taken as the target scheduling model; or can mean that after updating the parameters of the initial scheduling model based on the fourth update gradient, the parameters of the scheduling model are continuously updated in the manner of steps 701 to 703 until the updating process of the parameters of the scheduling model meets the third condition, and the model with the parameters obtained when the third condition is met is taken as the target scheduling model.

[0248] For example, the third condition can mean that the number of times of updating the parameters of the scheduling model reaches a fourth number threshold, can mean that the target loss converges, or can mean that the target loss is not less than a fourth loss threshold. The fourth number threshold and the fourth loss threshold are set according to experience or are flexibly adjusted according to an application scenario.

[0249] In the training process of the scheduling model, the standard sampling probability corresponding to the sample task is directly used to provide a supervision signal for the process of predicting the sampling probability by the scheduling model, which can update the parameters of the scheduling model in a manner of reducing the difference between the predicted sampling probability and the standard sampling probability, thereby intuitively improving the prediction accuracy of the sampling probability by the scheduling model.

[0250] The embodiments of the present application provide a data processing method. The method is applied to the implementation environment shown in the above Figure 1 For example, the data method is executed by a computer device, which can be the terminal 11 in the implementation environment shown in the above Figure 1 , or can be the server 12 in the implementation environment shown in the above Figure 1 , which is not limited in the embodiments of the present application. As shown in the above Figure 8 , the data method provided by the embodiments of the present application can include the following steps 801 and 802.

[0251] In step 801, the data to be processed matched with the target task and the target neural network model are obtained.

[0252] The target neural network model is obtained by the data processing method provided by the embodiments shown in the above Figure 3 .

[0253] The to-be-processed data matched with the target task refers to data to be processed by using the target neural network model. Since the target neural network model is obtained by accurately learning the target task, the target neural network model can accurately process the to-be-processed data matched with the target task, thereby ensuring the accuracy of data processing. Figure 3 The data processing method provided by the embodiment shown in the figure obtains the target neural network model by accurately learning the target task, so that the target neural network model can accurately process the to-be-processed data matched with the target task, thereby ensuring the accuracy of data processing.

[0254] The type of to-be-processed data is related to the type of the target task, and the embodiment of the present application does not limit it. Illustratively, the target task refers to a task of predicting a target attribute of an object of a reference type, and the to-be-processed data refers to representation information of a target object, wherein the type of the target object is the reference type. The representation information of the target object is used to represent the target object.

[0255] Next, the specific case of the representation information of the target object is illustrated in combination with the specific type of the task of predicting the target attribute of the object of the reference type. Illustratively, if the task of predicting the target attribute of the object of the reference type is a task of predicting the activity of a target compound on a target target protein, the representation information of the target object refers to the representation information of the target compound, and in this case, the to-be-processed data refers to the representation information of the target compound. If the task of predicting the target attribute of the object of the reference type is a task of predicting whether an image belongs to a target category, the representation information of the target object refers to the representation information of the target image, and in this case, the to-be-processed data refers to the representation information of the target image.

[0256] The target compound refers to a compound whose activity on a target target protein is unknown. The representation information of the target compound is used to represent the target compound, and illustratively, the representation information of the target compound can refer to the molecular formula of the target compound, or the molecular string of the target compound, or the Morgan molecular fingerprint feature of the target compound, etc. Illustratively, the form of the representation information of the target compound can be a vector or a matrix, and the dimension of the vector or the matrix is set according to experience or flexibly adjusted according to the application scenario, for example, the dimension is 1024. The target image refers to an image whose belonging to a target category is unknown. The representation information of the target image is used to represent the target image, and illustratively, the representation information of the target image can refer to the pixel value of a pixel point in the target image, or refer to the image feature of the target image, etc.

[0257] The embodiment of the present application does not limit the way of obtaining the to-be-processed data matched with the target task, and illustratively, the to-be-processed data matched with the target task can be extracted from a database, or extracted from a network, or received from a technician uploading the to-be-processed data matched with the target task, etc.

[0258] Exemplarily, the manner of obtaining the target neural network model can refer to obtaining the target neural network model in real time according to the method provided by the embodiment shown in the figure. Figure 3 The method provided by the embodiment shown in the figure can obtain the target neural network model in real time, and can also refer to extracting the target neural network model obtained in advance according to the method provided by the embodiment shown in the figure from storage, etc. The embodiment of the present application does not limit this. Figure 3 The method provided by the embodiment shown in the figure can obtain the target neural network model in real time, and can also refer to extracting the target neural network model obtained in advance according to the method provided by the embodiment shown in the figure from storage, etc. The embodiment of the present application does not limit this.

[0259] In step 802, the target neural network model is called to process the to-be-processed data to obtain a processing result corresponding to the to-be-processed data.

[0260] The process of calling the target neural network model to process the to-be-processed data to obtain a processing result corresponding to the to-be-processed data is an internal processing process of the target neural network model, which is related to the structure of the target neural network model, and the embodiment of the present application does not limit this. Exemplarily, the process of calling the target neural network model to process the to-be-processed data to obtain a processing result corresponding to the to-be-processed data can be: calling the target neural network model to extract features of the to-be-processed data, converting the extracted features to obtain a processing result corresponding to the to-be-processed data.

[0261] Exemplarily, the type of the processing result corresponding to the to-be-processed data is related to the type of the to-be-processed data. Exemplarily, the to-be-processed data is representation information of a target object, and the processing result corresponding to the to-be-processed data is a target attribute of the target object. The process of calling the target neural network model to process the to-be-processed data to obtain a processing result corresponding to the to-be-processed data includes: calling the target neural network model to process the representation information of the target object to obtain the target attribute of the target object.

[0262] Exemplarily, the representation information of the target object refers to representation information of a target compound, and the target attribute of the target object refers to a target activity of the target compound to a target target protein. In this case, the process of calling the target neural network model to process the to-be-processed data to obtain a processing result corresponding to the to-be-processed data includes: calling the target neural network model to process the representation information of the target compound to obtain the target activity of the target compound to the target target protein.

[0263] Exemplarily, the representation information of the target object refers to representation information of a target compound, and the target attribute of the target object refers to a target activity of the target compound to a target target protein. In this case, the process of calling the target neural network model to process the to-be-processed data to obtain a processing result corresponding to the to-be-processed data includes: calling the target neural network model to process the representation information of the target compound to obtain the target activity of the target compound to the target target protein.

[0264] In an example embodiment, after obtaining the processing result corresponding to the to-be-processed data, the processing result corresponding to the to-be-processed data can be applied. The manner of applying the processing result corresponding to the to-be-processed data is related to the type of the to-be-processed data and the processing result corresponding to the to-be-processed data.

[0265] For example, if the to-be-processed data is characterization information of a target compound, and the processing result corresponding to the to-be-processed data is a target activity of the target compound on a target target protein, the manner of applying the processing result corresponding to the to-be-processed data includes: determining a lead compound of the target target protein from the target compound based on the target activity of the target compound on the target target protein.

[0266] For example, a compound in the target compound with a target activity on the target target protein not less than an activity threshold is a lead compound of the target target protein. The activity threshold is set according to experience or flexibly adjusted according to an application scenario. For example, Q (Q is an integer not less than 1) compounds in the target compound with the largest target activity on the target target protein are lead compounds of the target target protein. The value of Q is set according to experience or flexibly adjusted according to an application scenario.

[0267] The above-mentioned application of determining the lead compound of the target target protein is a specific application in the application scenario of virtually screening the lead compound. The application of the data processing method provided in the embodiments of the present application in the application scenario of virtually screening the lead compound brings technical influence mainly reflected in improving the accuracy of drug virtual screening based on molecular structure. By reducing the sensitivity of the meta-learning framework to a considerable number of noisy measured activity experiments, improving the robustness of the model training process based on meta-learning in the case of a larger amount of noise, and improving the generalization ability of the parameters of the neural network model, the ideal activity prediction accuracy can be achieved on various diversified target proteins that may appear in various business scenarios, ensuring the quality of the virtually screened compounds, and discovering more optimal and accurate lead compounds, laying an important foundation for subsequent optimization of lead compounds and drug synthesis.

[0268] In an example embodiment, if the to-be-processed data is the representation information of the target image, the processing result corresponding to the to-be-processed data is the target classification result of the target image, and the manner of applying the processing result corresponding to the to-be-processed data includes: determining whether the target image belongs to a target category based on the target classification result of the target image. For example, the target classification result of the target image can be 0 or 1. If the target classification result is 0, it means that the target image does not belong to the target category. If the target classification result is 1, it means that the target image belongs to the target category. For example, the target classification result of the target image can also be a value between 0 and 1. If the value is greater than a classification threshold, it means that the target image belongs to the target category. If the value is not greater than the classification threshold, it means that the target image does not belong to the target category. The classification threshold is set according to experience or flexibly adjusted according to the application scenario. For example, the classification threshold can be 0.85, or 0.9, etc.

[0269] The target neural network model for processing the to-be-processed data in the data processing method provided by the embodiments of the present application is obtained by the method provided by the embodiments shown in Figure 3 In the method provided by the embodiments shown in Figure 3 In the method provided by the embodiments shown in the method, the sampling probability of the candidate task is obtained by considering the noise measurement index of the candidate task. The noise measurement index is used to measure the interference degree of the candidate task on the parameters of the updated model. The sampling probability obtained by considering the noise measurement index is a sampling probability matched with the interference degree of the candidate task on the parameters of the updated model. The reliability of the reference task selected from the candidate task based on this sampling probability is higher. The reliability of the parameters of the model updated based on the reference task with higher reliability is higher, which is conducive to improving the update quality of the parameters of the model. The model updated based on the parameters can learn the target task more accurately, thereby being conducive to improving the quality of the obtained target neural network model. The to-be-processed data can be processed more accurately by using such a target neural network model, and the accuracy of data processing is higher.

[0270] Referring to Figure 9 The embodiments of the present application provide a data processing device, which comprises:

[0271] The first obtaining unit 901 is configured to obtain a noise measurement index of a first candidate task based on an initial neural network model. The noise measurement index is used to measure the interference degree of the first candidate task on the parameters of the initial neural network model. The first candidate task is a task matched with a target task to be learned.

[0272] The second obtaining unit 902 is configured to obtain a sampling probability of the first candidate task based on the noise measurement index of the first candidate task, and select a reference task from the first candidate task, wherein the sampling probability of the reference task satisfies a first selection condition.

[0273] The updating unit 903 is configured to update parameters of the initial neural network model based on the reference task.

[0274] The third obtaining unit 904 is configured to learn the target task based on the neural network model after the parameter updating, to obtain a target neural network model, where the target neural network model is used for data processing.

[0275] In a possible implementation, the first candidate task corresponds to training data and test data; the first obtaining unit 901 is configured to obtain a task processing model corresponding to the first candidate task based on the training data corresponding to the first candidate task and the initial neural network model; test the task processing model corresponding to the first candidate task by using the test data corresponding to the first candidate task, to obtain a test performance of the task processing model corresponding to the first candidate task; obtain a first update gradient corresponding to the parameters of the initial neural network model based on the training data corresponding to the first candidate task; obtain a second update gradient corresponding to the parameters of the initial neural network model based on the test data corresponding to the first candidate task; and obtain a noise measurement index of the first candidate task based on similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task.

[0276] In a possible implementation, the first candidate task corresponds to training data and test data; the first obtaining unit 901 is configured to obtain a task processing model corresponding to the first candidate task based on the training data corresponding to the first candidate task and the initial neural network model; test the task processing model corresponding to the first candidate task by using the test data corresponding to the first candidate task, to obtain a test performance of the task processing model corresponding to the first candidate task; and obtain a second-order update gradient corresponding to the parameters of the initial neural network model based on the test performance of the task processing model corresponding to the first candidate task, and take the second-order update gradient as the noise measurement index of the first candidate task.

[0277] In a possible implementation, the second obtaining unit 902 is configured to call the target scheduling model to obtain a sampling probability of the first candidate task based on the noise measurement index of the first candidate task.

[0278] In a possible implementation, the second obtaining unit 902 is configured to encode a current parameter update round by using the target scheduling model, to obtain first encoding information; encode the noise measurement index of the first candidate task, to obtain second encoding information; splice the first encoding information and the second encoding information, to obtain target encoding information; and predict the sampling probability of the first candidate task based on the target encoding information.

[0279] In a possible implementation, the second obtaining unit 902 is further configured to invoke the initial scheduling model to obtain an initial sampling probability of the second candidate task based on the noise measurement index of the second candidate task; and select an initial task from the second candidate task, wherein the initial sampling probability of the initial task satisfies a second selection condition.

[0280] The updating unit 903 is further configured to obtain a to-be-verified model based on the initial task and the initial neural network model; obtain a verification performance of the to-be-verified model based on the verification task; obtain a third update gradient corresponding to a parameter of the initial scheduling model based on the verification performance; and update the parameter of the initial scheduling model based on the third update gradient to obtain a target scheduling model.

[0281] In a possible implementation, the verification task corresponds to training data and test data; and the updating unit 903 is further configured to train the to-be-verified model based on the training data corresponding to the verification task to obtain a task processing model corresponding to the verification task; test the task processing model corresponding to the verification task by using the test data corresponding to the verification task to obtain a test performance of the task processing model corresponding to the verification task; and obtain the verification performance of the to-be-verified model based on the test performance of the task processing model corresponding to the verification task.

[0282] In a possible implementation, the second obtaining unit 902 is further configured to obtain a sample noise measurement index of a sample task and a standard sampling probability of the sample task; and invoke the initial scheduling model to obtain an initial sampling probability of the sample task based on the sample noise measurement index.

[0283] The updating unit 903 is further configured to obtain a target loss based on the initial sampling probability and the standard sampling probability of the sample task; obtain a fourth update gradient corresponding to a parameter of the initial scheduling model based on the target loss; and update the parameter of the initial scheduling model based on the fourth update gradient to obtain the target scheduling model.

[0284] In a possible implementation, the second obtaining unit 902 is configured to map the noise measurement index of the first candidate task to a sampling probability of the first candidate task based on a mapping relationship between a noise measurement index and a sampling probability.

[0285] In a possible implementation, the number of reference tasks is at least one, and any reference task corresponds to training data and test data; the updating unit 903 is configured to obtain a task processing model corresponding to any reference task based on the training data corresponding to the reference task and the initial neural network model; test the task processing model corresponding to any reference task by using the test data corresponding to the reference task, to obtain a test performance of the task processing model corresponding to any reference task; and obtain a target update gradient corresponding to the parameters of the initial neural network model based on the test performances of the task processing models corresponding to the at least one reference task, and update the parameters of the initial neural network model based on the target update gradient.

[0286] In a possible implementation, the training data corresponding to the first candidate task includes training sample data and a standard processing result corresponding to the training sample data; the first obtaining unit 901 is configured to call a reference model to process the training sample data, to obtain a predicted processing result corresponding to the training sample data, the parameters of the reference model being the same as the parameters of the initial neural network model; obtain a first loss based on the predicted processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data; and update the parameters of the reference model based on the first loss, to obtain the task processing model corresponding to the first candidate task.

[0287] In a possible implementation, the first candidate task is a task of predicting the activity of a compound on a first candidate target protein, and the training sample data is representation information of the first compound; the predicted processing result is a predicted activity of the first compound on the first candidate target protein; the first obtaining unit 902 is configured to call a reference model to process the representation information of the first compound, to obtain the predicted activity of the first compound on the first candidate target protein; and obtain a first loss based on the predicted activity of the first compound on the first candidate target protein and a standard activity of the first compound on the first candidate target protein.

[0288] The data processing apparatus provided in the embodiments of the present application obtains the sampling probability of the candidate task by considering the noise measurement index of the candidate task, the noise measurement index is used to measure the interference degree of the candidate task on the parameters of the updated model, the sampling probability obtained by considering the noise measurement index is the sampling probability matched with the interference degree of the candidate task on the parameters of the updated model, the reliability of the reference task selected from the candidate task based on this sampling probability is higher, the reliability of the parameters of the model updated based on the reference task with higher reliability is higher, which is beneficial to improving the update quality of the parameters of the model, the model based on the parameter update can learn the target task more accurately, thereby being beneficial to improving the quality of the obtained target neural network model, and the target neural network model can be used to process the data matched with the target task more accurately.

[0289] Referring toFigure 10 The embodiment of the present application provides a data processing device, which comprises:

[0290] The fourth acquisition unit 1001 is configured to acquire to-be-processed data matched with a target task and a target neural network model, and the target neural network model is obtained through the method for processing data shown in the embodiment. Figure 3 The fourth acquisition unit 1001 is configured to acquire to-be-processed data matched with a target task and a target neural network model, and the target neural network model is obtained through the method for processing data shown in the embodiment.

[0291] The processing unit 1002 is configured to call the target neural network model to process the to-be-processed data, and obtain a processing result corresponding to the to-be-processed data.

[0292] In a possible implementation manner, the target task is a task of predicting an activity of a target compound on a target protein, the to-be-processed data is characterization information of the target compound, and the processing result corresponding to the to-be-processed data is a target activity of the target compound on the target protein; the processing unit 1002 is configured to call the target neural network model to process the characterization information of the target compound, and obtain the target activity of the target compound on the target protein.

[0293] In a possible implementation manner, the device further comprises:

[0294] The determining unit is configured to determine a lead compound of the target protein from the target compound based on the target activity of the target compound on the target protein.

[0295] The data processing device provided by the embodiment of the present application is used for processing to-be-processed data, and the target neural network model is obtained through the method provided in the embodiment. Figure 3 The method provided in the embodiment is used for obtaining the target neural network model. Figure 3 In the method provided in the embodiment, the sampling probability of the candidate task is obtained by considering the noise measurement index of the candidate task, the noise measurement index is used for measuring the interference degree of the candidate task on the parameters of the updated model, the sampling probability obtained by considering the noise measurement index is a sampling probability matched with the interference degree of the candidate task on the parameters of the updated model, the reliability of the reference task selected from the candidate task based on the sampling probability is relatively high, the reliability of the parameters of the updated model based on the reference task with relatively high reliability is relatively high, which is beneficial to improving the updating quality of the parameters of the model, the model based on the parameter updating can learn the target task more accurately, thereby being beneficial to improving the quality of the obtained target neural network model, and the target neural network model can be used to process the to-be-processed data more accurately, and the accuracy of the data processing is relatively high.

[0296] It should be noted that the apparatus provided in the above examples is only exemplified by the above division of each functional unit when realizing its function, and in actual application, the above functions can be completed by different functional units according to needs, that is, the internal structure of the device is divided into different functional units to complete all or part of the above described functions. In addition, the apparatus and method embodiments provided in the above examples belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be described here.

[0297] In an example embodiment, a computer device is also provided, which includes a processor and a memory having at least one computer program stored therein. The at least one computer program is loaded and executed by one or more processors to enable the computer device to implement any of the above data processing methods. The computer device can be a server or a terminal. Next, the structure of the server and the terminal will be introduced respectively.

[0298] Figure 11 is a structural schematic diagram of a server provided by an embodiment of the present application. The server can have great differences due to different configurations or performances, and can include one or more processors (Central Processing Units, CPU) 1101 and one or more memories 1102. The one or more memories 1102 store at least one computer program, which is loaded and executed by the one or more processors 1101 to enable the server to implement the data processing method provided by each method embodiment. Of course, the server can also have a wired or wireless network interface, a keyboard, an input and output interface, and other components for realizing the functions of the device, so as to perform input and output. The server can also include other components for realizing the functions of the device, which will not be described here.

[0299] Figure 12 is a structural schematic diagram of a terminal provided by an embodiment of the present application. The terminal can be a PC, a mobile phone, a smart phone, a PDA, a wearable device, a PPC, a tablet computer, a smart car machine, a smart television, a smart sound box, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, a VR device, and an AR device. The terminal can also be referred to as a user device, a portable terminal, a laptop terminal, a desktop terminal, and other names.

[0300] Generally, the terminal includes a processor 1501 and a memory 1502.

[0301] The processor 1501 can include one or more processing cores, such as a 4-core processor, an 8-core processor, and the like. The processor 1501 can be implemented in at least one of a hardware form of a DSP (Digital Signal Processing), an FPGA (Field-Programmable Gate Array), a PLA (Programmable Logic Array). The processor 1501 can also include a main processor and a coprocessor, the main processor being a processor for processing data in an awake state, also known as a CPU (Central Processing Unit), and the coprocessor being a low-power processor for processing data in a standby state. In some embodiments, the processor 1501 can be integrated with a GPU (Graphics Processing Unit) for rendering and drawing content required to be displayed by the display screen. In some embodiments, the processor 1501 can also include an AI (Artificial Intelligence) processor for processing computing operations related to machine learning.

[0302] The memory 1502 can include one or more computer-readable storage media that can be non-transitory. The memory 1502 can also include a high-speed random access memory, and a nonvolatile memory such as one or more disk storage devices, flash storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 1502 is used to store at least one instruction for being executed by the processor 1501 to enable the terminal to implement the data processing method provided by the method embodiment of the present application.

[0303] In some embodiments, the terminal can also optionally include a peripheral device interface 1503 and at least one peripheral device. The processor 1501, the memory 1502, and the peripheral device interface 1503 can be connected through a bus or a signal line. Each peripheral device can be connected to the peripheral device interface 1503 through a bus, a signal line, or a circuit board. Specifically, the peripheral device includes at least one of a radio frequency circuit 1504, a display screen 1505, a camera assembly 1506, an audio circuit 1507, and a power supply 1508.

[0304] The peripheral interface 1503 can be used to connect at least one I / O (Input / Output) related peripheral device to the processor 1501 and the memory 1502. In some embodiments, the processor 1501, the memory 1502 and the peripheral interface 1503 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1501, the memory 1502 and the peripheral interface 1503 can be implemented on a separate chip or circuit board, and the present embodiments are not limited in this regard.

[0305] The radio frequency circuit 1504 is configured to receive and send RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 1504 communicates with communication networks and other communication devices through electromagnetic signals. The radio frequency circuit 1504 converts electrical signals into electromagnetic signals for transmission, or converts electromagnetic signals received into electrical signals. Optionally, the radio frequency circuit 1504 includes an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a subscriber identity module card, and the like. The radio frequency circuit 1504 can communicate with other terminals through at least one wireless communication protocol. The wireless communication protocol includes but is not limited to: a metropolitan area network, various generations of mobile communication networks (2G, 3G, 4G and 5G), a wireless local area network and / or a WiFi (Wireless Fidelity) network. In some embodiments, the radio frequency circuit 1504 can also include NFC (Near Field Communication) related circuitry, and the present application is not limited in this regard.

[0306] The display screen 1505 is configured to display a UI (User Interface). The UI can include graphics, text, icons, video, and any combination thereof. When the display screen 1505 is a touch display screen, the display screen 1505 is further configured to capture touch signals on or above the surface of the display screen 1505. The touch signals can be input to the processor 1501 as control signals for processing. In this case, the display screen 1505 can also be configured to provide virtual buttons and / or virtual keyboard, also known as soft buttons and / or soft keyboard. In some embodiments, the display screen 1505 can be one, disposed on the front panel of the terminal; in other embodiments, the display screen 1505 can be at least two, respectively disposed on different surfaces of the terminal or in a folding design; in other embodiments, the display screen 1505 can be a flexible display screen, disposed on a curved surface or a folding surface of the terminal. Even, the display screen 1505 can also be disposed in an irregular shape, i.e., a special-shaped screen. The display screen 1505 can be made of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), etc.

[0307] The camera assembly 1506 is configured to capture images or videos. Optionally, the camera assembly 1506 includes a front camera and a rear camera. Typically, the front camera is disposed on the front panel of the terminal, and the rear camera is disposed on the back of the terminal. In some embodiments, the rear camera is at least two, which are any one of a main camera, a depth-of-field camera, a wide-angle camera, and a telephoto camera, to realize the background blur function by fusing the main camera and the depth-of-field camera, the panoramic shooting and VR (Virtual Reality) shooting function by fusing the main camera and the wide-angle camera, or other fusion shooting functions. In some embodiments, the camera assembly 1506 can further include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. The dual-color temperature flash refers to the combination of a warm light flash and a cold light flash, which can be used for light compensation under different color temperatures.

[0308] The audio circuit 1507 can include a microphone and a speaker. The microphone is used to collect sound waves of the user and the environment, and convert the sound waves into an electrical signal input to the processor 1501 for processing, or input to the radio frequency circuit 1504 to realize voice communication. For the purpose of stereo sound collection or noise reduction, the microphone can be multiple, respectively arranged at different parts of the terminal. The microphone can also be an array microphone or an omnidirectional collection type microphone. The speaker is used to convert the electrical signal from the processor 1501 or the radio frequency circuit 1504 into sound waves. The speaker can be a traditional diaphragm speaker, or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, not only can the electrical signal be converted into a sound wave that humans can hear, but also can be converted into a sound wave that humans cannot hear for ranging purposes. In some embodiments, the audio circuit 1507 can also include a headphone jack.

[0309] The power supply 1508 is used to supply power to each component in the terminal. The power supply 1508 can be alternating current, direct current, disposable battery or rechargeable battery. When the power supply 1508 includes a rechargeable battery, the rechargeable battery can support wired charging or wireless charging. The rechargeable battery can also be used to support fast charging technology.

[0310] In some embodiments, the terminal also includes one or more sensors 1509. The one or more sensors 1509 include but are not limited to: an acceleration sensor 1510, a gyroscope sensor 1511, a pressure sensor 1512, an optical sensor 1513, and a proximity sensor 1514.

[0311] The acceleration sensor 1510 can detect the acceleration in three coordinate axes of the coordinate system established by the terminal. For example, the acceleration sensor 1510 can be used to detect the components of the gravitational acceleration in three coordinate axes. The processor 1501 can control the display screen 1505 to display the user interface in a landscape view or a portrait view according to the gravitational acceleration signal collected by the acceleration sensor 1510. The acceleration sensor 1510 can also be used for game or user motion data collection.

[0312] The gyroscope sensor 1511 can detect the body direction and rotation angle of the terminal, and the gyroscope sensor 1511 can collect 3D actions of the user on the terminal in cooperation with the acceleration sensor 1510. The processor 1501 can realize the following functions according to the data collected by the gyroscope sensor 1511: motion sensing (such as changing the UI according to the user's tilt operation), image stabilization when shooting, game control, and inertial navigation.

[0313] The pressure sensor 1512 can be disposed on the side bezel of the terminal and / or the lower layer of the display screen 1505. When the pressure sensor 1512 is disposed on the side bezel of the terminal, it can detect the user's grip signal on the terminal, and the processor 1501 can perform left / right hand recognition or quick operation based on the grip signal collected by the pressure sensor 1512. When the pressure sensor 1512 is disposed on the lower layer of the display screen 1505, the processor 1501 can control the operable controls on the UI interface based on the user's pressure operation on the display screen 1505. The operable controls include at least one of button controls, scroll bar controls, icon controls, and menu controls.

[0314] Optical sensor 1513 is used to collect ambient light intensity. In one embodiment, processor 1501 can control the display brightness of display screen 1505 based on the ambient light intensity collected by optical sensor 1513. Specifically, when the ambient light intensity is high, the display brightness of display screen 1505 is increased; when the ambient light intensity is low, the display brightness of display screen 1505 is decreased. In another embodiment, processor 1501 can also dynamically adjust the shooting parameters of camera assembly 1506 based on the ambient light intensity collected by optical sensor 1513.

[0315] The proximity sensor 1514, also known as a distance sensor, is typically installed on the front panel of the terminal. The proximity sensor 1514 is used to detect the distance between the user and the front of the terminal. In one embodiment, when the proximity sensor 1514 detects that the distance between the user and the front of the terminal is gradually decreasing, the processor 1501 controls the display screen 1505 to switch from a screen-on state to a screen-off state; when the proximity sensor 1514 detects that the distance between the user and the front of the terminal is gradually increasing, the processor 1501 controls the display screen 1505 to switch from a screen-off state to a screen-on state.

[0316] Those skilled in the art will understand that Figure 12 The structure shown does not constitute a limitation on the terminal and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0317] In an exemplary embodiment, a computer-readable storage medium is also provided, which stores at least one computer program that is loaded and executed by a processor of a computer device to enable the computer to implement any of the above-described data processing methods.

[0318] In a possible implementation manner, the computer readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, or the like.

[0319] In the example embodiments, a computer program product is also provided, which includes a computer program or computer instructions loaded and executed by a processor to enable a computer to implement any of the above data processing methods.

[0320] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data for analysis, stored data, displayed data, etc.) and signals involved in the present application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions. For example, the characterization information, training data and test data of the target compound involved in the present application are all obtained under sufficient authorization.

[0321] It should be understood that "multiple" mentioned herein refers to two or more. The "and / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone. The character " / " generally represents that the associated objects before and after it are in an "or" relationship.

[0322] The above only describes the example embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A data processing method, characterized in that, The method includes: Based on the training data and initial neural network model corresponding to the first candidate task, a task processing model corresponding to the first candidate task is obtained; the training data corresponding to the first candidate task includes training sample data and standard processing results corresponding to the training sample data; wherein, when the first candidate task is a task to predict the activity of a compound against a first candidate target protein, the training sample data is the characterization information of the first compound, and the standard processing result is the standard activity of the first compound against the first candidate target protein; when the first candidate task is a task to predict whether an image belongs to a first candidate category, the training sample data is the characterization information of the first image, and the standard processing result is the standard classification result corresponding to the first image; The task processing model corresponding to the first candidate task is tested using the test data corresponding to the first candidate task, and the test performance of the task processing model corresponding to the first candidate task is obtained. Based on the test performance of the task processing model corresponding to the first candidate task, the second-order update gradient corresponding to the parameters of the initial neural network model is obtained, and the second-order update gradient is used as the noise measurement index of the first candidate task. The noise measurement index is used to measure the degree of interference of the first candidate task on updating the parameters of the initial neural network model. The first candidate task is a task that matches the target task to be learned. Based on the noise measurement index of the first candidate task, obtain the sampling probability of the first candidate task; select a reference task whose sampling probability satisfies the first selection condition from the first candidate task. The parameters of the initial neural network model are updated based on the reference task; the target neural network model is then learned based on the parameter-updated neural network model to obtain the target neural network model, which is used for data processing.

2. The method according to claim 1, characterized in that, The method further includes: The first update gradient corresponding to the parameters of the initial neural network model is obtained based on the training data corresponding to the first candidate task; the second update gradient corresponding to the parameters of the initial neural network model is obtained based on the test data corresponding to the first candidate task. Based on the similarity between the first update gradient and the second update gradient, and the test performance of the task processing model corresponding to the first candidate task, a noise metric for the first candidate task is obtained.

3. The method according to claim 1 or 2, characterized in that, The step of obtaining the sampling probability of the first candidate task based on the noise metric of the first candidate task includes: The target scheduling model is invoked to obtain the sampling probability of the first candidate task based on the noise metric of the first candidate task.

4. The method according to claim 3, characterized in that, The call target scheduling model obtains the sampling probability of the first candidate task based on the noise metric of the first candidate task, including: The target scheduling model is invoked to encode the current parameter update round number to obtain first encoded information; the noise measurement index of the first candidate task is encoded to obtain second encoded information. The first encoded information and the second encoded information are concatenated to obtain the target encoded information; the sampling probability of the first candidate task is predicted based on the target encoded information.

5. The method according to claim 3, characterized in that, Before the call target scheduling model obtains the sampling probability of the first candidate task based on the noise metric of the first candidate task, the method further includes: The initial scheduling model is invoked based on the noise metric of the second candidate task to obtain the initial sampling probability of the second candidate task; the initial task whose initial sampling probability satisfies the second selection condition is selected from the second candidate task. Based on the initial task and the initial neural network model, obtain the model to be verified; The verification performance of the model to be verified is obtained based on the verification task. The third update gradient corresponding to the parameters of the initial scheduling model is obtained based on the verification performance. The parameters of the initial scheduling model are updated based on the third update gradient to obtain the target scheduling model.

6. The method according to claim 5, characterized in that, The verification task corresponds to training data and test data; obtaining the verification performance of the model to be verified based on the verification task includes: The model to be verified is trained based on the training data corresponding to the verification task to obtain the task processing model corresponding to the verification task. The test data corresponding to the verification task is used to test the task processing model corresponding to the verification task, and the test performance of the task processing model corresponding to the verification task is obtained. Based on the test performance of the task processing model corresponding to the verification task, the verification performance of the model to be verified is obtained.

7. The method according to claim 3, characterized in that, Before the call target scheduling model obtains the sampling probability of the first candidate task based on the noise metric of the first candidate task, the method further includes: Obtain the sample noise metric for the sample task and the standard sampling probability of the sample task; The initial scheduling model is invoked to obtain the initial sampling probability of the sample task based on the sample noise metric; the target loss is obtained based on the initial sampling probability and the standard sampling probability of the sample task. The fourth update gradient corresponding to the parameters of the initial scheduling model is obtained based on the target loss; the parameters of the initial scheduling model are updated based on the fourth update gradient to obtain the target scheduling model.

8. The method according to claim 1 or 2, characterized in that, The step of obtaining the sampling probability of the first candidate task based on the noise metric of the first candidate task includes: Based on the mapping relationship between noise metrics and sampling probabilities, the noise metrics of the first candidate task are mapped to the sampling probabilities of the first candidate task.

9. The method according to any one of claims 1-2 and 4-7, characterized in that, The number of reference tasks is at least one, and each reference task corresponds to training data and test data; updating the parameters of the initial neural network model based on the reference tasks includes: Based on the training data corresponding to any of the reference tasks and the initial neural network model, obtain the task processing model corresponding to any of the reference tasks; The task processing model corresponding to any of the reference tasks is tested using the test data corresponding to any of the reference tasks to obtain the test performance of the task processing model corresponding to any of the reference tasks. Based on the test performance of the task processing model corresponding to at least one reference task, obtain the target update gradient corresponding to the parameters of the initial neural network model, and update the parameters of the initial neural network model based on the target update gradient.

10. The method according to claim 1 or 2, characterized in that, The step of obtaining the task processing model corresponding to the first candidate task based on the training data and initial neural network model corresponding to the first candidate task includes: The reference model is invoked to process the training sample data to obtain the prediction processing result corresponding to the training sample data. The parameters of the reference model are the same as those of the initial neural network model. Based on the prediction processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data, a first loss is obtained; based on the first loss, the parameters of the reference model are updated to obtain the task processing model corresponding to the first candidate task.

11. The method according to claim 10, characterized in that, When the first candidate task is to predict the activity of a compound against a first candidate target protein, the prediction result is the predicted activity of the first compound against the first candidate target protein. The step of calling the reference model to process the training sample data to obtain the prediction processing result corresponding to the training sample data includes: The reference model is invoked to process the characterization information of the first compound to obtain the predicted activity of the first compound against the first candidate target protein. The first loss is obtained based on the prediction processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data, including: The first loss is obtained based on the predicted activity of the first compound against the first candidate target protein and the standard activity of the first compound against the first candidate target protein.

12. A data processing method, characterized in that, The method includes: Acquire the data to be processed that matches the target task and the target neural network model, wherein the target neural network model is obtained by the data processing method as described in any one of claims 1-11; The target neural network model is invoked to process the data to be processed, and the processing result corresponding to the data to be processed is obtained; Wherein, when the target task is to predict the activity of a compound against a target protein, the data to be processed is the characterization information of the target compound, and the processing result corresponding to the data to be processed is the target activity of the target compound against the target protein. When the target task is to predict whether an image belongs to a target category, the data to be processed is the representation information of the target image, and the processing result corresponding to the data to be processed is the target classification result of the target image.

13. The method according to claim 12, characterized in that, When the target task is to predict the activity of a compound against a target protein, the step of calling the target neural network model to process the data to be processed and obtaining the processing result corresponding to the data to be processed includes: The target neural network model is invoked to process the characterization information of the target compound to obtain the target activity of the target compound against the target protein.

14. The method according to claim 13, characterized in that, The method further includes: Based on the target activity of the target compound against the target protein, a lead compound for the target protein is identified from the target compound.

15. A data processing apparatus, characterized in that, The device includes: The first acquisition unit is configured to acquire a task processing model corresponding to the first candidate task based on training data and an initial neural network model corresponding to the first candidate task; the training data corresponding to the first candidate task includes training sample data and standard processing results corresponding to the training sample data; wherein, when the first candidate task is a task to predict the activity of a compound against a first candidate target protein, the training sample data is the characterization information of the first compound, and the standard processing result is the standard activity of the first compound against the first candidate target protein; when the first candidate task is a task to predict whether an image belongs to a first candidate category, the training sample data is the characterization information of the first image, and the standard processing result is the standard classification result corresponding to the first image; the task processing model corresponding to the first candidate task is tested using test data corresponding to the first candidate task to obtain the test performance of the task processing model corresponding to the first candidate task; based on the test performance of the task processing model corresponding to the first candidate task, the second-order update gradient corresponding to the parameters of the initial neural network model is acquired, and the second-order update gradient is used as a noise metric for the first candidate task, the noise metric being used to measure the degree of interference of the first candidate task on updating the parameters of the initial neural network model, wherein the first candidate task is a task matching the target task to be learned. The second acquisition unit is used to acquire the sampling probability of the first candidate task based on the noise measurement index of the first candidate task; and select a reference task whose sampling probability satisfies the first selection condition from the first candidate task. An update unit is used to update the parameters of the initial neural network model based on the reference task; The third acquisition unit is used to learn the target task based on the neural network model after parameter update, and obtain the target neural network model, which is used for data processing.

16. The apparatus according to claim 15, characterized in that, The first acquisition unit is further configured to acquire a first update gradient corresponding to the parameters of the initial neural network model based on the training data corresponding to the first candidate task; acquire a second update gradient corresponding to the parameters of the initial neural network model based on the test data corresponding to the first candidate task; and acquire a noise measurement index of the first candidate task based on the similarity between the first update gradient and the second update gradient and the test performance of the task processing model corresponding to the first candidate task.

17. The apparatus according to claim 15 or 16, characterized in that, The second acquisition unit is used to call the target scheduling model to obtain the sampling probability of the first candidate task based on the noise measurement index of the first candidate task.

18. The apparatus according to claim 17, characterized in that, The second acquisition unit is used to call the target scheduling model to encode the current parameter update round number to obtain first encoding information; and to encode the noise measurement index of the first candidate task to obtain second encoding information. The first encoded information and the second encoded information are concatenated to obtain the target encoded information; The sampling probability of the first candidate task is predicted based on the target encoding information.

19. The apparatus according to claim 17, characterized in that, The second acquisition unit is further configured to call the initial scheduling model based on the noise measurement index of the second candidate task to obtain the initial sampling probability of the second candidate task; and select an initial task from the second candidate task whose initial sampling probability satisfies the second selection condition. The update unit is further configured to: obtain a model to be verified based on the initial task and the initial neural network model; obtain the verification performance of the model to be verified based on the verification task; obtain the third update gradient corresponding to the parameters of the initial scheduling model based on the verification performance; and update the parameters of the initial scheduling model based on the third update gradient to obtain the target scheduling model.

20. The apparatus according to claim 19, characterized in that, The verification task corresponds to training data and test data; the update unit is used to train the model to be verified based on the training data corresponding to the verification task to obtain the task processing model corresponding to the verification task; to test the task processing model corresponding to the verification task using the test data corresponding to the verification task to obtain the test performance of the task processing model corresponding to the verification task; and to obtain the verification performance of the model to be verified based on the test performance of the task processing model corresponding to the verification task.

21. The apparatus according to claim 17, characterized in that, The second acquisition unit is further configured to acquire the sample noise measurement index of the sample task and the standard sampling probability of the sample task; and call the initial scheduling model to acquire the initial sampling probability of the sample task based on the sample noise measurement index. The update unit is further configured to obtain a target loss based on the initial sampling probability and standard sampling probability of the sample task; obtain a fourth update gradient corresponding to the parameters of the initial scheduling model based on the target loss; and update the parameters of the initial scheduling model based on the fourth update gradient to obtain the target scheduling model.

22. The apparatus according to claim 15 or 16, characterized in that, The second acquisition unit is used to map the noise measurement index of the first candidate task to the sampling probability of the first candidate task based on the mapping relationship between the noise measurement index and the sampling probability.

23. The apparatus according to any one of claims 15-16 and 18-21, characterized in that, The number of reference tasks is at least one, and each reference task corresponds to training data and test data; the update unit is used to obtain the task processing model corresponding to each reference task based on the training data corresponding to each reference task and the initial neural network model. The task processing model corresponding to any reference task is tested using the test data corresponding to any reference task to obtain the test performance of the task processing model corresponding to any reference task; based on the test performance of the task processing model corresponding to at least one reference task, the target update gradient corresponding to the parameters of the initial neural network model is obtained, and the parameters of the initial neural network model are updated based on the target update gradient.

24. The apparatus according to claim 15 or 16, characterized in that, The first acquisition unit is used to call a reference model to process the training sample data and obtain the prediction processing result corresponding to the training sample data. The parameters of the reference model are the same as the parameters of the initial neural network model. Based on the prediction processing result corresponding to the training sample data and the standard processing result corresponding to the training sample data, a first loss is obtained; based on the first loss, the parameters of the reference model are updated to obtain the task processing model corresponding to the first candidate task.

25. The apparatus according to claim 24, characterized in that, When the first candidate task is to predict the activity of a compound against a first candidate target protein, the prediction result is the predicted activity of the first compound against the first candidate target protein. The first acquisition unit is used to call the reference model to process the characterization information of the first compound and obtain the predicted activity of the first compound against the first candidate target protein. The first loss is obtained based on the predicted activity of the first compound against the first candidate target protein and the standard activity of the first compound against the first candidate target protein.

26. A data processing apparatus, characterized in that, The device includes: The fourth acquisition unit is used to acquire the data to be processed that matches the target task and the target neural network model, wherein the target neural network model is obtained by the data processing method as described in any one of claims 1-11; The processing unit is used to call the target neural network model to process the data to be processed and obtain the processing result corresponding to the data to be processed; Wherein, when the target task is to predict the activity of a compound against a target protein, the data to be processed is the characterization information of the target compound, and the processing result corresponding to the data to be processed is the target activity of the target compound against the target protein. When the target task is to predict whether an image belongs to a target category, the data to be processed is the representation information of the target image, and the processing result corresponding to the data to be processed is the target classification result of the target image.

27. The apparatus according to claim 26, characterized in that, When the target task is to predict the activity of a compound against a target protein, the processing unit is used to call the target neural network model to process the characterization information of the target compound to obtain the target activity of the target compound against the target protein.

28. The apparatus according to claim 27, characterized in that, The device further includes: A determining unit is configured to determine, based on the target activity of the target compound against the target protein, a lead compound for the target protein from the target compound.

29. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to enable the computer device to implement the data processing method as described in any one of claims 1 to 11, or the data processing method as described in any one of claims 12 to 14.

30. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to implement the data processing method as described in any one of claims 1 to 11, or the data processing method as described in any one of claims 12 to 14.

31. A computer program product, characterized in that, The computer program product includes a computer program or computer instructions, which are loaded and executed by a processor to enable a computer to implement the data processing method as described in any one of claims 1 to 11, or the data processing method as described in any one of claims 12 to 14.

Images