WO2022121026A1

WO2022121026A1 - Collaborative learning method that updates central party, storage medium, terminal and system

Info

Publication number: WO2022121026A1
Application number: PCT/CN2020/140757
Authority: WO
Inventors: 戴晶帼; 杨旭; 陈�光; 苏新铎; 叶鹏
Original assignee: 广州广电运通金融电子股份有限公司
Priority date: 2020-12-10
Filing date: 2020-12-29
Publication date: 2022-06-16
Also published as: CN112686368A

Abstract

A collaborative learning method that dynamically updates a central party, a storage medium, a terminal and a system. The system comprises a data provider determination module, a performance evaluation module, a preference module which, according to a performance evaluation, independently selects the best data provider in a network to act as a central server and run a central program; a communication transmission module; and a dynamic updating learning module which, according to an initial status, determines whether to access a collaborative learning task or update the central server; if yes, the collaborative learning task is started and the risk for the central server is predicted; if not, the the central server is accessed for updating, and it is determined whether the collaborative learning task is continued, recovered or stopped until the task ends. Compared to the prior art, the present solution provides a method in which when a central server crashes, an optimal data provider is rapidly selected from among participating training tasks as the central server, which allows for rapid linking so that the collaborative learning can operate continuously.

Description

Update the cooperative learning method, storage medium, terminal and system of the central side

technical field

The invention relates to multi-party data joint processing, in particular to a cooperative learning method, a storage medium, a terminal and a system for dynamically updating a central party.

Background technique

In traditional centralized deep learning, the central server needs to collect a large amount of user data for training the neural network model (referred to as the model), but due to the large network communication overhead of data transmission, user data ownership and user data privacy issues , user data for deep learning is often difficult to obtain.

There are also some solutions to the above problems in the prior art, such as federated learning, which is an emerging machine learning framework, which adopts another way of training machine learning models: in one round of training, each A user uses its private data to train a local model, and then uploads the parameters of the local model to the central server. The central server fuses the parameters of all users to generate the parameters of the global model, and then sends the parameters of the global model to the user. The global model parameters update the local model, and so on for several rounds of training until the global model converges, and the training ends.

To sum up, federated (machine) learning is an emerging distributed machine learning paradigm in which various computing entities (mobile/edge devices, institutions across geographies) are coordinated at a central server (e.g., a service provider) jointly train a machine learning model. Federated learning reduces the privacy risks and data transfer costs of traditional centralized machine learning because the data always resides locally on the computing entity. As a new type of artificial intelligence basic technology, federated learning has received extensive attention from academia and industry in recent years, and has become a new trend in the development and application of machine learning.

Based on this technology, federated learning can enable multiple users to jointly train machine learning models and complete specified learning tasks, such as image classification, text prediction, etc., without leaving the user's private training data locally. The problem of difficult access to user data in centralized machine learning.

However, there are also some security risks in federated learning. Once the central server goes down, it is necessary to start all over again, conduct training for all parties, and then collect data to the new central server, which is costly and time-consuming. Specifically, a model training in federated learning usually includes multiple iteration rounds, and each iteration round includes four steps of model distribution, model calculation, model aggregation and model update (may include entity selection step when the number of computational entities is large) . Model distribution means that the central server distributes the latest model to each participating node; model calculation means that the participating nodes obtain the model update amount or gradient after calculating based on the latest model and local data; model aggregation means that the participating nodes aggregate the calculated model update amount or gradient to Central server; model update means that the central server updates the global model according to the aggregated model update amount or gradient. The model training process repeats the above four steps continuously until the global model converges (that is, the accuracy of the model on the standard test set reaches an ideal value). In the existing federated learning frameworks (such as TensorFlow Federated, FATE), the model distribution and model aggregation in the above steps generally adopts the Hub-and-spoke mode. In this mode, the central server acts as the only model distributor and aggregator. Periodically generate a large number of model communications with participating nodes. In the actual deployment environment, the central server and each participating node are usually distributed across regions, and the network between them is part of the cross-domain public network, which has the characteristics of limited bandwidth and heterogeneous dynamics. Therefore, the communication overhead caused by frequent and large model communication is the main bottleneck of federated learning training efficiency.

It can be seen that compression methods are generally studied from the level of efficient communication algorithms to reduce the amount of data in model communication, but these methods may reduce the quality of the model. As for the central service, it is difficult to quickly select the optimal data provider to participate in the cooperative learning task in the case of downtime, and it cannot be quickly connected, so that the model aggregation of cooperative learning cannot continue to run, which is an urgent problem to be solved at present.

SUMMARY OF THE INVENTION

In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a cooperative learning method, a readable storage medium, a terminal and a system for dynamically updating the center side, which can solve the above problems.

A cooperative learning method for dynamically updating the central party, the method includes the following steps:

Each data provider in S1 is ready, and the task initiator initiates the task;

Each data provider of S2 provides the data used in the cooperative learning task, and obtains the currently available data provider through the network connection status;

S3 evaluates the current performance of available data providers through metrics;

S4 selects the optimal data provider as the central server by comparing the performance of the data provider, and runs the central program;

The S5 central server establishes connections with all available data providers, that is, clients.

S6 judges whether the current task is in the initial state of the task, if yes, the cooperative learning task starts, and the flow proceeds to step S7a; if not, it is currently in the cooperative task progressing state, and the flow proceeds to step S7b;

S7a After the cooperative learning task starts, the task flow is as follows: 7a1) The central server initializes the model parameters; 7a2) The central server distributes the model to the clients; 7a3) Each client uses its local data for model training; 7a4) Each client will train The completed model is encrypted and sent to the central server, and the model is aggregated in the central server; after the aggregation is completed, the current model information is saved and sent to the public space, the stability of the central server is predicted, and the process goes to step S8;

S7b the state task process of the central server includes: 7b1) judging whether the current central server is updated; 7b2) if it is, the central server of the previous round of tasks is abnormal, and the model information is read from the public space on the newly selected optimal central server, Resume the cooperative learning task; if not, the abnormal risk of the central server has been eliminated at this time, and the cooperative learning task continues; 7b3) After one model aggregation is completed, determine whether the cooperative learning task stop condition is met, and if the stop condition is met, enter the process step S9, if the stop condition is not met, predict the stability of the central server, and enter the process step S8;

S8 predicts the stability of the central server. If the system is unstable and the central server is abnormal, the current round of tasks is terminated and the process returns to step S2. The central server is re-selected and transferred to S7a; if the system is stable, the cooperative learning task continues, and the process Go to S7b, repeat the above process until the task stop condition is met, and enter flow step S9;

S9 satisfies the task stop condition, and the task ends.

Preferably, the indicators in step S3 include computing power, bandwidth, and memory.

Preferably, in step S8, a probability graph model is used to analyze the reliability of the central server to predict its stability.

The present invention also provides a computer-readable storage medium on which computer instructions are stored, and when the computer instructions are executed, the steps of the aforementioned method are performed.

The present invention also provides a terminal, including a memory and a processor, the memory stores a registered picture and a computer instruction that can be run on the processor, and the processor executes the method of the foregoing method when the processor runs the computer instruction. step.

The present invention also provides a cooperative learning system based on the dynamic update center. The center server of the system is telecommunicationly connected with each data provider, and the aforementioned steps are executed. The system includes:

The data provider determination module determines the available data provider through the network connection status;

The performance evaluation module evaluates the performance of the data provider through parameters such as computing power, bandwidth, and memory;

The optimal module, according to the evaluation performance, independently selects the optimal data provider in the network as the central server and runs the central program;

Communication transmission module, the central server establishes connections with all available data providers, namely clients;

Dynamically update the learning module, determine whether to enter the cooperative learning task or the central server update through the initial state, if so, start the cooperative learning task and predict the risk of the central server; Abort until the task ends.

Compared with the prior art, the beneficial effect of the present invention is that: this solution provides a method for quickly selecting the optimal data provider in participating in cooperative learning tasks as the central server when the central server is down, which can be quickly connected, so that model training can be performed. Keep running.

Description of drawings

FIG. 1 is a flow chart of the cooperative learning method of the dynamic update center of the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

It should be understood that "system", "device", "unit" and/or "module" as used in this specification is a method used to distinguish different components, elements, parts, parts or assemblies at different levels. However, other words may be replaced by other expressions if they serve the same purpose.

As shown in the specification and claims, unless the context clearly dictates otherwise, the words "a", "an", "an" and/or "the" are not intended to be specific in the singular and may include the plural. Generally speaking, the terms "comprising" and "comprising" only imply that the clearly identified steps and elements are included, and these steps and elements do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

Flowcharts are used in this specification to illustrate operations performed by a system according to an embodiment of this specification. It should be understood that the preceding or following operations are not necessarily performed in the exact order. Instead, the various steps can be processed in reverse order or simultaneously. At the same time, other actions can be added to these procedures, or a step or steps can be removed from these procedures.

A large amount of information data is flooded in various industries such as economy, culture, education, medical care, and public management. Data processing and analysis such as data analysis, data mining, and trend forecasting are widely used in more and more scenarios. Among them, through data cooperation, multiple data owners can obtain better data processing results. For example, more accurate model parameters can be obtained through multi-way cooperative learning.

In some embodiments, the method of dynamically updating the cooperative learning of the central party can be applied to a scenario in which all parties cooperate to train a machine learning model for use by multiple parties under the condition of ensuring data security of all parties. In this scenario, multiple data parties have their own data, and they want to jointly use each other's data for unified modeling (eg, classification models, linear regression models, logistic regression models, etc.), but do not want their own data (especially privacy data) were leaked. For example, Internet savings institution A has a batch of user data, and bank B has another batch of user data. The training sample set determined based on the user data of A and B can be trained to obtain a machine learning model with better performance. Both A and B are willing to participate in model training through each other's user data, but for some reasons A and B do not want their user data information to be leaked, or at least do not want to let the other party know their user data information.

In some embodiments, cooperative learning can be performed using a federated learning approach. Federated Learning can carry out efficient machine learning among multiple parties or computing nodes. Federated learning enables multi-party data to perform model training without the local training samples, and only transfers the trained model or calculates the gradient, which protects the privacy of the training samples held by all parties.

In some embodiments, federated learning is often used in situations where the model is computationally intensive and has many parameters. In the embodiment of this scenario, due to the large amount of data transmission in the federated learning process, the pressure of communication transmission is relatively large. Therefore, in the scenario of federated learning, it is often necessary to adopt a certain method to reduce the communication pressure during the transmission process.

In some embodiments of this specification, during each iterative update of the model, the cooperative learning task judgment (including model gradient values or model parameters obtained by training) updated by the central server may be used for compression. Specifically, by restoring and continuing the updated task, the training of the client model can be uninterrupted and retraining is not required, thereby reducing the communication pressure. At the same time, the risk prediction is carried out for the abnormal situation of the central server to ensure the stability of the model.

first embodiment

A cooperative learning method for dynamically updating the central party, the method includes the following steps.

Each data provider in S1 is ready, and the task initiator initiates the task.

Each data provider of S2 provides the data used in the cooperative learning task, and obtains the currently available data provider through the network connection status.

S3 evaluates its current performance against available data providers through metrics.

S4 selects the optimal data provider as the central server by comparing the performance of the data provider to run the central program.

S6 judges whether the current task is in the task initial state, and if so, the cooperative learning task starts, and the flow goes to step S7a. If not, it is currently in a cooperative task progress state, and the flow proceeds to step S7b.

After the S7a cooperative learning task starts, the task flow is as follows:

7a1) The central server initializes model parameters. For example, before using the network data, the central server needs to initialize the model parameters, such as the weights and biases in the linear regression model. Import the init module from MXNet. This module provides various methods for initializing model parameters. Here init is the abbreviated form of initializer. Specify the weight parameter by init.Normal(sigma=0.01) Each element will be randomly sampled from a normal distribution with mean 0 and standard deviation 0.01 during initialization. The bias parameter is initialized to zero by default.

7a2) The central server distributes the model to the client.

7a3) Each client uses its local data for model training.

7a4) Each client encrypts the trained model and sends it to the central server, where the model is aggregated. After the aggregation is completed, the current model information is saved and sent to the public space, the stability of the central server is predicted, and the process goes to step S8.

The status task process of the S7b central server includes:

7b1) Determine whether the current central server is updated.

7b2) If the central server of the previous round of tasks is abnormal, the model information is read from the public space on the newly selected optimal central server, and the cooperative learning task is resumed. If not, the abnormal risk of the central server has been eliminated at this time, and the cooperative learning task continues.

7b3) After one model aggregation is completed, judge whether the cooperative learning task stop condition is met. If the stop condition is met, go to process step S9, if the stop condition is not met, predict the stability of the central server and go to process step S8.

S8 predicts the stability of the central server. If the system is unstable and the central server is abnormal, the current round of tasks is terminated and the process returns to step S2, where the central server is reselected and the process goes to S7a. If the system is stable and the cooperative learning task continues, the flow goes to S7b, the above process is repeated until the task stop condition is satisfied, and the flow goes to step S9.

S9 satisfies the task stop condition, and the task ends.

The indicators in step S3 include computing power, bandwidth, and memory.

Further, in step S7a, the standard for model parameter initialization is conventional requirements:

(1) The parameters cannot all be initialized to 0, nor can they all be initialized to the same value;

(2) It is best to ensure that the mean value of parameter initialization is 0, the positive and negative parameters are interleaved, and the positive and negative parameters are roughly equal in number.

Several common initialization methods are "random initialization according to normal distribution - corresponding to normal" and "random initialization according to uniform distribution - corresponding to uniform", which will not be repeated, in addition to:

①Normalized Glorot initialization - glorot_normal, which uses the Glorot normal distribution initializer, also known as the Xavier normal distribution initializer. It draws samples from a truncated normal distribution centered at 0 and with standard deviation stddev=sqrt(2/(fan_in+fan_out)), where fan_in is the number of input units in the weight tensor and fan_out is the output in the weight tensor the number of units. Normalized Glorot initialization - glorot_uniform, Glorot uniform initializer, also known as Xavier uniform initializer. It draws samples from a uniform distribution in [-limit, limit], where limit is sqrt(6/(fan_in+fan_out)), fan_in is the number of input units in the weights tensor, and fan_out is the number of output units in the weights tensor quantity.

② Kaiming initialization, also known as he initialization, also known as msra initialization: a normalized kaiming initialization - he_normal He normal distribution initializer. It draws samples from a truncated normal distribution centered at 0 and standard deviation stddev=sqrt(2/fan_in), where fan_in is the number of input units in the weight tensor, implemented in keras as keras.initializers.he_normal (seed=None). b Normalized kaiming initialization - he_uniform, He uniform variance scaling initializer. It draws samples from a uniform distribution in [-limit, limit], where limit is sqrt(6/fan_in), where fan_in is the number of input units in the weights tensor. keras.initializers.he_uniform(seed=None).

③lecun initialization, a standardized kaiming initialization - lecun_uniform, LeCun uniform initializer keras.initializers.lecun_uniform (seed=None). It draws samples from a uniform distribution in [-limit, limit], where limit is sqrt(3/fan_in) and fan_in is the number of input units in the weights tensor. b Normalized kaiming initialization -- lecun_normal, LeCun normal distribution initializer keras.initializers.lecun_normal(seed=None). It draws samples from a truncated normal distribution centered at 0 with standard deviation stddev=sqrt(1/fan_in), where fan_in is the number of input units in the weight tensor.

④Batch Normalization, BN is to change the input data distribution into Gaussian distribution, which can ensure that the input of each layer of neural network maintains the same distribution. As the number of network layers increases, the distribution gradually shifts, and the reason for the slow convergence is that the overall distribution approaches the upper and lower limits of the value range of the nonlinear function. This causes the gradient to vanish during backpropagation. BN is to forcibly pull the distribution of the input value of any neuron in each layer of neural network back to a standard normal distribution with a mean of 0 and a variance of 1 by means of normalization, so that the activation input value falls into the more sensitive area of the nonlinear function. It can make the gradient larger, the learning convergence speed is fast, and the convergence speed can be greatly accelerated. Scale and Shift act on γ and β. γ and β are learned parameters that make the standard normal distribution taller/fatder and skewed left and right.

In step S8, the reliability of the central server is analyzed by using a probabilistic graphical model to predict its stability. Among them, the probabilistic graphical model is a general term for a class of models that express probabilistic correlations in graphical patterns. The probabilistic graphical model combines the knowledge of probability theory and graph theory, and uses graphs to represent the joint probability distribution of variables related to the model. Common probabilistic graphical models include Bayesian network, Markov network and Hidden Markov network, and any model can be used in this scheme.

Second Embodiment

The present invention also provides a computer-readable storage medium on which computer instructions are stored, and when the computer instructions are executed, the steps of the aforementioned method are performed. Wherein, for the method, please refer to the detailed introduction in the foregoing part, and details are not repeated here.

Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, and the computer-readable medium includes a permanent Persistent and non-permanent, removable and non-removable media can be implemented by any method or technology for information storage. Information may be computer readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Flash Memory or other memory technology, Compact Disc Read Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, Magnetic tape cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

Third Embodiment

The present invention also provides a terminal, including a memory and a processor, the memory stores data provider information and computer instructions that can be executed on the processor, and the processor executes the aforementioned computer instructions when the processor executes the computer instructions. steps of the method. Wherein, for the method, please refer to the detailed introduction in the foregoing part, and details are not repeated here.

Fourth Embodiment

A cooperative learning system based on a dynamic update center, the system is used to make the center server telecommunication connection with each data provider, and the system includes:

Communication transmission module, the central server establishes connections with all available data providers, namely clients.

The key point of the present invention is the updated cooperative learning task state of the central server. This part mainly runs steps S7a and S7b in the pre-procedure method. During the process of the system executing the above method, the client model training will not be interrupted, and there is no need to restart Training, and can ensure the rapid update of the central server and the rapid recovery after an exception, until the learning task is completed.

It is to be understood that the systems and modules thereof described in one or more implementations of this specification can be implemented in a variety of ways. For example, in some embodiments, the system and its modules may be implemented in hardware, software, or a combination of software and hardware. Wherein, the hardware part can be realized by using dedicated logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. Those skilled in the art will appreciate that the methods and systems described above may be implemented using computer-executable instructions and/or embodied in processor control code, for example on a carrier medium such as a disk, CD or DVD-ROM, such as a read-only memory (firmware) ) or a data carrier such as an optical or electronic signal carrier. The system and its modules of the present application can not only be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc. , can also be implemented by, for example, software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software (eg, firmware).

It should be noted that different embodiments may have different beneficial effects, and in different embodiments, the possible beneficial effects may be any one or a combination of the above, or any other possible beneficial effects.

It should be noted that the above description of the processing device and its modules is only for convenience of description, and does not limit the present application to the scope of the illustrated embodiments. It can be understood that for those skilled in the art, after understanding the principle of the system, various modules may be combined arbitrarily, or a subsystem may be formed to connect with other modules without departing from the principle.

The basic concept has been described above. Obviously, for those skilled in the art, the above detailed disclosure is only an example, and does not constitute a limitation to the present application. Although not explicitly described herein, various modifications, improvements, and corrections to this application may occur to those skilled in the art. Such modifications, improvements, and corrections are suggested in this application, so such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

Meanwhile, the present application uses specific words to describe the embodiments of the present application. Such as "one embodiment," "an embodiment," and/or "some embodiments" means a certain feature, structure, or characteristic associated with at least one embodiment of the present application. Therefore, it should be emphasized and noted that two or more references to "an embodiment" or "one embodiment" or "an alternative embodiment" in various places in this specification are not necessarily referring to the same embodiment . Furthermore, certain features, structures or characteristics of the one or more embodiments of the present application may be combined as appropriate.

Furthermore, those skilled in the art will appreciate that aspects of this application may be illustrated and described in several patentable categories or situations, including any new and useful process, machine, product, or combination of matter, or combinations of them. of any new and useful improvements. Accordingly, various aspects of the present application may be performed entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software. The above hardware or software may be referred to as a "data block", "module", "engine", "unit", "component" or "system". Furthermore, aspects of the present application may be embodied as a computer product comprising computer readable program code embodied in one or more computer readable media.

A computer storage medium may contain a propagated data signal with the computer program code embodied therein, for example, on baseband or as part of a carrier wave. The propagating signal may take a variety of manifestations, including electromagnetic, optical, etc., or a suitable combination. Computer storage media can be any computer-readable media other than computer-readable storage media that can communicate, propagate, or transmit a program for use by coupling to an instruction execution system, apparatus, or device. Program code on a computer storage medium may be transmitted over any suitable medium, including radio, cable, fiber optic cable, RF, or the like, or a combination of any of the foregoing.

The computer program coding required for the operation of the various parts of this application may be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python etc., conventional procedural programming languages such as C language, VisualBasic, Fortran2003, Perl, COBOL2002, PHP, ABAP, dynamic programming languages such as Python, Ruby and Groovy, or other programming languages, etc. The program code may run entirely on the user's computer, or as a stand-alone software package on the user's computer, or partly on the user's computer and partly on a remote computer, or entirely on the remote computer or processing device. In the latter case, the remote computer can be connected to the user's computer through any network, such as a local area network (LAN) or wide area network (WAN), or to an external computer (eg, through the Internet), or in a cloud computing environment, or as a service Use eg software as a service (SaaS).

In addition, unless explicitly stated in the claims, the order of processing elements and sequences described in the present application, the use of numbers and letters, or the use of other names are not intended to limit the order of the procedures and methods of the present application. While the foregoing disclosure discusses by way of various examples some embodiments of the invention that are presently believed to be useful, it is to be understood that such details are for purposes of illustration only and that the appended claims are not limited to the disclosed embodiments, but rather The requirements are intended to cover all modifications and equivalent combinations falling within the spirit and scope of the embodiments of the present application. For example, although the system components described above may be implemented by hardware devices, they may also be implemented by software-only solutions, such as installing the described systems on existing processing devices or mobile devices.

Similarly, it should be noted that, in order to simplify the expressions disclosed in the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, in the drawings or descriptions thereof. However, this method of disclosure does not imply that the subject matter of the application requires more features than those mentioned in the claims. Indeed, there are fewer features of an embodiment than all of the features of a single embodiment disclosed above.

Some examples use numbers to describe quantities of ingredients and attributes, it should be understood that such numbers used to describe the examples, in some examples, use the modifiers "about", "approximately" or "substantially" to retouch. Unless stated otherwise, "about", "approximately" or "substantially" means that a variation of ±20% is allowed for the stated number. Accordingly, in some embodiments, the numerical parameters set forth in the specification and claims are approximations that can vary depending upon the desired characteristics of individual embodiments. In some embodiments, the numerical parameters should take into account the specified significant digits and use a general digit reservation method. Notwithstanding that the numerical fields and parameters used in some embodiments of the present application to confirm the breadth of their ranges are approximations, in particular embodiments such numerical values are set as precisely as practicable.

Each patent, patent application, patent application publication, and other material, such as articles, books, specifications, publications, documents, etc., cited in this application is hereby incorporated by reference in its entirety. Application history documents that are inconsistent with or conflict with the content of this application are excluded, as are documents (currently or hereafter appended to this application) that limit the broadest scope of the claims of this application. It should be noted that, if there is any inconsistency or conflict between the descriptions, definitions and/or terms used in the attached materials of this application and the content of this application, the descriptions, definitions and/or terms used in this application shall prevail .

Finally, it should be understood that the embodiments described in the present application are only used to illustrate the principles of the embodiments of the present application. Other variations are also possible within the scope of this application. Accordingly, by way of example and not limitation, alternative configurations of embodiments of the present application may be considered consistent with the teachings of the present application. Accordingly, the embodiments of the present application are not limited to the embodiments expressly introduced and described in the present application.

It should also be noted that the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device comprising a series of elements includes not only those elements, but also Other elements not expressly listed, or inherent to such a process, method, article of manufacture or apparatus are also included. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in the process, method, article of manufacture or device that includes the element.

It will be appreciated by those skilled in the art that the embodiments of the present application may be provided as methods, apparatuses, systems or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

A cooperative learning method for dynamically updating a central party, characterized in that the method comprises the following steps:

Each data provider in S1 is ready, and the task initiator initiates the task;

Each data provider of S2 provides the data used in the cooperative learning task, and obtains the currently available data provider through the network connection status;

S3 evaluates the current performance of available data providers through metrics;

S4 selects the optimal data provider as the central server by comparing the performance of the data provider, and runs the central program;

The S5 central server establishes connections with all available data providers, that is, clients.

S6 judges whether the current task is in the initial state of the task, if yes, the cooperative learning task starts, and the flow proceeds to step S7a; if not, it is currently in the cooperative task progressing state, and the flow proceeds to step S7b;

S7a After the cooperative learning task starts, the task flow is as follows: 7a1) The central server initializes the model parameters; 7a2) The central server distributes the model to the clients; 7a3) Each client uses its local data for model training; 7a4) Each client will train The completed model is encrypted and sent to the central server, and the model is aggregated in the central server; after the aggregation is completed, the current model information is saved and sent to the public space, the stability of the central server is predicted, and the process goes to step S8;

S7b the state task process of the central server includes: 7b1) judging whether the current central server is updated; 7b2) if it is, the central server of the previous round of tasks is abnormal, and the model information is read from the public space on the newly selected optimal central server, Resume the cooperative learning task; if not, the abnormal risk of the central server has been eliminated at this time, and the cooperative learning task continues; 7b3) After one model aggregation is completed, determine whether the cooperative learning task stop condition is met, and if the stop condition is met, enter the process step S9, if the stop condition is not met, predict the stability of the central server, and enter the process step S8;

S8 predicts the stability of the central server. If the system is unstable and the central server is abnormal, the current round of tasks is terminated and the process returns to step S2. The central server is re-selected and transferred to S7a; if the system is stable, the cooperative learning task continues, and the process Go to S7b, repeat the above process until the task stop condition is met, and enter flow step S9;

S9 satisfies the task stop condition, and the task ends.
The method according to claim 1, wherein the indicators in step S3 include computing power, bandwidth, and memory.
The method according to claim 1, characterized in that: in step S8, a probability graph model is used to analyze the reliability of the central server to realize its stability prediction.
A computer-readable storage medium on which computer instructions are stored, characterized in that: when the computer instructions are executed, the steps of the method according to any one of claims 1-3 are executed.
A terminal, comprising a memory and a processor, characterized in that: the memory stores data provider information and computer instructions that can be run on the processor, and the processor executes the claims when running the computer instructions Steps of any one of the methods 1-3.
A cooperative learning system based on a dynamic update center, wherein the center server of the system is telecommunicationly connected with each data provider, and runs the steps of the method described in any one of claims 1-3, wherein the system comprises:

The data provider determination module determines the available data provider through the network connection status;

The performance evaluation module evaluates the performance of the data provider through parameters such as computing power, bandwidth, and memory;

The preferred module, according to the evaluation performance, independently selects the optimal data provider in the network as the central server and runs the central program;

Communication transmission module, the central server establishes connections with all available data providers, namely clients;

Dynamically update the learning module, determine whether to enter the cooperative learning task or the central server update through the initial state, if so, start the cooperative learning task and predict the risk of the central server; Abort until the task ends.