CN114219338A - Resource allocation method and device based on joint learning - Google Patents
Resource allocation method and device based on joint learning Download PDFInfo
- Publication number
- CN114219338A CN114219338A CN202111574665.XA CN202111574665A CN114219338A CN 114219338 A CN114219338 A CN 114219338A CN 202111574665 A CN202111574665 A CN 202111574665A CN 114219338 A CN114219338 A CN 114219338A
- Authority
- CN
- China
- Prior art keywords
- resource
- learning
- participants
- participant
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06395—Quality analysis or management
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/08—Auctions
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Development Economics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Educational Administration (AREA)
- General Business, Economics & Management (AREA)
- Artificial Intelligence (AREA)
- Marketing (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Software Systems (AREA)
- Finance (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Tourism & Hospitality (AREA)
- Quality & Reliability (AREA)
- Accounting & Taxation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Operations Research (AREA)
- Game Theory and Decision Science (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Information Transfer Between Computers (AREA)
Abstract
The disclosure provides a resource allocation method and device based on joint learning. The method comprises the following steps: based on a joint learning framework, acquiring requirement information of a model demander, quantized values of virtual resource data to be distributed of the model demander, quantized values of learning metadata of a model training participator and quantized values of learning resource data of the model training participator according to resource information provided by a plurality of participators; determining participants of resource evaluation and corresponding excitation matching values based on the quantized values of the virtual resource data to be allocated and the quantized values of the learning resource data; determining the contribution amount of the participants of the resource evaluation based on the learning metadata quantitative value and the demand information of the participants of the resource evaluation, calculating the allocated resource proportion value of the participants of the resource evaluation, and determining the allocation resource scheme of the participants of the resource evaluation. The present disclosure combines participant resource output value and actual contribution measures to achieve a reasonable resource allocation for the participants.
Description
Technical Field
The present disclosure relates to the field of joint learning technologies, and in particular, to a resource allocation method and apparatus based on joint learning.
Background
With the increasing popularity of machine learning, big data-driven intelligent applications will be quickly applied to various aspects of our daily lives, including intelligent voice, medical treatment, traffic, and so on. However, in the conventional machine learning method, it is critical to ensure the accuracy of the training model to collect a sufficient amount of data, which may contain personal information about an individual.
Therefore, joint learning has significant advantages in ensuring data security and user privacy, however, most existing joint learning systems optimistically assume that there are enough users willing to participate in joint learning. In fact, since the participants need to spend a lot of computing and communication resources and to contribute local data, it is important to determine a reasonably fair resource allocation method in order to encourage each user to participate in federal learning. Therefore, how to reasonably and fairly allocate resources in the joint learning becomes an urgent technical problem to be solved.
Disclosure of Invention
In view of this, the embodiments of the present disclosure provide a resource allocation method and device based on joint learning, so as to solve the problem of how to reasonably and fairly perform resource allocation in joint learning.
In a first aspect of the embodiments of the present disclosure, a resource allocation method based on joint learning is provided, including:
acquiring resource information provided by a plurality of participants based on a joint learning architecture, wherein the resource information at least comprises: a model requiring party and a model training participating party;
acquiring requirement information of a model demander, a quantized value of virtual resource data to be distributed of the model demander, a quantized value of learning metadata of a model training participator and a quantized value of learning resource data of the model training participator according to resource information provided by the participator;
determining a participant of resource evaluation and an excitation ratio of the participant of resource evaluation based on the quantized value of the virtual resource data to be allocated and the quantized value of the learning resource data;
determining the contribution amount of the participants of the resource evaluation based on the learning metadata quantized value and the demand information of the participants of the resource evaluation;
calculating the allocation resource ratio of the participants of resource evaluation according to the quantized value and the contribution of the virtual resource data to be allocated;
and determining a resource allocation scheme of the participants of the resource evaluation based on the incentive allocation value and the allocated resource ratio value.
In a second aspect of the embodiments of the present disclosure, a resource allocation apparatus based on joint learning is provided, including:
a first obtaining module, configured to obtain resource information provided by multiple participants based on a joint learning architecture, where the resource information at least includes: a model requiring party and a model training participating party;
the second acquisition module is used for acquiring the requirement information of the model demander, the quantized value of the virtual resource data to be distributed of the model demander, the quantized value of the learning metadata of the model training participator and the quantized value of the learning resource data of the model training participator according to the resource information provided by the participator;
the excitation matching ratio determining module is used for determining a participant of resource evaluation and an excitation matching ratio corresponding to the participant of the resource evaluation based on the quantized value of the virtual resource data to be allocated and the quantized value of the learning resource data;
the contribution amount determining module is used for determining the contribution amount of the participant of the resource evaluation based on the learning metadata quantized value of the participant of the resource evaluation and the requirement information;
the allocation resource ratio determining module is used for calculating the allocation resource ratio of the participant of resource evaluation according to the to-be-allocated virtual resource data quantization value and the contribution amount;
and the allocation module is used for determining a resource allocation scheme of the participants of the resource evaluation based on the incentive allocation value and the allocated resource ratio value. .
In a third aspect of the embodiments of the present disclosure, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above method when executing the computer program.
In a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, which stores a computer program, which when executed by a processor, implements the steps of the above-mentioned method.
Compared with the prior art, the embodiment of the disclosure has the following beneficial effects: the method comprises the steps of determining an incentive proportion value based on learning resources and an allocation resource proportion value based on contribution of a participant respectively based on the learning resource data quantitative value and contribution of the participant after acquiring demand information and a to-be-allocated virtual resource data quantitative value of a model demander and learning metadata quantitative value and learning resource data quantitative value of the participant in model training, wherein the incentive proportion value based on learning resources can meet the requirement that the participant obtains allocation shares not lower than the learning resources paid by the participant in model training, so that the participant is ensured to have enthusiasm for participating in joint learning, and more allocation shares can be acquired based on the contribution (resource output value) of the participant through the allocation resource proportion value weighted by contribution, so that reasonable resource allocation is provided for the participant, the participation enthusiasm of the participant is better excited, and the participant creating more utilities for the model demander participates in joint learning more actively, and the model training quality or the learning task quality required by the model demander is further ensured.
Drawings
To more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings needed for the embodiments or the prior art descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without inventive efforts.
FIG. 1 is a scenario diagram of an application scenario of an embodiment of the present disclosure;
fig. 2 is a schematic flowchart of a resource allocation method based on joint learning according to an embodiment of the present disclosure;
FIG. 3 is a schematic flowchart of another resource allocation method based on joint learning according to an embodiment of the present disclosure;
fig. 4 is a schematic structural diagram of a resource allocation apparatus based on joint learning according to an embodiment of the present disclosure;
fig. 5 is a schematic structural diagram of an electronic device provided in an embodiment of the present disclosure.
Detailed Description
In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular system structures, techniques, etc. in order to provide a thorough understanding of the disclosed embodiments. However, it will be apparent to one skilled in the art that the present disclosure may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present disclosure with unnecessary detail.
Joint learning refers to comprehensively utilizing multiple AI (Artificial Intelligence) technologies on the premise of ensuring data security and user privacy, jointly mining data values by combining multiple parties, and promoting new intelligent business states and modes based on joint modeling. The joint learning has at least the following characteristics:
(1) and the participating nodes control the weak centralized joint training mode of own data, so that the data privacy security in the co-creation intelligent process is ensured.
(2) Under different application scenes, a plurality of model aggregation optimization strategies are established by utilizing screening and/or combined AI algorithm and privacy protection calculation so as to obtain a high-level and high-quality model.
(3) On the premise of ensuring data security and user privacy, the method for improving the efficiency of the joint learning engine is obtained based on a plurality of model aggregation optimization strategies, wherein the efficiency method can improve the overall efficiency of the joint learning engine by solving the problems of information interaction, intelligent perception, abnormal processing mechanisms and the like under the conditions of parallel computing architectures and large-scale cross-domain networks.
(4) The requirements of the users of multiple parties in each scene are acquired, the real contribution of each joint participant is determined and reasonably evaluated through a mutual trust mechanism, and distribution stimulation is carried out.
Based on the mode, the AI technical ecology based on the joint learning can be established, the industrial data value is fully exerted, and the falling of scenes in the vertical field is promoted.
A resource allocation method and apparatus based on joint learning according to an embodiment of the present disclosure will be described in detail below with reference to the accompanying drawings.
Fig. 1 is an architecture diagram of joint learning according to an embodiment of the present disclosure. As shown in fig. 1, the architecture of joint learning may include a server (central node) 101, as well as a participant 102, a participant 103, and a participant 104.
In the joint learning process, a basic model may be built by the server 101, and the server 101 sends the model to the participants 102, 103, and 104 with which communication connections are established. The basic model may also be uploaded to the server 101 after any participant has established the model, and the server 101 sends the model to other participants with whom communication connection is established. The participating party 102, the participating party 103 and the participating party 104 construct models according to the downloaded basic structures and model parameters, perform model training by using local data to obtain updated model parameters, and upload the updated model parameters to the server 101 in an encrypted manner. Server 101 aggregates the model parameters sent by participants 102, 103, and 104 to obtain global model parameters, and passes the global model parameters back to participants 102, 103, and 104. And the participants 102, 103 and 104 iterate the respective models according to the received global model parameters until the models finally converge, thereby realizing the training of the models. In the joint learning process, data uploaded by the participants 102, 103 and 104 are model parameters, local data are not uploaded to the server 101, and all the participants can share the final model parameters, so that common modeling can be realized on the basis of ensuring data privacy. It should be noted that the number of the participants is not limited to three, but may be set according to needs, and the embodiment of the disclosure does not limit this.
In the disclosed embodiments, incentives for joint learning may require model demanders, model training participants, and a joint learning platform (server). The model demand direction jointly learns the model demand to be distributed with virtual resources and resource demands by uploading the model demand, and the model demand can comprise a specific learning task, training precision and the like. The resource requirements may include the type and amount (quantum) of resources required, e.g., whether storage resources are needed, the quantum of bandwidth resources required, the amount of training data required, etc. And the joint learning platform issues information such as model requirements, resource requirements and virtual resources to be allocated of the model demanders to all the participants. And each participant decides whether to participate in the joint learning task according to the information issued by the joint learning platform, and if so, uploads learning metadata and learning resources, wherein the learning metadata comprises data volume, computing resources, storage resources, network bandwidth and the like. The learning resources include the overhead required based on the corresponding learning task and the corresponding training accuracy. The joint learning platform plans the allocation share allocation of the resource evaluation participants and the resource evaluation participants according to the joint learning incentive method provided by the embodiment of the disclosure based on the learning metadata and the learning resources uploaded by the participants, the resource evaluation participants can determine whether to participate in the joint learning or not based on the allocation share allocation, and if the participation is determined, the allocation shares are allocated to the resource evaluation participants according to the allocation shares after the joint learning is completed.
Referring to fig. 2, a resource allocation method based on joint learning provided by the embodiment of the present disclosure is shown, which is used for drawing up participants participating in resource evaluation of the joint learning this time and corresponding allocation share allocation results.
The method of fig. 2 may be performed by the server of fig. 1. As shown in fig. 2, the method includes:
s201, acquiring resource information provided by a plurality of participants based on a joint learning architecture, wherein the resource information at least comprises: a model requiring party and a model training participant.
Specifically, in the joint learning architecture, the model demander may also be one of many participants, and the participant at this time is shown as the demander.
S202, acquiring the requirement information of the model requiring party, the quantized value of the virtual resource data to be distributed, the quantized value of the learning metadata data of the model training participating party and the quantized value of the learning resource data of the model training participating party according to the resource information provided by the participating party.
In this embodiment, when information such as a virtual resource to be allocated, learning metadata, and a learning resource is obtained, the information may be represented in the form of a data quantization value. As an exemplary embodiment, the requirement information may include model requirement information and resource requirement information, where the model requirement information may include specific learning tasks, such as text recognition, image detection, etc., and training accuracy or model effect. In this embodiment, the model requirement information may represent a model training type and a model training difficulty. Resource requirement information may include training data type and data volume computing resources, storage resources, network bandwidth. Learning metadata for a participant may include training data type and data amount computing resources, storage resources, network bandwidth. The learning resources may include data overhead required for resource consumption, computational resource overhead, network bandwidth overhead, storage resource overhead, and the like, as well as training overhead required for model training ease and difficulty, and the like.
S203, determining the participant of resource evaluation and the excitation matching ratio of the participant of resource evaluation based on the quantized value of the virtual resource data to be allocated and the quantized value of the learning resource data.
As an exemplary embodiment, in order to satisfy the participation enthusiasm of the participants, the data quantization value of the virtual resource allocated to the participants should not be lower than the virtual resource data quantization value required by the learning resource of the participants, in this embodiment, the participants for resource evaluation may be determined based on a preset auction manner, for example, the participants for resource evaluation may be selected based on an auction manner such as a first price auction, a VCG auction, and the like, and in this embodiment, the learning platform may determine the participants for obtaining the participation qualification in the first price auction or the VCG auction manner based on the learning resource declared by each participant. And determining the incentive match ratio of each participant in resource evaluation based on the learning resources and the virtual resources to be allocated. As an exemplary embodiment, the virtual resource to be allocated should be greater than or equal to the sum of the resulting virtual resources of the participants of the respective resource evaluations.
And S204, determining the contribution amount of the participants of the resource evaluation based on the learning metadata quantized values and the demand information of the participants of the resource evaluation.
As an exemplary embodiment, after the participants of the resource assessment are obtained, the contribution amounts of the participants of the resource assessment may be measured, the model may be trained based on the participants of the resource assessment and the learning metadata of the participants of the resource assessment, and the contribution amounts of the participants of the resource assessment may be determined based on a preset contribution measurement manner.
S205, determining the allocation resource ratio of the participants of resource evaluation according to the data quantization value and the contribution amount of the virtual resource to be allocated.
As an exemplary embodiment, in order to improve the utility of the model demander, according to the utility of the participant on the model, one part of the virtual resources to be allocated may be used as the sum of the incentive share values, and the other part may be reserved as the allocation share of the further reward of the participant on the utility of the model. Therefore, the participators of the resource evaluation can be additionally stimulated based on the contribution amount of the participators of the resource evaluation to improve the model utility of the model demander, and meanwhile, the participators of the resource evaluation can obtain more allocated resource occupation values based on the contribution of the participators while ensuring the overhead coverage, so that the participation enthusiasm of the participators is better stimulated.
And S206, allocating the allocation share of the participants of the resource evaluation based on the incentive allocation value and the allocation resource ratio value.
As an exemplary embodiment, after determining allocation share allocation, the allocation share allocation result may be sent to the participants of resource evaluation, the participants of resource evaluation determine whether to participate in the joint learning, if so, upload a participation reply result, perform the joint learning based on a participation recovery result replied by the participants of resource evaluation, and after the joint learning is completed, pay allocation shares to the participants of resource evaluation according to drawn allocation share allocation.
In the embodiment, after acquiring the demand information of the model demander and the virtual resources to be allocated and the learning metadata and learning resources of the participators in the model training, respectively determining an incentive proportion value based on learning resources and an allocation resource proportion value based on contribution of the participator based on the learning resources and the contribution of the participator, wherein the incentive proportion value based on overhead can meet the condition that the allocation share obtained by the participator is not lower than the learning resources paid by the participator during model training, ensuring that the participator has the enthusiasm for participating in joint learning, and determining the allocation resource proportion value based on the contribution, more share allocation can be obtained based on the contribution of the participants, the participation enthusiasm of the participants is better stimulated, therefore, the participators creating more utilities for the model demander can participate in the joint learning more actively, and the model training quality or the learning task quality required by the model demander is further ensured.
As an exemplary embodiment, based on a preset screening condition in a joint learning framework, when the learning metadata quantization value matches with the requirement information, determining an initial selection participant; illustratively, the participants meeting the requirement can be preliminarily screened based on the requirement information such as the model requirement information and the resource requirement information issued by the model requirement party, the joint learning request can be sent to the participants meeting the requirement information, and after the responses of the participants are obtained, the participants which are selected preliminarily can be confirmed. After the initial selection participants are determined, determining participants of resource evaluation, participants of resource evaluation and corresponding resource evaluation values in the initial selection participants according to a preset resource value evaluation mode based on the quantized values of the virtual resource data to be allocated and the quantized values of the learning resource data, and calculating corresponding excitation allocation values of the participants of the corresponding resource evaluation according to the resource evaluation values of the participants of the resource evaluation; the quantized value of the virtual resource data obtained based on the excitation matching ratio is greater than or equal to the quantized value of the virtual resource data required by the learning resource, and the sum of the quantized values of the virtual resource data obtained by the participants of resource evaluation is less than or equal to the quantized value of the virtual resource data to be allocated. For example, the resource value evaluation mode may be a first price auction mode or a VCG auction mode, given an auction pricing budget, auctioning with the learning resources of each budget participant as the resource allocation requirements of the participants, and taking the participants meeting the resource allocation requirements as the participants of the resource evaluation.
As an exemplary embodiment, when determining the contribution amount of the participant of the resource evaluation based on the learning metadata and the demand information of the participant of the resource evaluation, the participant of the resource evaluation may be tried to train. And issuing the requirement information to the resource assessment participators, so that the resource assessment participators train the model based on the corresponding learning metadata. Model requirements and resource requirements such as learning tasks, training accuracy and the like in the requirement information are issued to the resource assessment participants, so that the resource assessment participants can perform trial training according to the model requirements to obtain model parameters. And evaluating the model training result according to the model training result fed back by the participant of the resource evaluation to obtain an evaluation result. And determining the contribution amount of the participant of the resource evaluation based on the evaluation result and a preset contribution amount measure.
For example, model parameters returned by the participants in the resource assessment and/or the type and amount of learning metadata participating in the training may be received; determining the contribution amount of the participant of the resource assessment in a manner that at least one of the type and amount of the learning metadata, marginal utility, or contribution value.
Illustratively, when contribution amounts are made according to the type and the amount of the learning metadata, the participant who can accept resource evaluation returns the type and the amount of the learning metadata that participates in training, the type of the learning metadata is matched with the type of the resource demand of the model demander to a higher degree, and the larger the amount or the larger the amount, the larger the corresponding contribution amount.
Illustratively, the contribution amount of the resource assessment participant is determined according to a marginal utility mode, and after model parameters returned by the resource assessment participant are obtained, the models are aggregated to obtain corresponding global model parameters, so as to obtain a global model. And calculating the loss of the global model when the model parameters of each participant are deleted respectively, and taking the loss as the contribution of the participants corresponding to the resource evaluation.
As an alternative embodiment, some low-quality models may not contribute, but may deteriorate the global model, and when measuring the contribution amount of the participants, the model training quality of the participants for resource assessment is also considered, for example, as shown in fig. 3:
s301, model parameters returned by the participants of resource evaluation and the types and the number of learning metadata participating in training are received.
And S302, aggregating the model parameters to obtain global model parameters, and determining the joint learning model.
And S303, respectively calculating the similarity between the model parameters of the participants of the resource evaluation and the global model parameters.
S304, predicting the loss of the joint learning model when the learning metadata of the types and the quantities corresponding to the participants of the resource evaluation are deleted or the model parameters corresponding to the participants of the resource evaluation.
S305, determining the contribution amount of each participant of resource evaluation based on the similarity and the loss. For example, the loss of the joint learning model may be adjusted based on the similarity, and the adjusted loss is used as the contribution amount of the participating party of the resource assessment. The higher the similarity is, the higher the model training quality is, and the loss amount should be additionally increased, so that the corresponding contribution amount is additionally increased.
For example, the similarity between the model parameters and the global model parameters may be calculated for the evaluation of the model quality, and for example, the similarity may be determined by calculating the euclidean distance between the model parameters and the global model parameters, the cosine included angle, and the like.
As an exemplary embodiment, the global model loss is determined in a marginal utility manner, and after model parameters returned by participants of resource evaluation are obtained, the models are aggregated to obtain corresponding global model parameters, so that a global model is obtained. And calculating the loss of the global model when the model parameters of all the participants are deleted respectively. As an optional embodiment, since the deletion order of the participant models has a large influence on the model performance, when determining the model loss by using marginal utility, the model parameters may be sorted based on the similarity of the model parameters, and the loss of the global model when deleting the model parameters of each participant respectively is calculated according to the principle that the lower the similarity is, the higher the priority of deletion is, so as to improve the accuracy of the calculation of the global model loss.
As another optional implementation, when determining the model loss amount, based on the idea of marginal utility, the corresponding loss of the learning metadata adopted by each participant in training the model is calculated respectively, so as to obtain the loss of each participant. In this embodiment, in order to protect privacy of the participants, the participants may upload information representing types and amounts of learning metadata participating in training, and predict, based on the loss prediction model, a loss of the joint learning model when learning metadata of types and amounts corresponding to the participants of each resource evaluation are deleted. The loss of the deleted participant to the joint learning model in the joint learning system can be objectively reflected.
As an optional implementation manner, before the virtual resources to be allocated are allocated according to the learning resources and the contribution amount, the virtual resources to be allocated may be initially allocated, and after the model demander proposes the virtual resources to be allocated, the virtual resources to be allocated may be divided into auction resource allocation shares and contribution resource allocation shares, where the auction resource allocation shares are used for satisfying the learning resources paid by the participant in the subsequent resource evaluation and the model training, and the contribution resource allocation shares are used for ensuring that the utility of the model demander is maximum.
In this embodiment, the classification may be performed based on a preset distribution coefficient, and for example, the preset distribution coefficient may be a preset distribution coefficient, and may be determined based on the demand information.
As an optional implementation, the requirement information includes model requirement information and resource requirement information; the model demand information can be information for characterizing model training difficulty degree such as learning task and training precision, the resource demand information can represent required data type, quality, quantity and the like, and required computing resources, storage resources, bandwidth resources and the like are used for representing resource consumption degree. The harder the learning, the more resources consumed, the higher the learning resources of the participants, and the more resources consumed, the greater the contribution of the participants tends to be. The auction pricing budget usually focuses on learning resources covering the participants, and the contribution measurement budget focuses on the participants with large additional incentive contribution, so that the allocation of the auction resource allocation shares and the contribution resource allocation shares should be different according to different model requirements of model demanders, so as to better encourage the participants to participate in training. For example, for a model that is easier to train but consumes more resources, when auctioning resource allocation shares and contribution resource allocation share allocations, the contribution resource allocation shares should be increased and the auctioning resource allocation shares should be decreased appropriately; for a model which is difficult to train but consumes less resources, when the auction resource allocation share and the contribution resource allocation share are allocated, the contribution resource allocation share should be reduced, and the auction resource allocation share should be properly increased; the allocation of allocation shares is made more rational.
Therefore, in the present embodiment, the determination of the preset distribution coefficient may be performed as follows:
determining a first distribution weight based on the model demand information, wherein the first distribution weight is a coefficient for representing the learning difficulty degree; determining a second distribution weight based on the resource demand information, wherein the second distribution weight is used for representing a coefficient of the resource consumption degree; the preset distribution coefficient is determined based on a ratio of the first distribution weight and the second distribution weight.
After the preset distribution coefficient is obtained, auction resource distribution shares and contribution resource distribution share distribution are determined according to the virtual resources to be distributed and the preset distribution coefficient, distribution share distribution can be more reasonable, the enthusiasm of participants is improved, and the fairness of the participants is guaranteed.
All the above-mentioned optional technical solutions can be combined arbitrarily to form the optional embodiments of the present disclosure, and are not limited herein.
The following are embodiments of the disclosed apparatus that may be used to perform embodiments of the disclosed methods. For details not disclosed in the embodiments of the apparatus of the present disclosure, refer to the embodiments of the method of the present disclosure.
Fig. 4 is a schematic diagram of an apparatus provided by an embodiment of the present disclosure. As shown in fig. 4, the apparatus includes:
a first obtaining module 401, configured to obtain resource information provided by multiple participants based on a joint learning architecture, where the resource information at least includes: a model requiring party and a model training participant.
A second obtaining module 402, configured to obtain, according to the resource information provided by the participant, requirement information of the model demander, a quantized value of virtual resource data to be allocated of the model demander, a quantized value of learning metadata of the participant in model training, and a quantized value of learning resource data of the participant in model training;
an excitation matching ratio determining module 403, configured to determine, based on the quantized value of the virtual resource data to be allocated and the quantized value of the learning resource data, a resource evaluation participant and an excitation matching ratio corresponding to the resource evaluation participant;
a contribution amount determination module 404, configured to determine a contribution amount of a participant of the resource evaluation based on the learning metadata quantized value of the participant of the resource evaluation and the requirement information;
a resource allocation ratio determining module 405, configured to calculate a resource allocation ratio of a participant in resource evaluation according to the quantized value of the virtual resource data to be allocated and the contribution amount;
an allocation module 406, configured to determine a participant allocation resource scheme for the resource assessment based on the incentive match value and the allocation resource fraction value.
According to the technical scheme provided by the embodiment of the disclosure, after the requirement information of the model demander and the virtual resources to be distributed as well as the learning metadata and the learning resources of the participators of the model training are acquired, respectively determining an excitation proportion value based on learning resources and an allocation resource proportion value based on contribution of the participants based on the expenditure and the contribution of the participants, wherein the excitation proportion value based on the learning resources can meet the condition that the allocation share obtained by the participants is not lower than the learning resources paid by the participants during model training, ensuring that the participants have the enthusiasm of participating in joint learning, and determining the allocation resource proportion value based on the contribution, more share allocation can be obtained based on the contribution of the participants, the participation enthusiasm of the participants is better stimulated, therefore, the participators creating more utilities for the model demander can participate in the joint learning more actively, and the model training quality or the learning task quality required by the model demander is further ensured.
In some embodiments, the incentive ratio value determining module comprises: the initial selection unit is used for determining an initial selection participant when the learning metadata quantization value is matched with the requirement information based on a preset screening condition in a joint learning framework; an excitation matching ratio determining unit, configured to determine the resource assessment participant and a corresponding resource evaluation value in the initially selected participant according to a preset resource value assessment manner based on the quantized value of the to-be-allocated virtual resource data and the quantized value of the learning resource data, and calculate an excitation matching ratio corresponding to the participant of the corresponding resource assessment according to the resource evaluation value of the resource assessment participant; and the sum of the excitation matching ratio values of the participants of the resource evaluation is less than or equal to the quantized value of the virtual resource to be allocated.
In some embodiments, the contribution amount determination module comprises: the evaluation unit is used for evaluating the model training result according to the model training result fed back by the participant of the resource evaluation to obtain an evaluation result; and the contribution amount determining unit is used for determining the contribution amount of the participant of the resource evaluation based on the evaluation result and a preset contribution amount measure.
In some embodiments, the contribution amount determination unit is further configured to receive model parameters returned by the participants of the resource evaluation and/or types and amounts of learning metadata participating in the training; determining an amount of contribution of the participant for the resource assessment by learning at least one of a type and amount of metadata, marginal utility, or contribution value.
In some embodiments, the contribution amount determination unit is further configured to receive model parameters returned by the participants of the resource evaluation and types and amounts of learning metadata participating in the training; aggregating the model parameters to obtain global model parameters, and determining a joint learning model; respectively calculating the similarity between the model parameters of the participants of resource evaluation and the global model parameters; predicting the loss of the joint learning model when the learning metadata of the type and the quantity corresponding to each resource evaluation participant is deleted or the model parameters corresponding to the resource evaluation participants are deleted; determining contribution amounts of participants of the respective resource evaluations based on the similarity and the loss.
In some embodiments, the assignment module comprises: the acquisition unit is used for acquiring a preset distribution coefficient; and the allocation unit is used for determining the resource allocation share and the contribution resource allocation share of the participants of the resource evaluation according to the virtual resources to be allocated and the preset allocation coefficient, wherein the resource allocation share is greater than or equal to the sum of the excitation ratio values, and the contribution resource allocation share is greater than or equal to the sum of the allocated resource ratio values.
In some embodiments, the demand information includes model demand information and resource demand information; the obtaining unit is further used for determining a first distribution weight based on the model requirement information, wherein the first distribution weight is a coefficient used for representing the learning difficulty degree; determining a second distribution weight based on the resource demand information, wherein the second distribution weight is used for representing a coefficient of the resource consumption degree; the preset distribution coefficient is determined based on a ratio of the first distribution weight and the second distribution weight.
It should be understood that, the sequence numbers of the steps in the foregoing embodiments do not imply an execution sequence, and the execution sequence of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.
Fig. 5 is a schematic diagram of an electronic device 5 provided by the embodiment of the present disclosure. As shown in fig. 5, the electronic apparatus 5 of this embodiment includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and operable on the processor 501. The steps in the various method embodiments described above are implemented when the processor 501 executes the computer program 503. Alternatively, the processor 501 implements the functions of the respective modules/units in the above-described respective apparatus embodiments when executing the computer program 503.
Illustratively, the computer program 503 may be partitioned into one or more modules/units, which are stored in the memory 502 and executed by the processor 501 to accomplish the present disclosure. One or more modules/units may be a series of computer program instruction segments capable of performing specific functions, which are used to describe the execution of the computer program 503 in the electronic device 5.
The electronic device 5 may be a desktop computer, a notebook, a palm computer, a cloud server, or other electronic devices. The electronic device 5 may include, but is not limited to, a processor 501 and a memory 502. Those skilled in the art will appreciate that fig. 5 is merely an example of the electronic device 5, and does not constitute a limitation of the electronic device 5, and may include more or less components than those shown, or combine certain components, or be different components, e.g., the electronic device may also include input-output devices, network access devices, buses, etc.
The Processor 501 may be a Central Processing Unit (CPU), other general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic device, discrete hardware component, or the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
The storage 502 may be an internal storage unit of the electronic device 5, for example, a hard disk or a memory of the electronic device 5. The memory 502 may also be an external storage device of the electronic device 5, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like provided on the electronic device 5. Further, the memory 502 may also include both internal storage units and external storage devices of the electronic device 5. The memory 502 is used for storing computer programs and other programs and data required by the electronic device. The memory 502 may also be used to temporarily store data that has been output or is to be output.
It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-mentioned division of the functional units and modules is illustrated, and in practical applications, the above-mentioned function distribution may be performed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules, so as to perform all or part of the functions described above. Each functional unit and module in the embodiments may be integrated in one processing unit, or each unit may exist alone physically, or two or more units are integrated in one unit, and the integrated unit may be implemented in a form of hardware, or in a form of software functional unit. In addition, specific names of the functional units and modules are only for convenience of distinguishing from each other, and are not used for limiting the protection scope of the present disclosure. The specific working processes of the units and modules in the system may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.
In the above embodiments, the descriptions of the respective embodiments have respective emphasis, and reference may be made to the related descriptions of other embodiments for parts that are not described or illustrated in a certain embodiment.
Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.
In the embodiments provided in the present disclosure, it should be understood that the disclosed apparatus/electronic device and method may be implemented in other ways. For example, the above-described apparatus/electronic device embodiments are merely illustrative, and for example, a module or a unit may be divided into only one logical function, and may be implemented in other ways, and multiple units or components may be combined or integrated into another system, or some features may be omitted or not implemented. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.
Units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.
In addition, functional units in the embodiments of the present disclosure may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.
The integrated modules/units, if implemented in the form of software functional units and sold or used as separate products, may be stored in a computer readable storage medium. Based on such understanding, the present disclosure may implement all or part of the flow of the method in the above embodiments, and may also be implemented by a computer program to instruct related hardware, where the computer program may be stored in a computer readable storage medium, and when the computer program is executed by a processor, the computer program may implement the steps of the above methods and embodiments. The computer program may comprise computer program code, which may be in the form of source code, object code, an executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying computer program code, recording medium, usb disk, removable hard disk, magnetic disk, optical disk, computer Memory, Read-Only Memory (ROM), Random Access Memory (RAM), electrical carrier wave signals, telecommunications signals, software distribution medium, and the like. It should be noted that the computer readable medium may contain suitable additions or additions that may be required in accordance with legislative and patent practices within the jurisdiction, for example, in some jurisdictions, computer readable media may not include electrical carrier signals or telecommunications signals in accordance with legislative and patent practices.
The above examples are only intended to illustrate the technical solutions of the present disclosure, not to limit them; although the present disclosure has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present disclosure, and are intended to be included within the scope of the present disclosure.
Claims (10)
1. A resource allocation method based on joint learning is characterized by comprising the following steps:
acquiring resource information provided by a plurality of participants based on a joint learning architecture, wherein the resource information at least comprises: a model requiring party and a model training participating party;
acquiring requirement information of a model demander, a quantized value of virtual resource data to be distributed of the model demander, a quantized value of learning metadata of a model training participator and a quantized value of learning resource data of the model training participator according to resource information provided by the participator;
determining a participant of resource evaluation and an excitation matching value corresponding to the participant of the resource evaluation based on the quantized value of the virtual resource data to be allocated and the quantized value of the learning resource data;
determining a contribution amount of a participant of the resource assessment based on the learning metadata quantitative value of the participant of the resource assessment and the demand information;
calculating the allocation resource ratio of the participants of resource evaluation according to the data quantization value of the virtual resource to be allocated and the contribution amount;
determining a participant allocation resource scheme for the resource assessment based on the incentive allocation value and the allocation resource fraction value.
2. The method of claim 1, wherein determining a participant in a resource assessment and a corresponding incentive match value for the participant in the resource assessment based on the quantized value of virtual resource data to be allocated and the quantized value of learning resource data comprises:
determining a primary selection participant when the learning metadata quantization value is matched with the demand information based on a preset screening condition in a joint learning framework;
determining the resource assessment participants and corresponding resource valuations in the initially selected participants according to a preset resource value assessment mode based on the quantized values of the virtual resource data to be allocated and the quantized values of the learning resource data;
calculating an excitation ratio value corresponding to the participant of the resource assessment according to the resource assessment value of the participant of the resource assessment;
and the excitation matching value is greater than or equal to the learning resource data quantized value, and the sum of the excitation matching values of the participants of the resource evaluation is less than or equal to the virtual resource quantized value to be allocated.
3. The method of claim 1, wherein determining the contribution amount of the participant of the resource assessment based on the learning metadata quantitative value and the demand information of the participant of the resource assessment comprises:
evaluating the model training result according to the model training result fed back by the participant of the resource evaluation to obtain an evaluation result;
and determining the contribution amount of the participant of the resource evaluation based on the evaluation result and a preset contribution amount measure.
4. The method of claim 3, wherein determining the contribution amount of the participant in the resource assessment based on the assessment results and a preset contribution amount measure comprises:
receiving model parameters returned by the participants of the resource assessment and/or types and quantities of learning metadata participating in training;
determining an amount of contribution of a participant of the resource assessment by way of at least one of a type and amount of the learning metadata, marginal utility, or a contribution value.
5. The method of claim 3, wherein determining the contribution amount of the participant in the resource assessment based on the assessment results and a preset contribution amount measure comprises:
receiving model parameters returned by the participants of the resource assessment and the types and the number of learning metadata participating in training;
aggregating the model parameters to obtain global model parameters, and determining a joint learning model;
respectively calculating the similarity between the model parameters of the participants of the resource assessment and the global model parameters;
predicting the loss of the joint learning model when learning metadata of types and quantities corresponding to the participants of the resource evaluation are deleted or model parameters corresponding to the participants of the resource evaluation are respectively predicted;
determining the contribution amounts of the participants of the respective resource evaluations based on the similarity and the loss.
6. The method of claim 1, further comprising:
acquiring a preset distribution coefficient;
and determining the resource allocation share and the contribution resource allocation share of the participants of the resource evaluation according to the virtual resources to be allocated and the preset allocation coefficient, wherein the resource allocation share is greater than or equal to the sum of the excitation ratio values, and the contribution resource allocation share is greater than or equal to the sum of the allocation resource ratio values.
7. The method of claim 6, wherein the demand information includes model demand information and resource demand information;
the acquiring of the preset distribution coefficient includes:
determining a first distribution weight based on the model demand information, wherein the first distribution weight is a coefficient for representing learning difficulty degree;
determining a second distribution weight based on the resource demand information, wherein the second distribution weight is used for representing a coefficient of resource consumption degree;
determining the preset distribution coefficient based on a ratio of the first distribution weight and the second distribution weight.
8. An apparatus for resource allocation based on joint learning, comprising:
a first obtaining module, configured to obtain resource information provided by multiple participants based on a joint learning architecture, where the resource information at least includes: a model requiring party and a model training participating party;
the second acquisition module is used for acquiring the requirement information of the model demander, the quantized value of the virtual resource data to be distributed of the model demander, the quantized value of the learning metadata of the model training participator and the quantized value of the learning resource data of the model training participator according to the resource information provided by the participator;
the excitation matching ratio determining module is used for determining a participant of resource evaluation and an excitation matching ratio corresponding to the participant of the resource evaluation based on the quantized value of the virtual resource data to be allocated and the quantized value of the learning resource data;
the contribution amount determining module is used for determining the contribution amount of the participant of the resource evaluation based on the learning metadata quantized value of the participant of the resource evaluation and the requirement information;
the allocation resource ratio determining module is used for calculating the allocation resource ratio of the participant of resource evaluation according to the to-be-allocated virtual resource data quantization value and the contribution amount;
and the allocation module is used for determining a resource allocation scheme of the participants of the resource evaluation based on the incentive allocation value and the allocated resource ratio value.
9. An electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any of claims 1 to 7 when executing the computer program.
10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111574665.XA CN114219338A (en) | 2021-12-21 | 2021-12-21 | Resource allocation method and device based on joint learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111574665.XA CN114219338A (en) | 2021-12-21 | 2021-12-21 | Resource allocation method and device based on joint learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114219338A true CN114219338A (en) | 2022-03-22 |
Family
ID=80704888
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111574665.XA Pending CN114219338A (en) | 2021-12-21 | 2021-12-21 | Resource allocation method and device based on joint learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114219338A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114417361A (en) * | 2022-03-31 | 2022-04-29 | 天聚地合(苏州)科技股份有限公司 | Block chain-based cross-domain AI (Artificial Intelligence) privacy calculation negotiation method and system |
-
2021
- 2021-12-21 CN CN202111574665.XA patent/CN114219338A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114417361A (en) * | 2022-03-31 | 2022-04-29 | 天聚地合(苏州)科技股份有限公司 | Block chain-based cross-domain AI (Artificial Intelligence) privacy calculation negotiation method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113992676B (en) | Incentive method and system for layered federal learning under terminal edge cloud architecture and complete information | |
| Yassine et al. | Cloudlet-based intelligent auctioning agents for truthful autonomous electric vehicles energy crowdsourcing | |
| Thakral | Matching with Stochastic Arrival. | |
| Han et al. | Tiff: Tokenized incentive for federated learning | |
| Long et al. | Equal-quantile rules in resource allocation with uncertain needs | |
| CN114219338A (en) | Resource allocation method and device based on joint learning | |
| CN115292413A (en) | Crowd sensing excitation method based on block chain and federal learning | |
| Amara-Ouali et al. | A benchmark of electric vehicle load and occupancy models for day-ahead forecasting on open charging session data | |
| Monteil et al. | Resource reservation within sliced 5G networks: A cost-reduction strategy for service providers | |
| CN113138847B (en) | Computer resource allocation scheduling method and device based on federal learning | |
| CN112598132A (en) | Model training method and device, storage medium and electronic device | |
| Chislov et al. | Intellectualization of logistic interaction of economic entities of transport and logistics chains | |
| CN116361542A (en) | Product recommendation method, device, computer equipment and storage medium | |
| Egan et al. | Hybrid mechanisms for on-demand transport | |
| CN113837677A (en) | Method, device and equipment for determining logistics line generation strategy | |
| Saengchote et al. | Capitalising the Network Externalities of New Land Supply in the Metaverse | |
| WO2023134181A1 (en) | Resource allocation method, apparatus and system based on federated learning | |
| Ghossoub et al. | Stackelberg equilibria with multiple policyholders | |
| CN116502812A (en) | Resource allocation method and device for joint learning | |
| Araúzo et al. | Agent based modelling and simulation of an auction market for airport slots allocation | |
| CN110264370B (en) | Mutual-help project expense settlement system, method and device | |
| CN116384501A (en) | Combined learning training method and device | |
| CN118333190A (en) | Single budget multi-round federal learning excitation mechanism method, device, system and medium | |
| CN117669772A (en) | Model sharing excitation method, device, equipment and storage medium | |
| CN116304652A (en) | Data heterogeneous-based joint learning model acquisition method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |