An edge intelligence approach that hybridizes distributed learning and centralized learning

Abstract

The present invention discloses an edge intelligence method for hybrid distributed learning and centralized learning, and relates to the field of wireless communications. In each round of communication, each user sends part of the collected data to the edge server, and some users selected to participate in federated learning also perform learning model training locally. After completing data transmission and local model training, the aggregated federated learning model is sent to the edge server using an air computing method. After receiving the user data and the aggregated model, the edge server generates a centralized learning model through training using the user data and mixes it with the federated learning model to obtain the final global model. The present invention provides an effective hybrid distributed learning and centralized learning method for fully utilizing the data collected by user terminals and the powerful computing power of edge computing servers.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and in particular to an edge intelligence method that combines distributed learning and centralized learning. Background Art

[0002] With the development of 5G technology and the widespread adoption of the Internet of Things (IoT), the data generated by geographically distributed mobile devices and IoT devices at the edge of networks will surpass that generated by large cloud data centers. Future wireless networks will need to support the simultaneous access of large numbers of devices. Furthermore, according to forecasts, by 2025, 70% of global IoT data will be processed at the network edge. Simultaneously, people's demands for the quality of service from these smart devices in their daily lives are further increasing. In this context, traditional centralized cloud processing models cannot efficiently process this network edge data and cannot meet users' demand for high-quality service from smart devices. Traditional cloud computing suffers from three shortcomings when processing data at the network edge: insufficient real-time performance, insufficient bandwidth, and high energy consumption. To address these issues, a more suitable approach is to process user needs directly at the edge of the network, giving rise to a new computing paradigm: edge computing. Edge computing moves cloud services from the network core to the network edge, closer to IoT devices and data sources. This technology enables real-time data processing at the network edge, accelerating data flows. In essence, compared with traditional cloud-based computing models, edge computing brings the physical distance between computing and data sources closer, greatly reduces the latency of data transmission, alleviates the pressure on network bandwidth, reduces the energy consumption of data communication, and greatly improves the quality of service for users.

[0003] In addition, the development and popularization of artificial intelligence has given rise to the idea of ​​combining edge computing with artificial intelligence, which has given rise to edge intelligence. Artificial intelligence is trained at the edge through data generated by end users at the edge of the network.

[0004] Methods for implementing edge intelligence include centralized learning by collecting end-user data on edge servers, and distributed learning, where users train locally. Federated learning, a common method for implementing distributed learning, takes a weighted average of each user's local model to obtain a global model. Federated learning only requires each user to transmit model parameters to the server, and does not require specific private data.

[0005] Over-the-air computing technology leverages the natural superposition of wireless multi-access channels to compute the target function during signal transmission. This technology is suitable for implementing federated learning model aggregation in wireless communications. To fully leverage the computing performance of edge servers and end-user data, distributed and centralized learning are combined to implement AI training for edge intelligent systems. Federated learning based on over-the-air computing technology is also used to implement distributed learning.

[0006] However, due to differences in computing power and data volume among end users, only a subset of devices must participate in federated learning to improve the performance of the resulting global AI model. Furthermore, to minimize errors in the aggregated model received by the edge server, the beamforming vectors and transmit power for each user must be correctly designed. Summary of the Invention

[0007] The purpose of this invention is to overcome the defects in the prior art, to make full use of the computing performance of the edge server and the data of the end user to improve the performance of the global artificial intelligence model generated by the edge intelligent system, and to propose an edge intelligence method that hybridizes distributed learning and centralized learning.

[0008] The specific technical solutions adopted in the present invention are as follows:

[0009] The present invention provides an edge intelligence method that combines distributed learning and centralized learning, comprising the following steps:

[0010] 1) Based on an edge intelligence system with K users and one edge server, the edge server obtains the channel state information of the kth user through channel estimation There is an error between the obtained channel state information and the real channel state information. Among them, e k is the channel error, ε k is the upper bound of the channel error;

[0011] 2) In each round of communication, the edge server obtains the channel state information Size of user collection dataset sent to edge server |D C |And the size of the dataset each user can use for federated learning|D F k |, design the receiving beam z for the edge server and the transmitting power p for the kth user k and the selection factor α k ∈{0, 1}, 0 represents not participating in federated learning, 1 represents participating in federated learning;

[0012] 3) Each user sends part of the local data to the edge server, and at the same time, the kth user selects the data based on the selection factor α.k , decide whether to conduct local model training; after local training is completed and local data is sent, the edge server uses the received user data for centralized learning, and at the same time, satisfies α k = 1 user transmits with power p k The model parameters w obtained by local training k The model parameters are aggregated by air calculations when sent to the edge server, and the edge server receives the aggregated signal with the designed receiving beam z;

[0013] 4) After the edge server receives the aggregated signal and completes the centralized training, it performs a weighted average of the model obtained by federated learning and the model obtained by centralized learning to obtain the final global model.

[0014] Preferably, in each round of communication in step 2), the design method of the edge server receiving beam, user transmission power and user selection factor is as follows:

[0015] a) Initialize the device selection vector α = [α1, α2, ... α K ] T =[1, 1, ... 1] T , initialize the transmission power p k =P max / 2,P max The maximum transmit power of the user;

[0016] b) The accuracy of the global model is expressed by the mean square error between the loss function of the expected model and the generated model. where w * is the optimal model, F(·) is the loss function, and the global model obtained after training is completed

[0017] is the global model generated in the previous round, The federated learning model obtained for the edge server, w C The model generated by centralized learning for base stations,

[0018] The upper bound of the mean square error is expressed as:

[0019] in G represents the mean square error between the aggregated federated learning model actually received by the base station and the ideal error-free aggregated federated learning model. 2 is the mean square upper bound of the gradient of the loss function, the Lipschitz constant of the loss function is L and is strongly convex to λ;

[0020] in, is the noise power;

[0021] After simplification, we get Upper bound of

[0022]

[0023] minimize This can be approximated by minimizing its upper bound; establish the optimization problem:

[0024] P:

[0025] st0≤p k ≤P max , k=1,2,...,K,

[0026] α k ∈{0,1},k=1,2,...,K;

[0027] c) Decompose the optimization problem P into two sub-problems and solve them through alternating optimization iterations; according to the initialization [α1, α2, ...α K ] T =[1, 1, ... 1] T , transform the problem P into the first sub-problem P1, fix the users participating in federated learning, optimize the receiving beam z and the transmitting power p k ,Right now:

[0028] P1:

[0029] st0≤p k ≤P max , k = 1, 2, ..., K;

[0030] Decompose the problem P1 into two sub-problems and solve them through alternating optimization; according to the initialization p k =P max / 2, problem P1 is transformed into problem P2:

[0031] P2:

[0032]

[0033] a k ≥0, k=1, 2, ..., K,

[0034] φ k ≥0, k=1, 2, ..., K,

[0035]

[0036] where a k ,φ k To introduce variables, solve problem P2 using the CVX toolkit to obtain z, and substitute the obtained z into problem P1 to obtain problem P3:

[0037] P3:

[0038]

[0039] a k ≥0, k=1, 2, ..., K,

[0040] φ k ≥0, k=1, 2, ..., K,

[0041] 0≤p k ≤P max , k=1,2,...,K,

[0042] P k 2 ≤p k

[0043]

[0044] Solve problem P2 using the CVX toolkit to obtain p k By alternately solving problems P2 and P3 and substituting the obtained optimization variable results into problem P1, new problems P2 and P3 are obtained to iteratively solve until convergence, and the solution z, p of problem P1 is obtained. k ;

[0045] Substituting the solution of problem P1 into problem P, we get problem P4:

[0046] P4:

[0047] stα k ∈{0,1},k=1,2,...,K;

[0048] Problem P4 is a 0-1 quadratic fractional programming problem, which is converted into a linear programming problem and solved to obtain α k By alternately solving problems P1 and P4, and substituting the obtained optimization variable results into problem P to obtain new problems P1 and P4, iteratively solve until convergence, and obtain the solution z, p of problem P k , α k .

[0049] Furthermore, in step c), the CVX toolkit is called to solve the problem in each round of alternating optimization iteration.

[0050] Furthermore, in step 3), the method of performing aggregation by air calculation is: k = 1, the user k uses the transmission power p obtained by optimization in step c). k Transmit signal For the convenience of representation, the signal sent in each time slot is recorded as s k ; The edge server receives the air aggregate signal according to the receive beam z optimized in step c) get n is the noise.

[0051] As a preferred method of obtaining the global model in step 4), the method is: using the data uploaded by the user to learn and obtain the model parameters w C , and then get the received signal ); global model is the global model generated in the previous round.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] The proposed edge intelligence approach, which combines distributed and centralized learning, fully utilizes the computing power of edge servers and user-collected data. Furthermore, some private user data can be fully utilized through federated learning without leaking it. The use of over-the-air computing to aggregate federated learning models significantly reduces communication and computing costs. This approach also improves the performance of the resulting global model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a block diagram of the method of the present invention;

[0055] Figure 2 is the mean square error between the loss function of the model parameters obtained under different signal-to-noise ratios for the method of the present invention and the device that only performs federated learning or performs hybrid learning without federated learning, and the loss function of the expected model;

[0056] Figure 3 It is the accuracy of the global model obtained by the present invention and the other two methods (i.e., only performing federated learning and performing hybrid learning without performing federated learning) under different signal-to-noise ratios. DETAILED DESCRIPTION

[0057] The present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention may be combined accordingly, provided that there is no conflict between them.

[0058] The block diagram of the edge intelligence method of hybrid distributed learning and centralized learning is as follows Figure 1 As shown, the edge server has N antennas, one for each user. Based on partial channel state information and the size of each user's training and upload datasets, the edge server selects which users to participate in federated learning and the transmit power used when each user transmits the model. It also designs a receive beam for itself. The present invention provides an edge intelligence method for hybrid distributed and centralized learning, comprising the following steps:

[0059] 1) K users and one edge server constitute an edge intelligence system;

[0060] 2) The edge server obtains the channel state information of the kth user through channel estimation There is an error between the obtained channel state information and the real channel state information. Among them, e k is the channel error, ε k is the upper bound of the channel error;

[0061] 3) In each round of communication, the edge server obtains the channel state information Size of user collection dataset sent to edge server |D C |And the size of the dataset each user can use for federated learning|D F k |, design the receiving beam z for the edge server and the transmitting power p for the kth user k and the selection factor α k ∈{0, 1}, 0 represents not participating in federated learning, and 1 represents participating in federated learning;

[0062] The design method of edge server receiving beam, user transmission power, and user selection in each round of communication in this step is:

[0063] a) Initialize the device selection vector α = [α1, α2, ... α K ] T =[1, 1, ... 1] T , initialize the transmit power p k =P max / 2,P max The maximum transmit power of the user;

[0064] b) The accuracy of the global model can be expressed by the mean square error between the loss function of the expected model and the generated model. where w * is the optimal model, F(·) is the loss function, and the global model obtained after training is completed is the global model generated in the previous round, The federated learning model obtained for the edge server, w C The model generated by centralized learning for base stations.

[0065] The upper bound of the mean square error can be expressed as:

[0066] in G represents the mean square error between the aggregation model received by the base station and the actual aggregation model. 2 is the mean square upper bound of the gradient of the loss function, the Lipschitz constant of the loss function is L and is strongly convex to λ.

[0067] in, is the noise power.

[0068] The inequality and Bring in

[0069] Simplify and obtain Upper bound of

[0070]

[0071] minimize It can be approximated by minimizing its upper bound. Establish the optimization problem P:

[0072] P:

[0073] st0≤p k ≤P max , k=1,2,...,K,

[0074] α k ∈{0,1},k=1,2,...,K;

[0075] c) The optimization problem P can be decomposed into two sub-problems and solved by alternating optimization iterations. K ] T =[1, 1, ... 1] T , transform the problem into the first sub-problem P1, fix the users participating in federated learning, optimize the receiving beam z and transmitting power p k ,Right now:

[0076] P1:

[0077] st0≤p k ≤P max , k=1, 2, ..., K.

[0078] Problem P1 can also be decomposed into two sub-problems and solved by alternating optimization. k =P max / 2, problem P1 is transformed into:

[0079] P2:

[0080]

[0081] a k ≥0, k=1, 2, ..., K,

[0082] φ k ≥0, k=1, 2, ..., K,

[0083]

[0084] where a k ,φ k To introduce variables, solve problem P2 using the CVX toolkit to obtain z, and substitute the obtained z into problem P1 to obtain problem P3:

[0085] P3:

[0086]

[0087] a k ≥0, k=1, 2, ..., K,

[0088] φ k ≥0, k=1, 2, ..., K,

[0089] 0≤p k ≤P max , k=1,2,...,K,

[0090] P k 2 ≤p k

[0091]

[0092] Solve problem P2 using the CVX toolkit to obtain p k By alternately solving P2, P3 and substituting the obtained optimization variable results into P1 to obtain new P2, P3, iteratively solve until convergence, and obtain the solution z, p of problem P1k .

[0093] Substituting the solution of problem P1 into problem P, we get problem P4:

[0094] P4:

[0095] stα k ∈{0,1},k=1,2,...,K.

[0096] This problem is a 0-1 quadratic fractional programming problem, which can be transformed into a linear programming problem to solve α k By alternately solving P1, P4 and substituting the obtained optimization variable results into P to obtain new P1, P4, iteratively solve until convergence, and obtain the solution z, p of problem P k , α k . The CVX toolkit is called to solve the problem in each round of iteration.

[0097] 4) Each user sends part of the local data to the edge server, and at the same time, the kth user selects the data based on the selection factor α. k , decide whether to conduct local model training. After local training is completed and local data is sent, the edge server uses the received user data for centralized learning, while satisfying α k =1 users with p k The model parameters w obtained by local training k The model parameters are aggregated by air calculations when sent to the edge server, and the edge server receives the aggregated signal with the designed receiving beam z;

[0098] The aggregation method of the aerial computing model in this step is: k = 1, the user k uses the transmission power p obtained by optimization in step c). k Transmit signal s k =|D F k |w k , For the convenience of representation, the signal sent in each time slot is recorded as s k The edge server receives the air aggregate signal according to the receiving beam z optimized in step c). get n is the noise.

[0099] 5) After the edge server receives the aggregated signal and completes the centralized training, it performs a weighted average of the federated learning model and the centralized learning model to obtain the final global model.

[0100] The method to obtain the global model in this step is: use the data uploaded by the user to learn the model parameters wc, and then obtain the model parameters wc according to the received signal. Global Model is the global model generated in the previous round.

[0101] To verify the effectiveness of this invention, a comparative experiment was conducted with two other methods (i.e., only federated learning and hybrid learning without federated learning). The results are as follows:

[0102] like Figure 2 As shown in the figure, the upper bound of the mean square error between the global machine learning model obtained by the proposed algorithm and the theoretical optimal model is significantly smaller than the upper bound of the mean square error between the models obtained by the other two methods and the theoretical optimal model, which shows that the global machine learning model obtained by the proposed algorithm is closer to the theoretical optimal model.

[0103] like Figure 3 As shown in the figure, the global machine learning model obtained by the proposed algorithm has a higher accuracy in handwritten digit recognition than the global models obtained by the other two methods, which shows that the proposed algorithm has a significant improvement on the global model.

[0104] The computer simulations above show that the proposed method achieves better global model performance than the other two methods, and the performance increases with increasing signal-to-noise ratio. Therefore, the proposed edge intelligence method, which combines distributed learning with centralized learning, provides a feasible and effective method for model training in edge intelligence systems.

[0105] The method of the present invention can generate a learning model based on a combination of centralized and distributed learning. It fully utilizes data collected by user terminals and the powerful computing power of edge computing servers. This method uses federated learning to implement distributed learning. To optimize the accuracy of the global model, the edge server optimizes the design of its receive beam, the selection of users participating in federated learning, and the transmit power used when users transmit models and data based on each user's channel state information and the size of the dataset available for local model training. In each communication round, each user sends a portion of their collected data to the edge server. Simultaneously, some users selected for federated learning also train their learning models locally. After data transmission and local model training are complete, the aggregated federated learning model is sent to the edge server using over-the-air computing. After receiving the user data and aggregated model, the edge server uses the user data for training to generate a centralized learning model, which is then combined with the federated learning model to produce the final global model. This invention provides an effective hybrid distributed and centralized learning method that fully utilizes the data collected by user terminals and the powerful computing power of edge computing servers.

[0106] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. An edge intelligence method that combines distributed learning and centralized learning, characterized in that: The steps include: 1) Based on an edge intelligence system with K users and an edge server, the edge server obtains the channel state information of the kth user through channel estimation There is an error between the obtained channel state information and the real channel state information. ||e k ||≤ε k ; Among them, e k is the channel error, ε k is the upper bound of the channel error; 2) In each round of communication, the edge server obtains the channel state information Size of user collection dataset sent to edge server |D C |And the size of the dataset each user can use for federated learning|D F k |, design the receiving beam z for the edge server and the transmitting power p for the kth user k and the selection factor α k ∈{0,1}, 0 represents not participating in federated learning, 1 represents participating in federated learning; 3) Each user sends part of the local data to the edge server, and at the same time, the kth user selects the data based on the selection factor α. k , decide whether to conduct local model training; after local training is completed and local data is sent, the edge server uses the received user data for centralized learning, and at the same time, satisfies α k = 1 user transmits with power p k The model parameters w obtained by local training k The model parameters are aggregated by air calculations when sent to the edge server, and the edge server receives the aggregated signal with the designed receiving beam z; 4) After the edge server receives the aggregated signal and completes centralized training, it performs a weighted average of the federated learning model and the centralized learning model to obtain the final global model. In each round of communication in step 2), the design method of the edge server receiving beam, user transmission power and user selection factor is as follows: a) Initialize the device selection vector α = [α1, α2, ... α K ] T =[1,1,…1] T , initialize the transmission power p k =P max / 2,P max The maximum transmit power of the user; b) The accuracy of the global model is expressed by the mean square error between the loss function of the expected model and the generated model. where w * is the optimal model, F(·) is the loss function, and the global model obtained after training is completed is the global model generated in the previous round, The federated learning model obtained for the edge server, w C The model generated by centralized learning for base stations, The upper bound of the mean square error is expressed as: in G represents the mean square error between the aggregated federated learning model actually received by the base station and the ideal error-free aggregated federated learning model. 2 is the mean square upper bound of the gradient of the loss function, the Lipschitz constant of the loss function is L and is strongly convex to λ; in, is the noise power; After simplification, we get Upper bound of minimize Approximate this by minimizing its upper bound; establish the optimization problem: s.t.0≤p k ≤P max ,k=1,2,…,K, a k ∈{0,1},k=1,2,…,K; c) Decompose the optimization problem P into two sub-problems and solve them through alternating optimization iterations; according to the initialization [α1, α2, ... α K ] T =[1,1,…1] T , transform the problem P into the first sub-problem P1, fix the users participating in federated learning, optimize the receiving beam z and the transmitting power p k ,Right now: s.t.0≤p k ≤P max ,k=1,2,…,K; Decompose the problem P1 into two sub-problems and solve them through alternating optimization; according to the initialization p k =P max / 2, problem P1 is transformed into problem P2: a k ≥0,k=1,2,…,K, f k ≥0,k=1,2,…,K, where a k ,φ k To introduce variables, solve problem P2 using the CVX toolkit to obtain z, and substitute the obtained z into problem P1 to obtain problem P3: a k ≥0,k=1,2,…,K, f k ≥0,k=1,2,…,K, 0≤p k ≤P max ,k=1,2,…,K, P k 2 ≤p k Solve problem P2 using the CVX toolkit to obtain p k By alternately solving problems P2 and P3 and substituting the obtained optimization variable results into problem P1, new problems P2 and P3 are obtained to iteratively solve until convergence, and the solution z, p of problem P1 is obtained. k ; Substituting the solution of problem P1 into problem P, we get problem P4: s.t.α k ∈{0,1},k=1,2,…,K; Problem P4 is a 0-1 quadratic fractional programming problem, which is converted into a linear programming problem and solved to obtain α k By alternately solving problems P1 and P4, and substituting the obtained optimization variable results into problem P to obtain new problems P1 and P4, iteratively solve until convergence, and obtain the solution z,p of problem P k ,α k .

2. The edge intelligence method of hybrid distributed learning and centralized learning according to claim 1 is characterized in that: In the step c), the CVX toolkit is called to solve the problem during each round of alternating optimization iteration.

3. The edge intelligence method of hybrid distributed learning and centralized learning according to claim 1, characterized in that: In step 3), the method of performing aggregation by air calculation is: k = 1, the user k uses the transmission power p obtained by optimization in step c). k Transmit signal s k =|D F k |w k , For the convenience of representation, the signal sent in each time slot is recorded as s k ; The edge server receives the air aggregate signal according to the receive beam z optimized in step c) get n is the noise.

4. The edge intelligence method of hybrid distributed learning and centralized learning according to claim 1, characterized in that: The method of obtaining the global model in step 4) is: using the data uploaded by the user to learn and obtain the model parameters w C , and then get the received signal Global Model is the global model generated in the previous round.

Images