CN114638357A - Edge computing system based on automatic federal learning and learning method thereof - Google Patents
Edge computing system based on automatic federal learning and learning method thereof Download PDFInfo
- Publication number
- CN114638357A CN114638357A CN202210186097.4A CN202210186097A CN114638357A CN 114638357 A CN114638357 A CN 114638357A CN 202210186097 A CN202210186097 A CN 202210186097A CN 114638357 A CN114638357 A CN 114638357A
- Authority
- CN
- China
- Prior art keywords
- edge
- cloud
- edge computing
- model
- data
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000003062 neural network model Methods 0.000 claims abstract description 65
- 238000012549 training Methods 0.000 claims abstract description 64
- 230000001537 neural effect Effects 0.000 claims abstract description 19
- 230000002787 reinforcement Effects 0.000 claims abstract description 19
- 238000012360 testing method Methods 0.000 claims description 31
- 238000010845 search algorithm Methods 0.000 claims description 17
- 230000001133 acceleration Effects 0.000 claims description 14
- 230000002776 aggregation Effects 0.000 claims description 11
- 238000004220 aggregation Methods 0.000 claims description 11
- 230000004044 response Effects 0.000 claims description 11
- 238000009826 distribution Methods 0.000 claims description 9
- 238000004891 communication Methods 0.000 claims description 7
- 238000003860 storage Methods 0.000 claims description 6
- 230000004931 aggregating effect Effects 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 6
- 238000005457 optimization Methods 0.000 abstract description 2
- 238000004364 calculation method Methods 0.000 description 9
- 230000006870 function Effects 0.000 description 9
- 238000010586 diagram Methods 0.000 description 5
- 238000005070 sampling Methods 0.000 description 5
- 238000012545 processing Methods 0.000 description 4
- 238000003709 image segmentation Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001575 pathological effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000007789 sealing Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5072—Grid computing
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/10—Interfaces, programming languages or software development kits, e.g. for simulating neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioethics (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an edge computing system based on automatic federal learning and a learning method thereof, wherein the edge computing system based on automatic federal learning comprises: the system comprises a public cloud, an edge computing box and intelligent terminal equipment; the public cloud gathers neural network model structures and parameters input by the edge cloud to perform parameter updating optimization of different weights, the edge cloud stores data uploaded by an edge computing box, and a neural architecture search model search optimal module combination based on reinforcement learning is adopted to perform localized model training, the trained neural network model is deployed by the edge box to provide real-time computing power for intelligent terminal equipment, and the intelligent terminal equipment acquires information such as images, voice and characters; the invention can effectively solve the problem that the preset network has poor effect on independent and uniformly distributed data, can well protect the privacy of users, can quickly respond to the data acquired by the intelligent terminal by the edge box, and has wide application prospect.
Description
Technical Field
The invention relates to the technical field of edge calculation and federal learning, in particular to an edge calculation system based on automatic federal learning and a learning method of the edge calculation system based on automatic federal learning.
Background
In the related art, the data volume generated by the intelligent device end is increased explosively, however, only cloud computing is relied on, data are uploaded to a cloud end for processing and computing, a large amount of bandwidth resources are consumed, meanwhile, the task with high time delay requirement cannot meet the requirement, the task is unloaded to the edge end for computing, a large amount of data pressure of the cloud end can be well relieved, corresponding computing power is provided near the user side for the edge computing, and network performance is improved. However, due to the diversity of the current intelligent tasks, the types of special equipment are more and more, different types of interfaces bring challenges to edge calculation, and a more universal and concise edge calculation box is urgently needed to be suitable for edge calculation of various tasks.
Meanwhile, the physical isolation in different service cells limits the function of mass data, the problem is well solved by the birth of federal learning, the federal learning can upload model parameters after the model is trained at a local end to carry out aggregation iterative optimization at a cloud end on the premise of not exposing data, and finally, the jointly optimized model is deployed at the local end; the federal learning architecture proposed by google is applied to personalized customization of input methods to medical and commercial aspects, and federal learning is widely applied. Various solutions have appeared in which the model accuracy rate is used as a main research content, mainly focusing on the data level, and selecting a data set with low degree of non-independence and same distribution by reducing clients with poor quality, but still mainly relying on a manually designed model in the aspect of the model. To find a more accurate model architecture, since the data distribution is not visible to the researcher, the developer must design or select multiple architectures and then remotely adjust the hyper-parameters to accommodate the scattered data, a process that is very expensive.
Disclosure of Invention
The present invention is directed to solving, at least to some extent, one of the technical problems in the art described above. Therefore, one purpose of the invention is to provide an edge computing system based on automatic federal learning, which can effectively solve the problem that a preset network has poor effect on independent and identically distributed data, can well protect the privacy of users, can quickly respond to data acquired by an intelligent terminal through an edge computing box, and has wide application prospects.
The second purpose of the invention is to provide a learning method of the edge computing system based on automatic federal learning.
In order to achieve the above object, a first aspect of the present invention provides an edge computing system based on automatic federal learning, which includes: the intelligent terminal equipment is used for acquiring data to be processed and sending the data to be processed to the corresponding edge computing box; the edge computing box is used for storing and preprocessing the data to be processed and transmitting the preprocessed data to be processed to the corresponding edge cloud; the edge cloud is used for generating a neural network model based on a neural architecture search algorithm of reinforcement learning, training the neural network model according to the received data to be processed to obtain trained model parameters, and uploading the neural network model and the trained model parameters to the corresponding public cloud; and the public cloud is used for aggregating the neural network models uploaded by all the edge clouds and the trained model parameters to carry out combined updating, and transmitting the updated parameters back to the edge clouds to carry out model testing.
According to the edge computing system based on automatic federal learning, data to be processed are obtained through intelligent terminal equipment, and the data to be processed are sent to corresponding edge computing boxes; the edge computing box stores and preprocesses the data to be processed, and transmits the preprocessed data to be processed to the corresponding edge cloud; the method comprises the steps that a neural network model is generated by an edge cloud based on a neural architecture search algorithm of reinforcement learning, the neural network model is trained according to received data to be processed to obtain trained model parameters, and the neural network model and the trained model parameters are uploaded to a corresponding public cloud; the public cloud is used for aggregating all the neural network models uploaded by the edge clouds and model parameters after training for joint updating, and returning the updated parameters to the edge clouds for model testing, so that the problem that the effect of the preset network on independent and distributed data is poor can be effectively solved, the privacy of a user can be well protected, the edge computing box can quickly respond to data collected by the intelligent terminal, and the intelligent terminal has a wide application prospect.
In addition, the edge computing system based on automatic federal learning proposed by the invention can also have the following additional technical features:
optionally, the edge computing box includes a CPU/GPU core processor module, an AI acceleration module, a wireless communication module, a network module, a storage module, an intelligent device interface module, and a power module.
Optionally, the edge computing box is connected to c intelligent terminal devices through the intelligent device interface module, and the AI acceleration module is configured to accelerate response time of the intelligent terminal devices.
Optionally, the edge cloud includes an edge management cloud and an edge training cloud, and the edge management cloud is configured to generate a neural network model based on a neural architecture search algorithm of reinforcement learning, and distribute the neural network model to the public cloud for parameter aggregation and weight selection, and meanwhile distribute the neural network model to the edge training cloud for local-end training; after the edge training cloud finishes local end training of a model, transmitting model parameters and test accuracy to the public cloud, wherein the public cloud and the edge training cloud jointly adopt an automatic federal average algorithm, and different weights are given to the data distribution situation of the edge training cloud so that the public cloud aggregates the neural network model uploaded by all the edge clouds and the trained model parameters to perform joint updating; after the public cloud completes the joint updating, the global accuracy is sent to the edge management cloud, so that the edge management cloud can generate a new network model for the policy network by taking the global accuracy as a reward; until the model or the strategy network with the global accuracy reaching the preset requirement is optimized to be converged.
Optionally, the intelligent terminal device is further configured to obtain data to be tested in an actual use stage, and upload the data to be tested to the edge computing box, so as to perform a test by running a model deployed on the edge computing box, so as to obtain a test result.
In order to achieve the above object, a second aspect of the present invention provides a learning method for an edge computing system based on automatic federal learning, wherein the edge computing system based on automatic federal learning includes public clouds, edge computing boxes and intelligent terminal devices, wherein c intelligent terminal devices correspond to one edge computing box, b edge computing boxes correspond to one edge cloud, a edge clouds correspond to one public cloud, and a, b, and c are integers greater than or equal to 1; the learning method includes the steps of: s101, the intelligent terminal equipment acquires data to be processed and sends the data to be processed to the corresponding edge computing box; s102, the edge computing box stores the data to be processed, preprocesses the data to be processed and transmits the preprocessed data to the corresponding edge cloud; s103, the edge cloud generates a neural network model based on a neural architecture search algorithm of reinforcement learning, trains the neural network model according to the received data to be processed to obtain model parameters after training, and uploads the neural network model and the model parameters after training to the corresponding public cloud; s104, the public cloud aggregates the neural network models uploaded by all edge clouds and the trained model parameters to carry out combined updating, and returns the updated parameters to the edge clouds to carry out model testing; and repeatedly executing S103-S104 until the optimal model structure is optimized.
The learning method of the edge computing system based on the automatic federal learning provided by the invention comprises the following steps: s101, the intelligent terminal equipment acquires data to be processed and sends the data to be processed to the corresponding edge computing box; s102, the edge computing box stores the data to be processed, preprocesses the data to be processed and transmits the preprocessed data to the corresponding edge cloud; s103, the edge cloud generates a neural network model based on a neural architecture search algorithm of reinforcement learning, trains the neural network model according to the received data to be processed to obtain model parameters after training, and uploads the neural network model and the model parameters after training to the corresponding public cloud; s104, the public cloud aggregates the neural network models uploaded by all edge clouds and the trained model parameters for joint updating, and returns the updated parameters to the edge clouds for model testing; and repeatedly executing S103-S104 until the optimal model structure is optimized. From this, can effectively solve and predetermine the network and to independent with the poor problem of distributed data effect, protection user privacy that simultaneously can be fine, the data that the edge computing box can quick response intelligent terminal gather have extensive application prospect.
In addition, the learning method of the edge computing system based on automatic federal learning according to the invention can also have the following additional technical characteristics:
optionally, the edge cloud includes an edge management cloud and an edge training cloud, wherein the edge management cloud generates a neural network model by using a neural architecture search algorithm based on reinforcement learning, and distributes the neural network model to the public cloud for parameter aggregation and weight selection, and meanwhile distributes the neural network model to the edge training cloud for local-end training; after the edge training cloud finishes local end training of a model, transmitting model parameters and test accuracy to the public cloud, wherein the public cloud and the edge training cloud jointly adopt an automatic federal average algorithm, and different weights are given to the data distribution situation of the edge training cloud so that the public cloud aggregates the neural network model uploaded by all the edge clouds and the trained model parameters to perform joint updating; after the public cloud completes the joint updating, the global accuracy is sent to the edge management cloud, so that the edge management cloud can generate a new network model for the policy network by taking the global accuracy as a reward; until the model with the global accuracy reaching the preset requirement or the strategy network is optimized to be converged.
Optionally, the edge computing box includes a CPU/GPU core processor module, an AI acceleration module, a wireless communication module, a network module, a storage module, an intelligent device interface module, and a power module.
Optionally, the edge computing box is connected to c intelligent terminal devices through the intelligent device interface module, and the AI acceleration module is configured to accelerate response time of the intelligent terminal devices.
Optionally, the intelligent terminal device is further configured to obtain data to be tested in an actual use stage, and upload the data to be tested to the edge computing box, so as to perform a test by running a model deployed on the edge computing box, so as to obtain a test result.
Drawings
FIG. 1 is a block diagram representation of an edge computing system based on automatic federated learning in accordance with an embodiment of the present invention;
FIG. 2 is a block diagram of an edge computing box according to one embodiment of the present invention;
FIG. 3 is an architectural diagram of automatic federated learning, in accordance with one embodiment of the present invention;
FIG. 4 is a schematic diagram of a policy network node generating a voice-processed LSTM network according to one embodiment of the present invention;
fig. 5 is a flowchart illustrating a learning method of an edge computing system based on automatic federated learning according to an embodiment of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.
In order to better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
In order to better understand the technical solution, the technical solution will be described in detail with reference to the drawings and the specific embodiments.
Referring to fig. 1, an edge computing system based on automatic federal learning according to an embodiment of the present invention includes a public cloud 10, an edge cloud 20, an edge computing box 30, and an intelligent terminal device 40.
The intelligent terminal devices 40 are used for acquiring data to be processed and sending the data to be processed to the corresponding edge computing boxes 30; the edge computing box 30 is used for storing and preprocessing data to be processed, and transmitting the preprocessed data to be processed to the corresponding edge cloud 20; the edge cloud 20 is used for generating a neural network model based on a neural architecture search algorithm of reinforcement learning, training the neural network model according to received data to be processed to obtain trained model parameters, and uploading the neural network model and the trained model parameters to the corresponding public cloud 10; the public cloud 10 is configured to aggregate all the neural network models uploaded by the edge clouds 20 and the trained model parameters for joint update, and return the updated parameters to the edge clouds 20 for model testing.
That is, the smart terminal device 40 is a device that interacts with a person or perceives information; the edge computing box 30 provides data processing and computational power unloading for the intelligent terminal device 40, accelerates the response time of the intelligent terminal device 40, and is compatible with various types of terminals through various deployed interfaces; the edge cloud 20 is a server of the edge computing box 30, which is centrally managed on the edge side and is responsible for managing local area data and providing computing power for model training; the public cloud 10 is a central server, and aggregates the parameter update model uploaded by the edge cloud 20 to realize a central scheduling coordination function, and meanwhile, the public cloud 10 is isolated from local data and only transmits related parameters.
As an embodiment, the intelligent terminal device 40 may be any one of a face recognition entrance guard, a smart home, a wearable device, a case reporter, a CT device, and an MRI device, and the invention is not limited in this regard, and the intelligent terminal device 40 is responsible for performing human-computer interaction or sensing information with a user, and inputting generated data such as images, voices, and characters into the edge computing box 30.
As one embodiment, as shown in fig. 2, the edge computing box 30 includes a CPU/GPU core processor module, an AI acceleration module, a wireless communication module, a network module, a storage module, a smart device interface module, and a power module.
As an embodiment, the edge computing box 30 is connected to c intelligent terminal devices through an intelligent device interface module, and the AI acceleration module is used to accelerate the response time of the intelligent terminal devices.
That is to say, the edge computing box 30 may connect multiple or multiple devices through the smart device interface module of the edge computing box 30, and meanwhile, the edge cloud 20 may also establish a connection with one or more edge computing boxes 30, and the public cloud 10 is responsible for parameter aggregation update and information interaction with one or more edge clouds 20.
It should be noted that, modules with different edge computing capabilities, container/APP functions, rich internet of things interfaces and various communication protocols and visual interfaces, through a high-performance CPU or low-power-consumption GPU, an NPU is a core processor, a plurality of built-in AI acceleration engines are used for edge AI computing acceleration, low-power-consumption and high-credibility edge AI computing capabilities are provided, an artificial intelligent edge computing model and an algorithm are deployed, various kinds of scenarization services are realized in a computing power and flow unloading manner, low-delay and fast-response multi-device scene linkage capabilities are provided for multi-connection devices, computing power unloading capabilities are directly provided for intelligent terminal device 40, response time of intelligent terminal device 40 is accelerated, and various types of terminals are compatible through various deployed interfaces.
As an embodiment, as shown in fig. 3, the edge cloud 20 includes an edge management cloud and an edge training cloud, the edge management cloud is configured to generate a neural network model based on a neural architecture search algorithm of reinforcement learning, and distribute the neural network model to a public cloud for parameter aggregation and weight selection, and simultaneously distribute the neural network model to the edge training cloud for local-end training;
after the edge training cloud finishes local end training of the model, model parameters and test accuracy are transmitted to the public cloud, the public cloud and the edge training cloud jointly adopt an automatic federal average algorithm, different weights are given to the data distribution situation of the edge training cloud, and therefore the public cloud aggregates neural network models uploaded by all edge clouds and trained model parameters to carry out combined updating;
after the public cloud completes the joint updating, the overall accuracy is sent to the edge management cloud, so that the edge management cloud uses the overall accuracy as a reward to enable the policy network RNN to generate a new network model with the probability of P;
until the model with the global accuracy reaching the preset requirement or the strategy network is optimized to be converged.
Specifically, the neural architecture search algorithm deployed by the edge management cloud comprises the following steps:
(1) constructing a search space: for different tasks, a large amount of work has proven the corresponding effective structure, and in order to reduce the search space, a manually designed modular structure chain is adopted, and the search space is the internal structure of the module.
The specific embodiment constructs a search space for processing voice signals and adopts an LSTM module; the search space for constructing the classification model is divided into a conventional convolution module and a down-sampling module; the search space for constructing the segmentation model is constructed as a conventional convolution module, a downsampling module and an upsampling module.
(2) Adopting a search strategy based on reinforcement learning, adopting an RNN structure by a strategy network, using the global accuracy of the edge cloud as a reinforcement learning reward signal, taking the controller generation network as an action, and maximizing the expectation of reward R by adjusting the RNN structure, wherein a1:tIs a set of procedures to design a sub-network structure; m is the number of model structures sampled by the controller on 1 batch; t is the number of network structure hyperparameters predicted by the controller; rkIs the verification accuracy of the network structure at the Kth time after training, thetacAre parameters of the model. The gradient formula of the reward R is:
the specific embodiment 1 adopts an edge computing architecture based on automatic federal learning to generate voice signal processing, a main common algorithm for the connection of voice signals to preceding and following documents is an LSTM algorithm, and an RNN structure is mainly as follows:
ht＝tanh(W1xt+W2ht-1)
the search space constructed by the method is an LSTM module, and the Block operation of the LSTM module is divided into addition, dot multiplication and activation functions (ReLU, sigmoid, tanh and the like). The input is xt,ht-1And ct-1The unit state, as shown in fig. 4, adopts a modular structure of two nodes, and each node corresponds to the RNN policy network to generate the optimal model performance by using the RNN policy network.
In the specific embodiment 2, an edge calculation architecture based on automatic federal learning is adopted to generate image segmentation, the algorithm flow is shown in fig. 3, and the constructed search space is a convolution module, a down-sampling module and an up-sampling module; the Block operation of the model is divided into 3 × 3 convolution kernels, 5 × 5 convolution kernels, pooling, deconvolution and activation functions (ReLU, sigmoid, tanh and the like), an RNN strategy network is constructed to generate a corresponding down-sampling module, an up-sampling module and a conventional convolution module, and the optimal image segmentation neural network model is searched by matching with a UNet structure verified on medical image segmentation.
The model training strategy is not limited to a full-automatic neural architecture search framework, a small data set can be used for searching modules suitable for specific tasks according to the condition of a large data set, then the modules are connected according to an excellent module chain according to the prior, the large data set is trained, the searching time of the model is effectively shortened, and the searching efficiency of the model is greatly improved on the premise of not reducing the accuracy of the model.
Specifically, the public cloud performs parameter aggregation by using an automatic federal average algorithm, and dynamically adjusts the aggregation weight according to data distribution among edge clouds and the training progress of the model to solve the problem that non-independent data training parameters are the same in weight, and the loss function of the model can be expressed as:
wherein alpha iskWeight for each edge cloud model parameter:
the automatic federal average algorithm is characterized by comprising the following steps:
(1) initializing omega0，β0And alpha0＝softmax(β0)，β0Is omega0Initial non-normalized weights of (a);
(2) edge cloud training local model, loss function of kth model: l isk(ω；x)；
(3) All the data are updated
Respectively transmitting the calculated L (beta i; x) into edge clouds k-1, 2 and … N;
(4) updating
Wherein s is the number of iterations of β;
(5) uploading the edge cloud data to a public cloud for aggregation:
(6) and (5) is circulated until the model converges to obtain omegaTIs the best parameter of the current model.
As an embodiment, the intelligent terminal device 40 is further configured to obtain data to be tested in an actual use stage, and upload the data to be tested to the edge computing box 30, so as to perform a test by running a model deployed on the edge computing box 30, so as to obtain a test result.
That is to say, in the actual use stage, the intelligent terminal device 40 only needs to upload data to the edge computing box 30 and run the model deployed on the edge computing box 30 for testing, so that the test result can be efficiently returned, and data sealing in the testing environment is realized to ensure the security of the user data.
It should be noted that the edge computing system based on automatic federal learning can be deployed in hospitals, cells and other occasions according to different scenarios to meet different requirements, for example, the edge computing system is deployed in the home of a cell user to provide face recognition and voice recognition functions, and the edge computing system is deployed in a hospital to provide OCR text recognition, face recognition and pathological diagram analysis.
In summary, according to the edge computing system based on automatic federal learning provided by the present invention, the data to be processed is obtained through the intelligent terminal device, and the data to be processed is sent to the corresponding edge computing box; the edge computing box stores and preprocesses the data to be processed, and transmits the preprocessed data to be processed to the corresponding edge cloud; the method comprises the steps that a neural network model is generated by an edge cloud based on a neural architecture search algorithm of reinforcement learning, the neural network model is trained according to received data to be processed to obtain trained model parameters, and the neural network model and the trained model parameters are uploaded to a corresponding public cloud; the public cloud is used for aggregating all the neural network models uploaded by the edge clouds and model parameters after training for joint updating, and returning the updated parameters to the edge clouds for model testing, so that the problem that the effect of the preset network on independent and distributed data is poor can be effectively solved, the privacy of a user can be well protected, the edge computing box can quickly respond to data collected by the intelligent terminal, and the intelligent terminal has a wide application prospect.
In addition, as shown in fig. 5, an embodiment of the present invention further provides a learning method for an edge computing system based on automatic federal learning, where the edge computing system based on automatic federal learning includes public clouds, edge computing boxes, and intelligent terminal devices, where c intelligent terminal devices correspond to one edge computing box, b edge computing boxes correspond to one edge cloud, a edge clouds correspond to one public cloud, and a, b, and c are integers greater than or equal to 1; the learning method comprises the following steps:
s101, the intelligent terminal equipment acquires data to be processed and sends the data to be processed to a corresponding edge calculation box;
s102, the edge computing box stores data to be processed, preprocesses the data to be processed and transmits the preprocessed data to corresponding edge clouds;
s103, the edge cloud generates a neural network model based on a neural architecture search algorithm of reinforcement learning, trains the neural network model according to received data to be processed to obtain model parameters after training, and uploads the neural network model and the model parameters after training to the corresponding public cloud;
s104, the public cloud aggregates all neural network models uploaded by the edge cloud and trained model parameters to carry out combined updating, and the updated parameters are returned to the edge cloud to carry out model testing;
and repeatedly executing S103-S104 until the optimal model structure is optimized.
As an embodiment, the edge cloud comprises an edge management cloud and an edge training cloud, wherein the edge management cloud generates a neural network model by adopting a neural architecture search algorithm based on reinforcement learning, and distributes the neural network model to a public cloud for parameter aggregation and weight selection, and simultaneously distributes the neural network model to the edge training cloud for local-end training; after local end training of the model is completed by the edge training cloud, model parameters and test accuracy are transmitted to the public cloud, the public cloud and the edge training cloud jointly adopt an automatic Federal averaging algorithm, different weights are given to the data distribution situation of the edge training cloud, and therefore the public cloud aggregates neural network models uploaded by all edge clouds and trained model parameters to conduct combined updating; after the public cloud completes the joint updating, the overall accuracy is sent to the edge management cloud, so that the edge management cloud uses the overall accuracy as a reward to enable the policy network to generate a new network model; until the model with the global accuracy reaching the preset requirement or the strategy network is optimized to be converged.
As one embodiment, the edge computing box includes a CPU/GPU core processor module, an AI acceleration module, a wireless communication module, a network module, a storage module, a smart device interface module, and a power module.
As an embodiment, the edge computing box is connected to the c intelligent terminal devices through the intelligent device interface module, and the AI acceleration module is used to accelerate the response time of the intelligent terminal devices.
As an embodiment, the intelligent terminal device is further configured to obtain data to be tested in an actual use stage, and upload the data to be tested to the edge computing box, so as to perform a test by running a model deployed on the edge computing box, so as to obtain a test result.
It should be noted that the foregoing explanation for the embodiment of the edge computing system based on automatic federated learning is also applicable to the learning method of the edge computing system based on automatic federated learning of this embodiment, and is not repeated here.
In summary, the learning method of the edge computing system based on automatic federal learning provided by the invention comprises the following steps: s101, the intelligent terminal equipment acquires data to be processed and sends the data to be processed to the corresponding edge computing box; s102, the edge computing box stores the data to be processed, preprocesses the data to be processed and transmits the preprocessed data to the corresponding edge cloud; s103, the edge cloud generates a neural network model based on a neural architecture search algorithm of reinforcement learning, trains the neural network model according to the received data to be processed to obtain model parameters after training, and uploads the neural network model and the model parameters after training to the corresponding public cloud; s104, the public cloud aggregates the neural network models uploaded by all edge clouds and the trained model parameters to carry out combined updating, and returns the updated parameters to the edge clouds to carry out model testing; and repeatedly executing S103-S104 until the optimal model structure is optimized. From this, can effectively solve and predetermine the network to independent with the poor problem of distributed data effect, protection user privacy that simultaneously can be fine, the data that the edge calculation box can quick response intelligent terminal gathered have extensive application prospect.
In the description of the present invention, it is to be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", and the like, indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, and are used merely for convenience of description and simplification of the description, but do not indicate or imply that the device or element referred to must have a particular orientation, be constructed in a particular orientation, and be operated, and thus, are not to be construed as limiting the present invention.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the present invention, "a plurality" means two or more unless specifically defined otherwise.
In the present invention, unless otherwise expressly stated or limited, the terms "mounted," "connected," "secured," and the like are to be construed broadly and can, for example, be fixedly connected, detachably connected, or integrally formed; can be mechanically or electrically connected; either directly or indirectly through intervening media, either internally or in any other relationship. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations.
In the present invention, unless otherwise expressly stated or limited, "above" or "below" a first feature means that the first and second features are in direct contact, or that the first and second features are not in direct contact but are in contact with each other via another feature therebetween. Also, the first feature being "on," "above" and "over" the second feature includes the first feature being directly on and obliquely above the second feature, or merely indicating that the first feature is at a higher level than the second feature. A first feature being "under," "below," and "beneath" a second feature includes the first feature being directly under and obliquely below the second feature, or simply meaning that the first feature is at a lesser elevation than the second feature.
In the description of the specification, reference to the description of "one embodiment," "some embodiments," "an example," "a specific example," or "some examples" or the like means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above should not be understood to necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples described in this specification can be combined and combined by those skilled in the art.
Although embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and not to be construed as limiting the present invention, and that changes, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.
Claims (10)
1. An edge computing system based on automated federated learning, comprising: the intelligent terminal equipment is used for acquiring data to be processed and sending the data to be processed to the corresponding edge computing box; the edge computing box is used for storing and preprocessing the data to be processed and transmitting the preprocessed data to be processed to the corresponding edge cloud; the edge cloud is used for generating a neural network model based on a neural architecture search algorithm of reinforcement learning, training the neural network model according to the received data to be processed to obtain trained model parameters, and uploading the neural network model and the trained model parameters to the corresponding public cloud; and the public cloud is used for aggregating the neural network models uploaded by all the edge clouds and the trained model parameters to carry out combined updating, and transmitting the updated parameters back to the edge clouds to carry out model testing.
2. The automatic federal learning based edge computing system as claimed in claim 1, wherein the edge computing box comprises a CPU/GPU core processor module, an AI acceleration module, a wireless communication module, a network module, a storage module, a smart device interface module, and a power module.
3. The automatic federal learning-based edge computing system as claimed in claim 2, wherein the edge computing box is connected to c intelligent end devices through the intelligent device interface module, and the AI acceleration module is used to accelerate the response time of the intelligent end devices.
4. The automatic federated learning-based edge computing system of claim 1, wherein the edge cloud comprises an edge management cloud and an edge training cloud, the edge management cloud being configured to generate a neural network model based on a reinforcement learning neural architecture search algorithm and distribute the neural network model to the public cloud for parameter aggregation and weight selection while distributing to the edge training cloud for local-end training;
after the edge training cloud finishes local end training of a model, transmitting model parameters and test accuracy to the public cloud, wherein the public cloud and the edge training cloud jointly adopt an automatic federal average algorithm, and different weights are given to the data distribution situation of the edge training cloud so that the public cloud aggregates the neural network model uploaded by all the edge clouds and the trained model parameters to perform joint updating;
after the public cloud completes the joint updating, the global accuracy is sent to the edge management cloud, so that the edge management cloud can generate a new network model for the policy network by taking the global accuracy as a reward;
until the model with the global accuracy reaching the preset requirement or the strategy network is optimized to be converged.
5. The automatic federal learning based edge computing system as claimed in claim 4, wherein the intelligent terminal device is further configured to obtain data to be tested in an actual use stage, and upload the data to be tested to the edge computing box, so as to perform testing by running a model deployed on the edge computing box to obtain a testing result.
6. The learning method of the edge computing system based on the automatic federal learning is characterized in that the edge computing system based on the automatic federal learning comprises public clouds, edge computing boxes and intelligent terminal equipment, wherein c intelligent terminal equipment corresponds to one edge computing box, b edge computing boxes correspond to one edge cloud, a edge clouds correspond to one public cloud, and a, b and c are integers which are more than or equal to 1; the learning method includes the steps of:
s101, the intelligent terminal equipment acquires data to be processed and sends the data to be processed to the corresponding edge computing box;
s102, the edge computing box stores the data to be processed, preprocesses the data to be processed and then transmits the preprocessed data to the corresponding edge cloud;
s103, the edge cloud generates a neural network model based on a neural architecture search algorithm of reinforcement learning, trains the neural network model according to the received data to be processed to obtain model parameters after training, and uploads the neural network model and the model parameters after training to the corresponding public cloud;
s104, the public cloud aggregates the neural network models uploaded by all edge clouds and the trained model parameters to carry out combined updating, and returns the updated parameters to the edge clouds to carry out model testing;
and repeatedly executing S103-S104 until the optimal model structure is optimized.
7. The learning method for the edge computing system based on automatic federated learning of claim 6, wherein the edge cloud comprises an edge management cloud and an edge training cloud, wherein the edge management cloud generates a neural network model using a neural architecture search algorithm based on reinforcement learning and distributes the neural network model to the public cloud for parameter aggregation and weight selection, and simultaneously to the edge training cloud for local-end training;
after the edge training cloud finishes local end training of a model, transmitting model parameters and test accuracy to the public cloud, wherein the public cloud and the edge training cloud jointly adopt an automatic federal average algorithm, and different weights are given to the data distribution situation of the edge training cloud so that the public cloud aggregates the neural network model uploaded by all the edge clouds and the trained model parameters to perform joint updating;
after the public cloud completes the joint updating, the global accuracy is sent to the edge management cloud, so that the edge management cloud can generate a new network model for the policy network by taking the global accuracy as a reward;
until the model with the global accuracy reaching the preset requirement or the strategy network is optimized to be converged.
8. The learning method for an edge computing system based on automatic federated learning of claim 6, wherein the edge computing box includes a CPU/GPU core processor module, an AI acceleration module, a wireless communication module, a network module, a storage module, a smart device interface module, and a power module.
9. The learning method of the edge computing system based on automatic federal learning of claim 8, wherein the edge computing box is connected to c intelligent terminal equipments through the intelligent equipment interface module, and the AI acceleration module is used to accelerate the response time of the intelligent terminal equipments.
10. The learning method for the edge computing system based on automatic federal learning as claimed in claim 6, wherein the intelligent terminal device is further configured to obtain the data to be tested in the actual using stage, and upload the data to be tested to the edge computing box, so as to perform the test by running the model deployed on the edge computing box, so as to obtain the test result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210186097.4A CN114638357B (en) | 2022-02-28 | Edge computing system based on automatic federal learning and learning method thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210186097.4A CN114638357B (en) | 2022-02-28 | Edge computing system based on automatic federal learning and learning method thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114638357A true CN114638357A (en) | 2022-06-17 |
| CN114638357B CN114638357B (en) | 2024-05-31 |
Family
ID=
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115830400A (en) * | 2023-02-10 | 2023-03-21 | 南昌大学 | Data identification method and system based on federal learning mechanism |
| CN116070714A (en) * | 2023-01-10 | 2023-05-05 | 天津大学 | Cloud edge cooperative training method and system based on federal learning and neural architecture search |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200293887A1 (en) * | 2019-03-11 | 2020-09-17 | doc.ai, Inc. | System and Method with Federated Learning Model for Medical Research Applications |
| CN112232528A (en) * | 2020-12-15 | 2021-01-15 | 之江实验室 | Method and device for training federated learning model and federated learning system |
| CN113010305A (en) * | 2021-02-08 | 2021-06-22 | 北京邮电大学 | Federal learning system deployed in edge computing network and learning method thereof |
| CN113469325A (en) * | 2021-06-09 | 2021-10-01 | 南京邮电大学 | Layered federated learning method, computer equipment and storage medium for edge aggregation interval adaptive control |
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200293887A1 (en) * | 2019-03-11 | 2020-09-17 | doc.ai, Inc. | System and Method with Federated Learning Model for Medical Research Applications |
| CN112232528A (en) * | 2020-12-15 | 2021-01-15 | 之江实验室 | Method and device for training federated learning model and federated learning system |
| CN113010305A (en) * | 2021-02-08 | 2021-06-22 | 北京邮电大学 | Federal learning system deployed in edge computing network and learning method thereof |
| CN113469325A (en) * | 2021-06-09 | 2021-10-01 | 南京邮电大学 | Layered federated learning method, computer equipment and storage medium for edge aggregation interval adaptive control |
Non-Patent Citations (1)
| Title |
|---|
| 王亚珅;: "面向数据共享交换的联邦学习技术发展综述", 无人系统技术, no. 06, 15 November 2019 (2019-11-15), pages 62 - 66 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116070714A (en) * | 2023-01-10 | 2023-05-05 | 天津大学 | Cloud edge cooperative training method and system based on federal learning and neural architecture search |
| CN116070714B (en) * | 2023-01-10 | 2024-03-08 | 天津大学 | Cloud edge cooperative training method and system based on federal learning and neural architecture search |
| CN115830400A (en) * | 2023-02-10 | 2023-03-21 | 南昌大学 | Data identification method and system based on federal learning mechanism |
| CN115830400B (en) * | 2023-02-10 | 2023-05-16 | 南昌大学 | Data identification method and system based on federal learning mechanism |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2022105714A1 (en) | Data processing method, machine learning training method and related apparatus, and device | |
| CN108280670A (en) | Seed crowd method of diffusion, device and information jettison system | |
| CN108986872A (en) | More granularity attribute weight Spark methods for big data electronic health record reduction | |
| CN113191479A (en) | Method, system, node and storage medium for joint learning | |
| CN113992692B (en) | Method and system for layered federal learning under terminal edge cloud architecture and incomplete information | |
| CN113485409A (en) | Unmanned aerial vehicle path planning and distribution method and system for geographic fairness | |
| CN111343602B (en) | Joint layout and task scheduling optimization method based on evolutionary algorithm | |
| Peng et al. | Deep reinforcement learning based freshness-aware path planning for uav-assisted edge computing networks with device mobility | |
| Song et al. | DRL-Based Backbone SDN Control Methods in UAV-Assisted Networks for Computational Resource Efficiency | |
| CN114638357A (en) | Edge computing system based on automatic federal learning and learning method thereof | |
| Domingo | Power Allocation and energy cooperation for UAV-enabled mmwave networks: A multi-agent deep reinforcement learning approach | |
| CN114638357B (en) | Edge computing system based on automatic federal learning and learning method thereof | |
| Cao et al. | HCOME: Research on hybrid computation offloading strategy for MEC based on DDPG | |
| Reddy et al. | A Beginner's Guide to Federated Learning | |
| CN109101329A (en) | The finegrained tasks distribution method and system of data are acquired by multiple mobile terminals | |
| Hossain et al. | Fedavo: Improving communication efficiency in federated learning with african vultures optimizer | |
| Yu et al. | 5G Multi-Slices Bi-Level Resource Allocation by Reinforcement Learning | |
| Zheng et al. | Research on Energy-Saving Routing Technology Based on Deep Reinforcement Learning | |
| CN115834580B (en) | Ocean big data-oriented distributed data processing method, device and equipment | |
| Li et al. | Privacy Preserved Federated Learning for Skin Cancer Diagnosis | |
| CN117474130B (en) | Federal learning system, method and device based on multi-cloud sharing | |
| CN110175645A (en) | A kind of method and computing device of determining protective device model | |
| CN117557870B (en) | Classification model training method and system based on federal learning client selection | |
| CN115240171B (en) | Road structure sensing method and device | |
| CN117474128B (en) | Distributed online machine learning model construction method based on multi-cloud server |
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 | ||
| GR01 | Patent grant |