1 Federated Learning and Meta Learning Approaches Applications and Directions

2025-04-30 0 0 4.78MB 48 页 10玖币

侵权投诉

Federated Learning and Meta Learning:

Approaches, Applications, and Directions

Xiaonan Liu, Member, IEEE, Yansha Deng, Senior Member, IEEE,

Arumugam Nallanathan and Mehdi Bennis, Fellow, IEEE

Abstract—Over the past few years, signiﬁcant advancements

have been made in the ﬁeld of machine learning (ML) to

address resource management, interference management, auton-

omy, and decision-making in wireless networks. Traditional ML

approaches rely on centralized methods, where data is collected

at a central server for training. However, this approach poses

a challenge in terms of preserving the data privacy of devices.

To address this issue, federated learning (FL) has emerged as

an effective solution that allows edge devices to collaboratively

train ML models without compromising data privacy. In FL,

local datasets are not shared, and the focus is on learning a

global model for a speciﬁc task involving all devices. However, FL

has limitations when it comes to adapting the model to devices

with different data distributions. In such cases, meta learning

is considered, as it enables the adaptation of learning models to

different data distributions using only a few data samples. In this

tutorial, we present a comprehensive review of FL, meta learning,

and federated meta learning (FedMeta). Unlike other tutorial

papers, our objective is to explore how FL, meta learning, and

FedMeta methodologies can be designed, optimized, and evolved,

and their applications over wireless networks. We also analyze

the relationships among these learning algorithms and examine

their advantages and disadvantages in real-world applications.

Index Terms—Centralized learning, distributed learning, fed-

erated learning, meta learning, federated meta learning, wireless

networks.

I. INTRODUCTION

The rapid advancement of technology and the increasing

proliferation of devices, including IoT sensors, mobile phones,

and tablets, have resulted in an unprecedented growth in data

generation [1], [2]. According to a report issued by the SG

Analytics, 2.5 quintillion bytes of data were generated every

day in 2020 [3]. To extract meaningful information and enable

dynamic decision-making for various tasks, machine learning

(ML) algorithms are used to analyze these datasets [4], [5],

X. Liu is with the Department of Engineering, King’s College London,

U.K. (e-mail:{liuxiaonan19931107}@gmail.com). (This work is performed

at KCL.)

Y. Deng is with the Department of Engineering, King’s College London,

London, WC2R 2LS, U.K. (e-mail:{yansha.deng}@kcl.ac.uk). (Correspond-

ing author: Yansha Deng).

A. Nallanathan is with the School of Electronic Engineering and Com-

puter Science, Queen Mary University of London (QMUL), U.K. (e-

mail:{a.nallanathan}@qmul.ac.uk).

M. Bennis is with Faculty of Information Technology and Electrical En-

gineering, Centre for Wireless communications, University of Oulu, Finland.

(e-mail:{mehdi.bennis}@oulu.ﬁ).

This work was supported in part by Engineering and Physical Sciences Re-

search Council (EPSRC), U.K., under Grant EP/W004348/1 , EP/W004100/1,

and in part by UKRI under the UK government’s Horizon Europe funding

guarantee (grant number 10061781), as part of the European Commission-

funded collaborative project VERGE, under SNS JU program (grant number

101096034).

such as controlling self-driving cars [6], pattern recognition

[7], or prediction of user behavior [8].

Data Device

Model Training

Central Server .

…

Data

Uploading

Output

Downloading

D2D

(a) Centralized Machine Learning

Fig. 1. The canonical architecture of centralized machine learning, device to

device communication, and vehicle to vehicle communication.

According to [9], in traditional cloud computing systems,

various types of datasets, such as images, videos, audio ﬁles,

and location information, acquired from the internet of thing

(IoT) devices or mobile devices, are transmitted to a cloud-

based server or data center [10], where centralized ML models

are trained. As shown in Fig. 1, in centralized learning,

the devices communicate with a central server to upload

their data by wired/wireless connections. In particular, the

devices transmit their local data to a central server, and the

central server performs all the computational tasks to train the

data. Then, the output of the centralized learning model is

delivered to devices. Centralized learning is computationally-

efﬁcient for devices as they are not required to be equipped

with high computation capability. However, the traditional

centralized training method may not be suitable for the ever-

growing complexity and heterogeneity of wireless networks,

such as device to device (D2D) and vehicle to vehicle (V2V)

communications, as shown in Fig. 1 (b) and (c), respectively,

due to the following reasons:

•The presence of limited bandwidth, dynamic wireless

channels, and high interference can result in signiﬁcant

transmission latency, which can negatively affect real-

time applications, such as wireless virtual reality (VR)

arXiv:2210.13111v2 [cs.LG] 4 Nov 2023

[11]–[13], D2D, and self-driving car systems [14]. These

applications require immediate and responsive data pro-

cessing, but the delays caused by limited bandwidth

and the unstable wireless channels negatively affect their

performance and reliability.

•The process of transmitting a large amount of data to

the cloud leads to signiﬁcant communication overhead, as

well as increased storage and computational costs [15],

[16].

•The data collected from users, such as medical and

ﬁnancial information, can often be private and sensitive in

nature. Transmitting this data to the cloud raises concerns

on privacy and security, as it exposes users’ personal

information to potential risks. This situation becomes

particularly problematic when it comes to its compliant

with data privacy legislation, such as the European Com-

mission’s General Data Protection Regulation (GDPR)

[17] and the Consumer Privacy Bill of Rights in the U.S.

[18].

In order to overcome the aforementioned challenges, dis-

tributed learning has emerged as a solution to effectively

and efﬁciently learn models from distributed data. Distributed

learning refers to the use of multi-node machine learning algo-

rithms and systems that are speciﬁcally designed to address the

computational challenges associated with complex centralized

algorithms operating on large-scale datasets [19], [20]. By

employing distributed learning algorithms, multiple learning

models can be trained based on distributed datasets. This ap-

proach offers several advantages over centralized approaches,

particularly when the number of datasets is large. One no-

table advantage is the potential for reducing biases, as the

distributed learning enables the utilization of diverse datasets

and mitigates the impact of individual dataset characteristics

[21]–[23].

Conventional machine learning algorithms, such as k-

nearest neighbor [24], support vector machine [25], [26],

Bayesian networks [27], [28], and decision trees [29], [30],

can be trained in a distributed manner by exchanging raw data,

which can hardly protect the user data privacy [31]. In order to

ensure the privacy of data and facilitate collaborative machine

learning (ML) involving complex models distributed across

IoT or mobile devices, a method called federated learning

(FL) was initially introduced in [32] by McMahan et al.

The standard steps of FL include: 1) each device trains a

local model using its own dataset; 2) the devices send their

local models to a central server for model aggregation; 3) the

server updates the global model and transmits it back to the

devices. These steps are repeated iteratively until convergence.

In addition, FL offers several advantages over traditional

centralized learning and distributed learning approaches that

rely on the exchange of raw data. Some of these advantages

include:

•Data Privacy: FL prioritizes user privacy by exchanging

only model weights between the server and the devices,

rather than the datasets themselves. This approach ef-

fectively protects user privacy during the collaborative

learning process.

•Data Diversity: FL enables the utilization of heteroge-

neous data by allowing the aggregation of models from

different devices, leading to the development of an en-

hanced global model. This aggregation process enhances

the overall performance and effectiveness of the model

from various devices.

In addition to the aforementioned advantages, FL has

demonstrated its effectiveness in various applications, includ-

ing training predictive models for human trajectory/behavior

by mobile devices [33], automatically learning users’ behavior

patterns through smart home IoT [34], and enabling collab-

orative diagnosis in health artiﬁcial intelligence (AI) among

multiple hospitals and government agencies [35]–[37]. How-

ever, these applications usually suffer from data heterogeneity

(i.e., data with different statistical characteristics) and system

heterogeneity (i.e., systems with different types of computation

units), which severely affect the convergence and accuracy of

FL algorithms. It is worth noting that FL primarily focuses on

learning a single ML task across multiple devices [38], [39]

and only develops a common output for all devices. Therefore,

it does not adapt the model to each device. This is an important

missing feature, especially given the heterogeneity of the un-

derlying data distribution for various devices. To address these

limitations, integrating meta learning, also known as learning

to learn, with FL becomes crucial, resulting in what is known

as federated meta learning (FedMeta). FedMeta aims to solve

multi-task learning problems. Meta learning aims to create AI

models that are able to adapt to new tasks and improve their

learning performance over time, without the need for extensive

retraining. In other words, meta learning typically involves

training a learning model on a variety of different tasks,

with the goal of learning generalizable knowledge that can be

transferred to new tasks. This is different from traditional ML,

where a learning model is typically trained on a single task

and then used for that task alone. In general, a meta learning

algorithm is trained using the outputs, e.g., model predictions,

and metadata of ML algorithms. After ﬁnishing training, its

learning models are delivered for testing and used to make

ﬁnal predictions. In addition, meta learning includes tasks

such as observing the learning performance of different ML

models on learning tasks, learning from metadata, and faster

learning process for new tasks. While both federated learning

(FL) and meta learning can be used in distributed learning

systems and involve sharing a global model among multiple

devices, they differ in three key aspects: (1) Data Distribution:

In FL, devices have their own datasets but perform the same

task. On the other hand, meta learning involves multiple tasks,

each with its corresponding dataset. (2) Update Mechanism:

FL utilizes local updates deployed by devices to enhance

learning performance, whereas meta learning employs inner-

loop updates for each task to improve learning performance.

(3) Aggregation and Parameter Updates: FL applies model

aggregation to improve the global performance of all devices.

In contrast, meta learning employs an outer-loop to update

global parameters for all tasks. Initialization-based meta-

learning algorithms, in particular, are known for fast adaptation

and good generalization to new tasks. These distinctions

highlight the different approaches and objectives of FL and

meta learning. While FL focuses on collaborative learning with

devices having different data distributions but performing the

same task, meta learning guarantees the adaptation of models

to multiple tasks with their corresponding datasets, often

emphasizing fast adaptation and generalization capabilities,

especially by initialization-based algorithms [40]. The goal

of FedMeta is to collaboratively meta-train a learning model

using data from different tasks distributed among devices.

The server maintains the initialized model and updates it by

collecting testing loss from a mini batch of devices. The

transmitted information in learning process consists of the

model parameter initialization (from a server to devices) and

testing loss (from devices to the server), and no data is required

to be delivered to the server. Compared to FL, FedMeta has

the following advantages:

•FedMeta brings a reduction in the required communica-

tion cost because of faster convergence, and an increase

in learning accuracy. Additionally, it is highly adaptable

and can be employed for arbitrary tasks across multiple

devices. This ﬂexibility allows for efﬁcient and accurate

learning in various scenarios.

•FedMeta enables model sharing and local model training

without substantial expansion in model size. As a result,

it does not consume a large amount of memory, and

the resulting global model can be personalized for each

device. This aspect allows for efﬁcient utilization of

resources and customization of the global model to satisfy

the speciﬁc requirements of devices.

A. Related Works

This section provides a brief review of relevant surveys and

tutorials on FL and meta-learning. Additionally, it highlights

the novel contributions of this paper.

1) FL: In the past 5 years, numerous surveys and tutorials

have been published on FL methodologies [41]–[52] and their

applications over wireless [53]–[65] networks. To differentiate

our tutorial from these existing surveys and tutorials, we have

classiﬁed them into different categories based on their primary

focus in Table I. We then compare and summarize the content

of these surveys and tutorials, aligning them with the structure

of our own tutorial, as presented in Tables II and III. Table I

highlights that surveys and tutorials for FL methodologies pri-

marily focus on either fundamental deﬁnitions, architectures,

challenges, future directions, and applications of FL [41], [43],

[44], or a speciﬁc subﬁeld of FL, such as emerging trends of

FL (FL in the intersection with other learning paradigms) [45],

communication efﬁciency of FL (challenges and constraints

caused by limited bandwidth and computation ability) [46],

fairness-aware FL (client selection, optimization, contribution

evaluation, incentive distribution, and performance metrics)

[47], federated reinforcement learning (FRL) (deﬁnitions, evo-

lution, and advantages of horizontal FRL and vertical FRL)

[48], FL for natural language processing (NLP) (algorithm

challenges, system challenges as well as privacy issues) [49],

incentive schemes for FL (stackelberg game, auction, contract

theory, Shapley value, reinforcement learning, and blockchain)

[50], [51], unlabeled data mining in FL (potential research

directions, application scenarios, and challenges) [42], and

FL over next-generation Ethernet Passive Optical Networks

[52]. On the other hand, FL surveys and tutorials for wire-

less networks mainly focus on either fundamental theories,

key techniques, challenges, future directions, and applications

[54]–[57], [64], or a speciﬁc subﬁeld of FL applied in wireless

networks, including FL in IoT (data sharing, ofﬂoading and

caching, attack detection, localization, mobile crowdsensing,

and privacy) [58], [61], [62], communication-efﬁcient FL un-

der various challenges incurred by communication, computing,

energy, and data privacy issues [53], [63], FL for mobile edge

computing (MEC) (communication cost, resource allocation,

data privacy, data security, and implementation) [59], collab-

orative FL (deﬁnitions, advantages, drawbacks, usage condi-

tions, and performance metrics) [60], and asynchronous FL

(device heterogeneity, data heterogeneity, privacy and security,

and applications) [65]. To the best of the authors’ knowledge,

this is the ﬁrst paper considering FL methodologies in different

research areas, including model aggregation, gradient descent,

communication efﬁciency, fairness, Bayesian learning, and

clustering, and how FL algorithms evolve from the canonical

one, namely, federated averaging (FedAvg), in detail. Also,

we present a qualitative comparison considering advantages

and disadvantages among different FL algorithms in the same

research area.

2) Meta Learning: There have been several surveys and

tutorials focusing on meta learning methodologies over the

past 20 years [66]–[72]. However, to the best of the authors’

knowledge, there are no existing surveys and tutorials in meta

learning over wireless networks. To distinguish the scope of

our tutorial from existing literature, we classify the available

meta learning surveys and tutorials into different categories

based on their focus, as presented in Table I. Furthermore, we

compare and summarize the content of these existing surveys

and tutorials in Table IV, aligning them with the structure

and content of our own tutorial. Table I illutrates that the

existing meta learning surveys and tutorials focused either on

the general advancement, including deﬁnitions, models, chal-

lenges, research directions, and applications, of meta learning

[66], [67], [70], [71], or exploring the detailed applications

of meta learning in a speciﬁc meta learning ﬁeld, including

algorithm selection (transfer learning, few-shot learning, and

beyond supervised learning) for data mining [68], NLP (es-

pecially few-shot applications, including deﬁnitions, research

directions, and some common datasets) [69], and multi-modal

meta learning in terms of the methodologies (few-shot learning

and zero-shot learning) and applications [72]. To the best

of the authors’ knowledge, no existing surveys or tutorials

have covered the application of FedMeta in wireless networks,

making our tutorial a unique contribution to this area of

research.

Tables II, III, and IV provide a comparative analysis of

existing surveys and tutorials in relation to our tutorial. It

is observed that the existing literature only covers a limited

number of subtopics related to our tutorial and offers only

brief descriptions of the corresponding learning algorithms. In

contrast, our tutorial goes beyond these limitations by provid-

TABLE I

ANOVERVIEW OF SELECTED SURVEYS AND TUTORIALS ON FL AND META LEARNING

Subject Ref. Contributions

Methodologies

[41], [43], [44] Deﬁnitions, architectures, challenges, future directions, and applications of the FL framework

[45] Survey on emerging trends in FL, including FL in the intersection with other learning paradigms

[46] Survey on communication efﬁciency of FL, including challenges and constraints caused by limited bandwidth and computation ability

[47] Survey on the fairness-aware FL, including client selection, optimization, contribution evaluation, incentive distribution, and performance metrics

[48] Survey on federated reinforcement learning (FRL), including deﬁnitions, evolution, and advantages of horizontal FRL and vertical FRL

[49] Survey on FL for natural language processing (NLP), including algorithm challenges, system challenges as well as privacy issues

[50], [51] Surveys on incentive schemes for FL in terms of Stackelberg game, auction, contract theory, Shapley value, reinforcement learning, and blockchain

[42] Survey on unlabeled data in FL, including potential research directions, application scenarios, and challenges

[52] FL over ethernet passive optical networks, considering dynamic wavelength and bandwidth allocation for quality of service provisioning

FL in

Wireless

Networks

[54]–[57], [64] Fundamental theories, key techniques, challenges, future directions, and applications for FL over wireless networks

[58], [61], [62] Surveys on FL in IoT, including data sharing, ofﬂoading and caching, attack detection, localization, mobile crowdsensing, and privacy

[53], [63] Surveys on communication-efﬁcient FL under various challenges incurred by communication, computing, energy, and data privacy issues

[59] Survey on FL in MEC, including communication cost, resource allocation, data privacy, data security, and implementation

[60] Survey on collaborative FL, including the deﬁnitions, advantages, drawbacks, usage conditions, and performance metrics

[65] Survey on asynchronous FL, including device heterogeneity, data heterogeneity, privacy and security, and applications

Meta Learning

Methodologies

[66], [67], [70], [71] Introductory tutorial on meta learning, e.g., deﬁnitions, models, challenges, research directions, and applications

[68] Tutorial on meta learning on algorithm selection (transfer learning, few-shot learning, and beyond supervised learning) for data mining

[69] Survey on meta learning for NLP, especially few-shot applications, including deﬁnitions, research directions, and some common datasets

[72] Tutorial on multimodality-based meta-learning in terms of the methodologies (few-shot learning and zero-shot learning) and applications

TABLE II

COMPARISON OF SURVEYS AND TUTORIALS ON FL METHODOLOGIES

Reference [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] This paper

Year 2019 2020 2021 2021 2021 2021 2021 2021 2021 2021 2021 -

Section III:

Methodologies

(III.A): Model Aggregation ✓✓✓✓✓ ✓✓✓✓ ✓

(III.B): Gradient Descent ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

(III.C): Communication Efﬁciency ✓ ✓

(III.D): Fairness ✓ ✓ ✓ ✓

(III.E): Bayesian Machinery ✓ ✓ ✓

(III.F): Clustering ✓ ✓

TABLE III

COMPARISON OF SURVEYS AND TUTORIALS ON FL OVER WIRELESS NETWORKS

Reference [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] This paper

Year 2019 2020 2020 2020 2020 2020 2020 2020 2021 2021 2021 2021 2021 -

Section

IV:

Wireless

Networks

(IV.A.1): Device

Selection ✓ ✓ ✓ ✓ ✓ ✓

(IV.A.2.a): Communication

Efﬁciency ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

(IV.A.2.b): Computation

Efﬁciency ✓ ✓ ✓ ✓

(IV.A.2.c): Energy

Efﬁciency ✓ ✓ ✓

(IV.A.2): Packet

Error ✓ ✓ ✓ ✓

(IV.A.3): Resource

Allocation ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

(IV.A.4): Asynchronous ✓ ✓ ✓ ✓ ✓ ✓

(IV.B): Over the Air

Computation ✓ ✓ ✓ ✓ ✓

ing a detailed introduction to the underlying design concepts

of the relevant algorithms. Notably, our tutorial stands out by

analyzing the relationship and evolution of these algorithms

speciﬁcally and their applications over wireless networks.

This crucial aspect has not been extensively investigated in

other surveys or tutorials. By exploring the interplay and

advancements of these algorithms and their applications over

wireless networks, our tutorial contributes novel insights and

addresses research gaps in the existing literature.

B. Summary of Contributions

Given the privacy-preserving characteristics of federated

learning (FL) and the ability of meta learning to quickly adapt

to different tasks, researchers from both academia and industry

are now exploring the joint design of FL, meta learning, and

federated meta learning (FedMeta) and wireless networks. This

paper provides the ﬁrst comprehensive tutorial that highlights

the research areas, relationships, advancements, challenges,

and opportunities associated with these three learning concepts

and their applications over wireless network environments.

The main contributions of this article can be summarized as

follows:

TABLE IV

COMPARISON OF SURVEYS AND TUTORIALS ON META LEARNING

Reference [66] [67] [68] [69] [70] [71] [72] This paper

Year 2002 2015 2018 2020 2020 2020 2021 -

Section V:

Meta

Learning

Methodologies

(V.A.1): Siamese ✓✓✓✓✓ ✓

(V.A.2): Matching ✓ ✓ ✓ ✓ ✓

(V.A.3): Prototypical ✓✓✓✓✓ ✓

(V.A.4): Relation ✓✓✓✓ ✓

(V.B.1): Memory-augmented Neural Network ✓ ✓ ✓

(V.B.2): Meta Network ✓

(V.B.3): Recurrent Meta-learner ✓ ✓

(V.B.4): Simple Neural Attentive Meta-learner ✓ ✓ ✓

(V.C.1): Model-Agnostic Meta-Learning (MAML) ✓✓✓✓✓ ✓

(V.C.2): Meta-SGD ✓ ✓ ✓

(V.C.3): Reptile ✓ ✓ ✓ ✓

(V.C.4): Bayesian MAML ✓ ✓ ✓

(V.C.5): Laplace Approximation for Meta Adaptation ✓ ✓

(V.C.6): Latent Embedding Optimization ✓ ✓

(V.C.7): MAML with Implicit Gradients ✓

(V.C.8): Online MAML ✓ ✓ ✓

Section VI:

Meta Learning

in Wireless Networks

(VI.A): Trafﬁc Prediction ✓

(VI.B): Rate Maximization ✓

(VI.C): MIMO Detectors ✓

•Based on the foundational FedAvg algorithm, we outline

six key research areas of FL methodologies, includ-

ing model aggregation, gradient descent, communication

efﬁciency, fairness, Bayesian learning, and clustering.

We provide a detailed analysis of the relationships and

evolutionary developments of the respective learning al-

gorithms within these areas. Furthermore, we introduce

two research areas that explore the interplay between FL

and wireless factors over wireless networks, including

digital and analog over-the-air computation schemes.

•We summarize three key research areas in meta-learning,

including metric-based, model-based, and gradient-based

meta-learning. Based on the gradient-based meta-learning

paradigm, we focus on the fundamental scheme known

as model-agnostic meta-learning (MAML). We discuss

the evolution of MAML, providing a detailed analysis

of its advancements and applications. Additionally, we

explore the potential of meta-learning in solving wireless

communication problems. By leveraging meta-learning

techniques, we investigate how they can be utilized to

tackle various challenges and optimize wireless commu-

nication systems.

•The fundamental principle of federated meta learning

(FedMeta) and its evolution are comprehensively sum-

marized. Furthermore, we introduce two speciﬁc research

areas of FedMeta over wireless networks, including de-

vice selection and energy efﬁciency. We explore how

FedMeta can be applied to address challenges related

to device selection, such as optimizing the selection of

devices for participation in the FL process. Additionally,

we discuss how FedMeta can contribute to enhancing

energy efﬁciency in wireless networks, thereby improving

the overall energy consumption of the system.

•In addition to the previously discussed topics, we present

several other important aspects related to the imple-

mentation of FL, meta-learning, and FedMeta. We ex-

plore different implementation platforms that facilitate

the practical deployment of these learning methodologies.

Furthermore, we present real-world applications where

FL, meta-learning, and FedMeta have demonstrated their

effectiveness and potential impact. In addition, we iden-

tify and highlight open research problems that present

exciting opportunities for future research in the ﬁeld of

FL, meta-learning, and FedMeta. These research prob-

lems have the potential to drive innovation and pave

the way for new directions and advancements in these

learning concepts. By addressing these open challenges,

researchers can contribute to the further development and

practical applications of FL, meta-learning, and FedMeta.

The rest of this paper is organized as follows. In Section

II, we introduce the background and fundamentals of dis-

tributed learning. In Sections III and IV, we present important

research ﬁelds in FL methodologies and their applications

over wireless networks, respectively. In Sections V and VI,

we introduce research areas in meta learning methodologies

and their applications over wireless networks, respectively.

Section VII presents principle of FedMeta and its applications

over wireless networks. Section VIII introduces open research

problems and future directions in FL, meta learning, and

FedMeta. A graphical illustration of the content of Sections II

to VII is provided in Fig. 2. Finally, conclusions are drawn in

Section IX.

II. BACKGROUND AND FUNDAMENTALS OF DISTRIBUTED

LEARNING

In distributed learning, two main approaches are commonly

used for learning across servers or devices: data parallelizing

and model parallelizing. In the data parallelizing approach,

the data is divided into multiple datasets, and each server or

device applies the same machine learning (ML) algorithm to a

different dataset. This approach allows for parallel processing

of data, enabling faster training and improved scalability. On

the other hand, the model parallelizing approach involves

segmenting the ML model into different sub-models. Each

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1FederatedLearningandMetaLearning:Approaches,Applications,andDirectionsXiaonanLiu,Member,IEEE,YanshaDeng,SeniorMember,IEEE,ArumugamNallanathanandMehdiBennis,Fellow,IEEEAbstract—Overthepastfewyears,significantadvancementshavebeenmadeinthefieldofmachinelearning(ML)toaddressresourcemanagement,interfere...

展开>> 收起<<

1 Federated Learning and Meta Learning Approaches Applications and Directions.pdf

共48页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Federated Learning and Meta Learning Approaches Applications and Directions

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: