東華大學圖書館 |

Language: English

Help

回圖書館首頁

手機版館藏查詢

Back

Switch To: Labeled | MARC Mode | ISBD

Linked to FindBook

Google Book

Amazon

博客來

Over-The-Air Statistical Estimation.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Over-The-Air Statistical Estimation./
Author:	Lee, Chuan-Zheng.
Published:	Ann Arbor : ProQuest Dissertations & Theses, : 2021,
Description:	106 p.
Notes:	Source: Dissertations Abstracts International, Volume: 83-05, Section: B.
Contained By:	Dissertations Abstracts International83-05B.
Subject:	Wireless networks. -
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28827987
ISBN:	9798494462114

Over-The-Air Statistical Estimation.
Lee, Chuan-Zheng.

Over-The-Air Statistical Estimation. - Ann Arbor : ProQuest Dissertations & Theses, 2021 - 106 p.

Source: Dissertations Abstracts International, Volume: 83-05, Section: B.

Thesis (Ph.D.)--Stanford University, 2021.

This item must not be sold to any third party vendors.

The data fueling today's rise in machine learning is often generated by devices at the edge of a network, like sensors or mobile devices. To use all this data to train a common model, devices need to communicate something about their own data to a central server. But physical communication channels have limits-and these constraints are increasingly becoming the bottleneck in distributed and federated learning systems. Can we improve such learning algorithms by explicitly incorporating the physical communication layer into their design?In this thesis, we explore this question using a new framework that draws on wireless communication theory and statistical estimation. We consider distributed minimax statistical estimation over a Gaussian multiple-access channel (MAC) under squared error loss. First, we develop "analog" joint estimation-communication schemes that leverage the additive nature of the Gaussian MAC, and we characterize their minimax risk for two estimation tasks: the Gaussian location and product Bernoulli models. We then compare this risk to lower bounds for risk in digital schemes, in which nodes transmit bits noiselessly at the Shannon capacity. We show that in both cases, the analog approach brings about exponentially smaller estimation error than the digital one. This suggests that drastic improvements in performance can be obtained by using analog schemes that work in tandem with the physical layer, rather than digital schemes using a physical abstraction.Second, having established that analog schemes can beat digital lower bounds, we derive more general lower bounds for minimax statistical estimation in this setting. These "analog" lower bounds are within a logarithmic factor of the aforementioned achievability results, showing at most a small gap between our analog schemes and the optimal ones.Third, we implement these ideas in federated machine learning experiments, considering the aggregation step of federated averaging to be a distributed statistical estimation problem of the type we are studying. We train logistic regression and image classification tasks, using both analog and digital estimation schemes in the aggregation step. Our results show that the ideas from our theoretical analysis can translate to performance gains in a federated learning context.Finally, we develop analog schemes for sparse Gaussian location and sparse product Bernoulli models, for the setting where there are much fewer channel uses available than the dimension of the samples. These analog schemes use compressed sensing to reduce the number of channel uses needed, while still exploiting the superposition in the Gaussian MAC. Like their dense counterparts, when compared to lower bounds for digital schemes, these analog schemes bring similarly substantial reductions in estimation error. This bolsters our finding that analog schemes that design the estimation and communication protocols jointly, vastly outperform all digital schemes that use a physical abstraction.

ISBN: 9798494462114Subjects--Topical Terms:

1531264
Wireless networks.

Over-The-Air Statistical Estimation.
LDR:04074nmm a2200361 4500 001 2346916
005 20220706051322.5
008 241004s2021 ||||||||||||||||| ||eng d
020 $a 9798494462114
035 $a (MiAaPQ)AAI28827987
035 $a (MiAaPQ)STANFORDvj749dt7188
035 $a AAI28827987
040 $a MiAaPQ $c MiAaPQ
100 1 $a Lee, Chuan-Zheng. $3 3686122
245 1 0 $a Over-The-Air Statistical Estimation.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2021
300 $a 106 p.
500 $a Source: Dissertations Abstracts International, Volume: 83-05, Section: B.
500 $a Advisor: Ozgur, Ayfer.
502 $a Thesis (Ph.D.)--Stanford University, 2021.
506 $a This item must not be sold to any third party vendors.
520 $a The data fueling today's rise in machine learning is often generated by devices at the edge of a network, like sensors or mobile devices. To use all this data to train a common model, devices need to communicate something about their own data to a central server. But physical communication channels have limits-and these constraints are increasingly becoming the bottleneck in distributed and federated learning systems. Can we improve such learning algorithms by explicitly incorporating the physical communication layer into their design?In this thesis, we explore this question using a new framework that draws on wireless communication theory and statistical estimation. We consider distributed minimax statistical estimation over a Gaussian multiple-access channel (MAC) under squared error loss. First, we develop "analog" joint estimation-communication schemes that leverage the additive nature of the Gaussian MAC, and we characterize their minimax risk for two estimation tasks: the Gaussian location and product Bernoulli models. We then compare this risk to lower bounds for risk in digital schemes, in which nodes transmit bits noiselessly at the Shannon capacity. We show that in both cases, the analog approach brings about exponentially smaller estimation error than the digital one. This suggests that drastic improvements in performance can be obtained by using analog schemes that work in tandem with the physical layer, rather than digital schemes using a physical abstraction.Second, having established that analog schemes can beat digital lower bounds, we derive more general lower bounds for minimax statistical estimation in this setting. These "analog" lower bounds are within a logarithmic factor of the aforementioned achievability results, showing at most a small gap between our analog schemes and the optimal ones.Third, we implement these ideas in federated machine learning experiments, considering the aggregation step of federated averaging to be a distributed statistical estimation problem of the type we are studying. We train logistic regression and image classification tasks, using both analog and digital estimation schemes in the aggregation step. Our results show that the ideas from our theoretical analysis can translate to performance gains in a federated learning context.Finally, we develop analog schemes for sparse Gaussian location and sparse product Bernoulli models, for the setting where there are much fewer channel uses available than the dimension of the samples. These analog schemes use compressed sensing to reduce the number of channel uses needed, while still exploiting the superposition in the Gaussian MAC. Like their dense counterparts, when compared to lower bounds for digital schemes, these analog schemes bring similarly substantial reductions in estimation error. This bolsters our finding that analog schemes that design the estimation and communication protocols jointly, vastly outperform all digital schemes that use a physical abstraction.
590 $a School code: 0212.
650 4 $a Wireless networks. $3 1531264
650 4 $a Communication. $3 524709
650 4 $a Algorithms. $3 536374
650 4 $a Ethernet. $3 3683702
650 4 $a Electrical engineering. $3 649834
650 4 $a Multiuser systems. $3 3686123
650 4 $a Parameter estimation. $3 567557
650 4 $a Artificial intelligence. $3 516317
650 4 $a Computer science. $3 523869
650 4 $a Engineering. $3 586835
650 4 $a Information science. $3 554358
690 $a 0544
690 $a 0459
690 $a 0800
690 $a 0984
690 $a 0537
690 $a 0723
710 2 $a Stanford University. $3 754827
773 0 $t Dissertations Abstracts International $g 83-05B.
790 $a 0212
791 $a Ph.D.
792 $a 2021
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28827987