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

Algorithmic Strategies for Learning ...

Saha, Rajarshi,

Linked to FindBook

Google Book

Amazon

博客來

Algorithmic Strategies for Learning and Inference in Resource-Constrained Environments /

Record Type:	Electronic resources : Monograph/item
Title/Author:	Algorithmic Strategies for Learning and Inference in Resource-Constrained Environments // Rajarshi Saha.
Author:	Saha, Rajarshi,
Description:	1 electronic resource (340 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
Contained By:	Dissertations Abstracts International86-03B.
Subject:	Decomposition. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31460356
ISBN:	9798384345121

Algorithmic Strategies for Learning and Inference in Resource-Constrained Environments /
Saha, Rajarshi,

Algorithmic Strategies for Learning and Inference in Resource-Constrained Environments /Rajarshi Saha. - 1 electronic resource (340 pages)

Source: Dissertations Abstracts International, Volume: 86-03, Section: B.

Machine Learning (ML) algorithms require significant computation, communication and memory resources, which is challenging to provide in many environments. Such resource-constrained environments include edge computing devices such as smartphones and Internet-of-Things (IoT) devices, remote-sensing equipment on drones and satellites, and healthcare monitoring wearables. The unique challenges these systems present necessitate solutions that can efficiently process data despite constraints on power, memory, and connectivity. This thesis tackles these significant challenges by concentrating on two critical areas: (i) ML subject to significant computational limitations such as those on edge devices, and (ii) the impediments to distributed training of ML models due to restricted communication capabilities.Firstly, this thesis investigates the potential of low-precision modeling techniques for edge inference, employing randomized embeddings to achieve efficient model compression. Utilizing randomized embeddings significantly reduces the computational demands of deploying high-dimensional models on edge devices, which are typically constrained by limited processing power and memory. The proposed algorithms are designed to maintain an acceptable level of accuracy, ensuring that the performance of the models do not degrade despite the reduction in computational complexity. This is critical to ensure that the trade-off between performance and efficiency are carefully balanced to achieve practical deployment. Furthermore, the information-theoretic optimality of the proposed algorithms is determined, and it is shown that randomized-embedding based algorithms are minimax optimal up to constant factors for statistical estimation tasks under bit-budget constraints. Through detailed experiments and theoretical analyses, this thesis demonstrates how randomized low-precision techniques provide a viable solution for enhancing the scalability and usability of machine learning applications in resource-constrained settings.Secondly, this thesis explores the challenges associated with distributed training over unreliable communication networks marked by intermittent connectivity. Collaborative relaying is proposed as a solution to enhance the reliability of data transmission and ensure data privacy despite the unpredictable nature of the communication links. Furthermore, to counteract additional sources of network unreliability, such as noisy and bandwidth-limited communication between agents, this thesis adapts popular distributed training algorithms, including stochastic gradient descent and lazy mirror descent. These adaptations incorporate strategies like power control and randomized-embedding-based gradient quantization to mitigate adverse effects. Collectively, these strategies effectively address the critical challenges of ensuring robust and resilient model training in unreliable distributed environments, while minimizing overhead.

English

ISBN: 9798384345121Subjects--Topical Terms:

3561186
Decomposition.

Algorithmic Strategies for Learning and Inference in Resource-Constrained Environments /
LDR:04330nmm a22003853i 4500 001 2434734
005 20250930134351.5
006 m o d
007 cr|nu||||||||
008 261011s2024 miu||||||m |||||||eng d
020 $a 9798384345121
035 $a (MiAaPQD)AAI31460356
035 $a (MiAaPQD)Stanfordsr466pj3629
035 $a AAI31460356
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Saha, Rajarshi, $e author. $3 3824814
245 1 0 $a Algorithmic Strategies for Learning and Inference in Resource-Constrained Environments / $c Rajarshi Saha.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 1 electronic resource (340 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
500 $a Advisors: Goldsmith, Andrea; Pilanci, Mert Committee members: Ozgur, Ayfer; Weissman, Tsachy.
502 $b Ph.D. $c Stanford University $d 2024.
520 $a Machine Learning (ML) algorithms require significant computation, communication and memory resources, which is challenging to provide in many environments. Such resource-constrained environments include edge computing devices such as smartphones and Internet-of-Things (IoT) devices, remote-sensing equipment on drones and satellites, and healthcare monitoring wearables. The unique challenges these systems present necessitate solutions that can efficiently process data despite constraints on power, memory, and connectivity. This thesis tackles these significant challenges by concentrating on two critical areas: (i) ML subject to significant computational limitations such as those on edge devices, and (ii) the impediments to distributed training of ML models due to restricted communication capabilities.Firstly, this thesis investigates the potential of low-precision modeling techniques for edge inference, employing randomized embeddings to achieve efficient model compression. Utilizing randomized embeddings significantly reduces the computational demands of deploying high-dimensional models on edge devices, which are typically constrained by limited processing power and memory. The proposed algorithms are designed to maintain an acceptable level of accuracy, ensuring that the performance of the models do not degrade despite the reduction in computational complexity. This is critical to ensure that the trade-off between performance and efficiency are carefully balanced to achieve practical deployment. Furthermore, the information-theoretic optimality of the proposed algorithms is determined, and it is shown that randomized-embedding based algorithms are minimax optimal up to constant factors for statistical estimation tasks under bit-budget constraints. Through detailed experiments and theoretical analyses, this thesis demonstrates how randomized low-precision techniques provide a viable solution for enhancing the scalability and usability of machine learning applications in resource-constrained settings.Secondly, this thesis explores the challenges associated with distributed training over unreliable communication networks marked by intermittent connectivity. Collaborative relaying is proposed as a solution to enhance the reliability of data transmission and ensure data privacy despite the unpredictable nature of the communication links. Furthermore, to counteract additional sources of network unreliability, such as noisy and bandwidth-limited communication between agents, this thesis adapts popular distributed training algorithms, including stochastic gradient descent and lazy mirror descent. These adaptations incorporate strategies like power control and randomized-embedding-based gradient quantization to mitigate adverse effects. Collectively, these strategies effectively address the critical challenges of ensuring robust and resilient model training in unreliable distributed environments, while minimizing overhead.
546 $a English
590 $a School code: 0212
650 4 $a Decomposition. $3 3561186
650 4 $a Codes. $3 3560019
650 4 $a Error analysis. $3 3562845
650 4 $a Neural networks. $3 677449
650 4 $a Linear algebra. $3 2923381
650 4 $a Mathematics. $3 515831
690 $a 0800
690 $a 0405
710 2 $a Stanford University. $e degree granting institution. $3 3765820
720 1 $a Goldsmith, Andrea $e degree supervisor.
720 1 $a Pilanci, Mert $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-03B.
790 $a 0212
791 $a Ph.D.
792 $a 2024
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31460356