國立虎尾科技大學 |

Navigating the Challenges of Imperfect Supervision for Machine Learning in the Real-World.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Navigating the Challenges of Imperfect Supervision for Machine Learning in the Real-World./
Author:	Liu, Kangning.
Description:	1 online resource (270 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
Contained By:	Dissertations Abstracts International85-12B.
Subject:	Statistics. -
Online resource:	click for full text (PQDT)
ISBN:	9798383165775

Navigating the Challenges of Imperfect Supervision for Machine Learning in the Real-World.
Liu, Kangning.

Navigating the Challenges of Imperfect Supervision for Machine Learning in the Real-World. - 1 online resource (270 pages)

Source: Dissertations Abstracts International, Volume: 85-12, Section: B.

Thesis (Ph.D.)--New York University, 2024.

Includes bibliographical references

Machine learning, particularly deep learning, has markedly revolutionized numerous domains, including computer vision, natural language processing, healthcare, and autonomous driving. The success of these technologies is deeply reliant on the availability of well-curated datasets. Yet, this requirement presents a considerable obstacle in practical applications, where different forms of imperfect supervision-such as weak, noisy, or insufficient labeling-are commonly encountered. This thesis explores innovative approaches to adapt machine learning systems to navigate these challenges. We begin by developing a novel neural network architecture aimed at improving segmentation models for high-resolution mammography images, utilizing only image-level labels. Moreover, this thesis investigates how to adapt self-supervised pretrained models to downstream tasks under weak supervision. Despite lacking specific instance labels, the enhanced model excels in generating high-quality instance-level representations within a multiple instance learning framework. Further exploration is conducted on the learning dynamics of deep segmentation models trained with noisy annotations, leading to the design of a class-specific corrective algorithm. This method progressively corrects erroneously labeled areas in the annotations, significantly advancing model accuracy. With the emergence of foundation models, the machine learning community has witnessed a paradigm shift to data-centric AI approaches, where designing scalable ways to address imperfect supervision is crucial. Lastly, we reassess the training methodology of the Segment Anything Model, a foundational segmentation model that relies on a vast amount of unlabeled data with minimal human supervision. Given the inherent challenges posed by potentially inaccurate model-generated pseudo-labels, we propose a new framework that quantifies the uncertainty in these labels. This framework leverages the quantified uncertainty to perform uncertainty-aware fine-tuning, substantially improving model performance.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798383165775Subjects--Topical Terms:

556824
Statistics.
Subjects--Index Terms:

Machine learningIndex Terms--Genre/Form:

554714
Electronic books.

Navigating the Challenges of Imperfect Supervision for Machine Learning in the Real-World.
LDR:03467ntm a22003857 4500 001 1150149
005 20241022111559.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798383165775
035 $a (MiAaPQ)AAI30997229
035 $a AAI30997229
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Liu, Kangning. $3 1476583
245 1 0 $a Navigating the Challenges of Imperfect Supervision for Machine Learning in the Real-World.
264 0 $c 2024
300 $a 1 online resource (270 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: 85-12, Section: B.
500 $a Advisor: Fernandez-Granda, Carlos;Geras, Krzysztof Jerzy.
502 $a Thesis (Ph.D.)--New York University, 2024.
504 $a Includes bibliographical references
520 $a Machine learning, particularly deep learning, has markedly revolutionized numerous domains, including computer vision, natural language processing, healthcare, and autonomous driving. The success of these technologies is deeply reliant on the availability of well-curated datasets. Yet, this requirement presents a considerable obstacle in practical applications, where different forms of imperfect supervision-such as weak, noisy, or insufficient labeling-are commonly encountered. This thesis explores innovative approaches to adapt machine learning systems to navigate these challenges. We begin by developing a novel neural network architecture aimed at improving segmentation models for high-resolution mammography images, utilizing only image-level labels. Moreover, this thesis investigates how to adapt self-supervised pretrained models to downstream tasks under weak supervision. Despite lacking specific instance labels, the enhanced model excels in generating high-quality instance-level representations within a multiple instance learning framework. Further exploration is conducted on the learning dynamics of deep segmentation models trained with noisy annotations, leading to the design of a class-specific corrective algorithm. This method progressively corrects erroneously labeled areas in the annotations, significantly advancing model accuracy. With the emergence of foundation models, the machine learning community has witnessed a paradigm shift to data-centric AI approaches, where designing scalable ways to address imperfect supervision is crucial. Lastly, we reassess the training methodology of the Segment Anything Model, a foundational segmentation model that relies on a vast amount of unlabeled data with minimal human supervision. Given the inherent challenges posed by potentially inaccurate model-generated pseudo-labels, we propose a new framework that quantifies the uncertainty in these labels. This framework leverages the quantified uncertainty to perform uncertainty-aware fine-tuning, substantially improving model performance.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Statistics. $3 556824
650 4 $a Computer science. $3 573171
653 $a Machine learning
653 $a Imperfect supervision
653 $a Segmentation models
653 $a Segment Anything Model
655 7 $a Electronic books. $2 local $3 554714
690 $a 0463
690 $a 0984
690 $a 0800
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a New York University. $b Center for Data Science. $3 1468521
773 0 $t Dissertations Abstracts International $g 85-12B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30997229 $z click for full text (PQDT)