본문

Safe Machine Learning-Based Perception Via Closed-Loop Analysis- [electronic resource]
Safe Machine Learning-Based Perception Via Closed-Loop Analysis - [electronic resource]
내용보기
Safe Machine Learning-Based Perception Via Closed-Loop Analysis- [electronic resource]
자료유형  
 학위논문파일 국외
최종처리일시  
20240214101625
ISBN  
9798380470414
DDC  
629.13309
저자명  
Katz, Sydney Michelle.
서명/저자  
Safe Machine Learning-Based Perception Via Closed-Loop Analysis - [electronic resource]
발행사항  
[S.l.]: : Stanford University., 2023
발행사항  
Ann Arbor : : ProQuest Dissertations & Theses,, 2023
형태사항  
1 online resource(171 p.)
주기사항  
Source: Dissertations Abstracts International, Volume: 85-04, Section: B.
주기사항  
Advisor: Kochenderfer, Mykel.
학위논문주기  
Thesis (Ph.D.)--Stanford University, 2023.
사용제한주기  
This item must not be sold to any third party vendors.
초록/해제  
요약Machine learning has the potential to unlock a variety of new capabilities for the automation of safety-critical systems. For example, it can be used in the aviation domain to automate tasks such as aircraft taxiing and aircraft collision avoidance. It can also be used in the driving domain for tasks such as scene recognition and lane detection. Because these systems are safety-critical, their automation tends to be heavily regulated and requires signifcant validation efort. However, machine learning-based systems pose new challenges in safety validation due to their complex nature. For this reason, the early adoption of machine learning into these systems will likely only involve tasks that cannot be solved entirely by traditional automation techniques. This thesis focuses on the task of perception.Perception systems use sensor data to estimate information about the state of the world around them. To increase automation to the point where human operators are no longer required, perception systems will need to replace visual tasks typically performed by the operator by processing high-dimensional sensor data from cameras or lidar units. Machine learning techniques are well-suited for this task, but they add complexity to the safety validation process. To address this added complexity, the contributions of this thesis focus on the safe design and formal verifcation of machine learning-based perception systems. Each contribution makes use of closed-loop analysis, which allows us to ensure that machine learning systems meet the high-level safety requirements of the systems in which they operate.Our frst two contributions relate to the safe design of machine learning-based perception systems. Specifcally, we develop techniques to translate high-level, closed-loop safety properties to perception system design requirements. We frst propose a technique for risk-driven design of perception systems, which accounts for the efect of perceptual errors on the performance of the fully integrated, closedloop system. We show how we can use our approach during perception system training and data collection to design safer perception systems. Next, we develop an approach to efciently determine perception system performance requirements given a high-level safety property and a black-box simulator of the closed-loop system. We combine elements of common black-box estimation techniques such as Gaussian processes and multi-armed bandits. While the safe design methods we develop encourage safe behavior of machine learning models, they do not provide guarantees. Once designed, the perception systems should still be put through additional testing and safety validation using techniques such as formal verifcation.Our fnal two contributions apply to the formal verifcation of machine learningbased perception systems. We frst develop a method to extend existing closed-loop neural network verifcation techniques to provide probabilistic safety guarantees on systems operating in stochastic environments. While this approach does not apply directly to perception systems, it provides a foundation for our fnal contribution related to the verifcation of machine learning-based perception systems. In particular, we develop an approach to extend existing state-based neural network verifcation techniques to work with image-based neural networks. We demonstrate the techniques developed in this thesis on the realistic examples of vision-based taxi navigation and vision-based aircraft collision avoidance.
일반주제명  
Aviation.
일반주제명  
Design.
일반주제명  
Collisions.
일반주제명  
Neural networks.
기타저자  
Stanford University.
기본자료저록  
Dissertations Abstracts International. 85-04B.
기본자료저록  
Dissertation Abstract International
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.
신착도서 더보기
최근 3년간 통계입니다.

소장정보

  • 예약
  • 소재불명신고
  • 나의폴더
  • 우선정리요청
  • 비도서대출신청
  • 야간 도서대출신청
소장자료
등록번호 청구기호 소장처 대출가능여부 대출정보
TF07306 전자도서
마이폴더 부재도서신고 비도서대출신청

* 대출중인 자료에 한하여 예약이 가능합니다. 예약을 원하시면 예약버튼을 클릭하십시오.

해당 도서를 다른 이용자가 함께 대출한 도서

관련 인기도서

로그인 후 이용 가능합니다.