서브메뉴
검색
Efficient and Reliable Optimization for Deep Learning and Media Generation
Efficient and Reliable Optimization for Deep Learning and Media Generation
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103529
- ISBN
- 9798288866272
- DDC
- 004
- 저자명
- Bai, Yatong.
- 서명/저자
- Efficient and Reliable Optimization for Deep Learning and Media Generation
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 297 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Sojoudi, Somayeh.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약This dissertation develops new methodologies for improving the safety, efficiency, and alignment of deep learning models. Among this broad topic, we focus on four facets of understanding model robustness and enhancing media generation.We first address optimization challenges of neural networks arising from their non-convexity and the bi-level min-max formulation necessitated by robust training methods. Unlike convex optimization problems, which can be efficiently solved to global optimality, the complicated neural network training formulations often become stuck at spurious local optima. By navigating the challenging loss landscapes via surrogate convex training optimization problems with provable global convergence, we enable tractable learning with global optimality guarantees. We customize efficient optimization algorithms for convex training and extend the framework to the adversarial training problem, simultaneously ensuring robustness, explainability, and training speed.While working on the optimization challenges, we identified robust neural classifiers' generalization limitations as an equally critical issue -- the accuracy-robustness trade-off makes robust models from academia unattractive for practitioners to implement, leaving real-world systems unsafe. To this end, we propose a flexible mixed classifier framework and develop plug-and-play methods like adaptive smoothing and MixedNUTS. These methods mix the output probabilities of a robust model and an accurate (generally non-robust) model, leveraging the benign confidence property of robust classifiers to balance accuracy and robustness. Our methods assume both base classifiers to be already trained, thus compatible with other advancements in the field. Their state-of-the-art accuracy-robustness balance incentivizes practical deployment of robust models.Then, we analyze the vulnerability of large language models (LLMs) coupled with content retrieval systems to form conversational search engines, focusing on the e-commerce scenario of product promotion. We show that when LLMs are fed with fetched product websites, their recommendation order depends on product name (prior knowledge), website content (excluding the name), and the website input order. Despite the intertwined influences of these factors, we can force the model to recommend a product at the top by embedding algorithmically determined adversarial strings into the product website source code. This observation unveils an influential yet lucrative backdoor of LLMs, calling for further research on LLM robustness.Next, we shift focus to media generation, where diffusion models -- the "workhorses" of the field -- suffer from a mismatch between training objective (denoise) and target goal (creative generation). Not only are they misaligned with the target goal reward functions, but they also require a painfully slow iterative denoising inference process. To tackle this inefficiency and unreliability, we propose ConsistencyTTA to distill diffusion models' iterative inference procedure into a single model pass.In addition to ensuring efficiency, this non-recursiveness enables end-to-end fine-tuning to align with the target goal rewards.Finally, we solve the misalignment from another angle and develop DRAGON, a general-purpose framework that optimizes media creation toward desired outcomes. DRAGON is compatible with reward functions that evaluate generation individuals or distributions. We leverage this flexibility to propose exemplar-based reward functions, with which DRAGON enhances human-perceived media generation quality without relying on human preference annotations or additional high-quality data.Together, these contributions advance the efficiency and dependability of modern deep learning, particularly in settings where training does not fully model deployment scenarios.
- 일반주제명
- Computer science
- 일반주제명
- Engineering
- 키워드
- Audio generation
- 키워드
- Diffusion model
- 키워드
- Neural network
- 기타저자
- University of California, Berkeley Mechanical Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357559
■00520260202103529
■006m o d
■007cr#unu||||||||
■020 ▼a9798288866272
■035 ▼a(MiAaPQ)AAI32039737
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aBai, Yatong.
■24510▼aEfficient and Reliable Optimization for Deep Learning and Media Generation
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a297 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Sojoudi, Somayeh.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aThis dissertation develops new methodologies for improving the safety, efficiency, and alignment of deep learning models. Among this broad topic, we focus on four facets of understanding model robustness and enhancing media generation.We first address optimization challenges of neural networks arising from their non-convexity and the bi-level min-max formulation necessitated by robust training methods. Unlike convex optimization problems, which can be efficiently solved to global optimality, the complicated neural network training formulations often become stuck at spurious local optima. By navigating the challenging loss landscapes via surrogate convex training optimization problems with provable global convergence, we enable tractable learning with global optimality guarantees. We customize efficient optimization algorithms for convex training and extend the framework to the adversarial training problem, simultaneously ensuring robustness, explainability, and training speed.While working on the optimization challenges, we identified robust neural classifiers' generalization limitations as an equally critical issue -- the accuracy-robustness trade-off makes robust models from academia unattractive for practitioners to implement, leaving real-world systems unsafe. To this end, we propose a flexible mixed classifier framework and develop plug-and-play methods like adaptive smoothing and MixedNUTS. These methods mix the output probabilities of a robust model and an accurate (generally non-robust) model, leveraging the benign confidence property of robust classifiers to balance accuracy and robustness. Our methods assume both base classifiers to be already trained, thus compatible with other advancements in the field. Their state-of-the-art accuracy-robustness balance incentivizes practical deployment of robust models.Then, we analyze the vulnerability of large language models (LLMs) coupled with content retrieval systems to form conversational search engines, focusing on the e-commerce scenario of product promotion. We show that when LLMs are fed with fetched product websites, their recommendation order depends on product name (prior knowledge), website content (excluding the name), and the website input order. Despite the intertwined influences of these factors, we can force the model to recommend a product at the top by embedding algorithmically determined adversarial strings into the product website source code. This observation unveils an influential yet lucrative backdoor of LLMs, calling for further research on LLM robustness.Next, we shift focus to media generation, where diffusion models -- the "workhorses" of the field -- suffer from a mismatch between training objective (denoise) and target goal (creative generation). Not only are they misaligned with the target goal reward functions, but they also require a painfully slow iterative denoising inference process. To tackle this inefficiency and unreliability, we propose ConsistencyTTA to distill diffusion models' iterative inference procedure into a single model pass.In addition to ensuring efficiency, this non-recursiveness enables end-to-end fine-tuning to align with the target goal rewards.Finally, we solve the misalignment from another angle and develop DRAGON, a general-purpose framework that optimizes media creation toward desired outcomes. DRAGON is compatible with reward functions that evaluate generation individuals or distributions. We leverage this flexibility to propose exemplar-based reward functions, with which DRAGON enhances human-perceived media generation quality without relying on human preference annotations or additional high-quality data.Together, these contributions advance the efficiency and dependability of modern deep learning, particularly in settings where training does not fully model deployment scenarios.
■590 ▼aSchool code: 0028.
■650 4▼aComputer science
■650 4▼aEngineering
■653 ▼aAdversarial robustness
■653 ▼aAudio generation
■653 ▼aConvex optimization
■653 ▼aDiffusion model
■653 ▼aLarge language model
■653 ▼aNeural network
■690 ▼a0984
■690 ▼a0796
■690 ▼a0537
■71020▼aUniversity of California, Berkeley▼bMechanical Engineering.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357559▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


