자료유형  

 학위논문

Control Number  

0017164865

International Standard Book Number  

9798346389989

Dewey Decimal Classification Number  

519.54

Main Entry-Personal Name  

Ethayarajh, Kawin.

Publication, Distribution, etc. (Imprint  

[S.l.] : Stanford University., 2024

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Physical Description  

176 p.

General Note  

Source: Dissertations Abstracts International, Volume: 86-05, Section: B.

General Note  

Advisor: Jurafsky, Dan.

Dissertation Note  

Thesis (Ph.D.)--Stanford University, 2024.

Summary, Etc.  

요약Machine learning is often characterized as a sterile process, but in reality, it is shaped by workers, firms, states, and many other actors. In fact, it is often the behavior of these actors that principally determines what machine learning can accomplish in the real world, even more so than traditional concerns like the amount of memory or compute. Therefore, I propose that much in the way we characterize processes as being compute-bound or memory-bound, we should also think of them as being behavior-bound. By then formalizing real-world behavior, we can create machine learning pipelines that are compatible with actual actors and not just idealized ones. For inspiration, I turn to economics, a field that has undergone a similar revolution over the past half-century, moving away from a conception of people as perfectly rational agents making optimal decisions to studying the complex reality of how they actually behave. I first explain how conflicting interests between actors leads to datasets being much simpler than the problems they purport to reflect, and how we can diagnose this using the theory of usable information. I then use this framework to create one of the largest datasets for aligning large language models with human feedback. The next chapter argues that the methods used for aligning generative models are uniquely effective because they capture systematic biases in the way that humans themselves make decisions, and that human-aware alignment methods can be cheaper, faster, and more robust. Lastly, I present an evaluation-as-a-service platform that evaluates these models in a cost-sensitive way instead of treating them as if they exist in a vacuum. Together, these contributions lay the foundation for conceptualizing machine learning as a part of a larger sociotechnical whole.

Subject Added Entry-Topical Term  

Standard scores.

Subject Added Entry-Topical Term  

Large language models.

Subject Added Entry-Topical Term  

Statistics.

Added Entry-Corporate Name  

Stanford University.

Host Item Entry  

Dissertations Abstracts International. 86-05B.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:654313