Instructor: Zachary Lipton (Director of ACMI lab)
Philosopher: Mel Andrews
TA: Charvi Rastogi
Time: Thursdays 3:30pm - 5:50pm
Location: GHC 4215 (in person)
Reading List: see schedule

What is this field? What are its normative aims? What are its modes of inquiry? What are (and have been) its intellectual and ideological commitments? What foundational questions is it in dialogue with, and what foundational obstacles obstruct its progress? Finally: What are our responsibilities as researchers & practitioners deploying this technology?

The pursuit of machine intelligence holds philosophical significance, both because of the field’s own philosophical commitments, because it seeks (knowingly or not) to resolve questions of longstanding philosophical interest, and because of its role as an object of interest to philosophers of science, mind, and ethics. No field can completely avoid its philosophical questions. Whenever reviewers argue about whether a paper is "in scope" they are engaged in the conceptual question of "what is machine learning?" Whenever authors argue that a method is well-motivated they make assertions about what problems are important, and what are sound/plausible ways to idealize them. However, these matters are seldom addressed head-on and play little role in the formal training of ML's practitioners. They are absent from machine learning curricula, seldom addressed in machine learning’s conference proceedings, and often visible to only a small subset of machine learning scientists. Moreover, while its easy to see the career benefit to prominent scientists who engage in the philosophical debates of their times, it can be harder for junior researchers to see the incentive to pursue these contributions. The disconnect between these conceptual questions and technical work in the field has consequences: too many papers purport to address questions that they do not (or cannot) address, and a burgeoning subfield of AI ethics surprisingly out of touch with most basic concerns of ethics; a gap so extreme that many philosophers (reasonably) do not recognize large swaths of the AI ethics literature as being engaged in ethics research of any kind. This is an exploratory seminar-style course aimed at inlining the philosophical problems surrounding machine intelligence into the machine learning curriculum.

In this course, we will address the origins of the field through the foundational writings of (e.g.,) Turing, Weiner, McCarthy, Simon, Minsky, Vapnik, Rumelhart/McClelland/Hinton, Pearl, etc. We will address the fundamental problem of learning from observation, including both the problem of induction (setting Popper in dialogue with Vapnik and Wolpert) and the formation, evolution, and abandonment of concepts/kinds/theories (e.g.,) through the writings of (e.g.,) Kuhn, Hacking, Hofstadter. We will address the very technical language used to formulate our inquiry, including probability and causality through (e.g.,) Polya, Cox, Cartwright, Pearl, Halpern. And we will discuss the ethical questions associated with deploying data driven models to automate decisions in consequential domains. We may also focus on some especially timely topics, like ML's generative turn and attendant questions about what it means for a model to know something or to be creative.

There are no formal prerequisites for this class. It is open to PhD students and MS and undergraduate students may enroll with permission from the instructor. The course will not involve much theorem proving or engineering but will be considerably more reading-intensive than a typical computer science course. Students are expected to keep up with each week’s readings, write thoughtful short responses digesting the main points and relating them to modern practice in the field, and to lead or co-lead the presentation/discussion of at least one reading during the semester.

Schedule of readings

• Every participant will be required to present papers, we will rotate among the enrolled / participating students and instructors with each participant presenting roughly # of papers / # of people papers over the course of the semester (roughly 3).
• Every student will be required to do all the readings each week and to preregister a list of 3 questions before each session (links to each week's question form are availabel on the schedule page. Students are allowed to miss 3 weeks no questions asked, beyond that failure to register questions or attend class may impact grades.
• Grading will be based on participation, attendance, delivering thoughtful questions and will mostly be generous for students participating in good faith and doing the readings. There will be no exams. Students can optionally write a paper and use time after class to get feedback.