time: Mon/Wed 3:30-5pm

location: AGH 105

instructors: Osbert Bastani and Rajeev Alur

teaching assistants: Seewon Choi and Avishree Khare

collaboration policy: You are responsible for knowing Penn's Code of Academic Integrity. In particular, copying solutions from other students or other resources (e.g. the web or from students who have taken the class in previous years) is NOT allowed. Making answers to homeworks or exams available to others either directly or by posting on the web is also NOT allowed. We will not have a sense of humor about violations of this policy!

links: We will use Ed Discussion for questions and communication, and GradeScope to submit assignments. We encourage students to use Google Colab for coding assignments.

attendance: We expect students to attend classes regularly. However, please do not come to class if you are not feeling well or test positive for Covid-19. We will do our best to provide lecture slides (on this website) for students unable to make it to class.

description: Recent advances in machine learning---in particular deep neural networks and large language models, are transforming the design and implementation of decision making systems. However, due to their black-box nature, brittleness, and lack of safety guarantees, significant challenges remain in their adoption in critical and potentially high payoff applications such as autonomous systems and healthcare. The vibrant field of "Trustworthy ML" is developing methods and tools to address questions such as: how can we ensure that a decision recommended by an ML system is always safe? how can we explain the decision made by an ML system to a stakeholder? how can we ensure that an ML system makes its decisions in a fair and ethical manner? The goal of this course is to introduce students to state-of-the-art research in trustworthy ML.

topics: Here is a tentative list of topics: (i) adversarial and distributional robustness, (ii) uncertainty quantification, (iii) interpretability and explainability, and (iv) AI alignment. Each section will include a mix of broadly applicable techniques, as well as how these techniques specialize to Generative AI. See the course schedule for additional details.

prerequisites: The main prerequisite is either CIS 5200 (Machine Learning), or CIS 4190/5190 (Applied Machine Learning) AND CIS 3333 (Mathematics of Machine Learning). The course requires both mathematical maturity, including experience with mathematical proofs, and familiarity with machine learning libraries. It is appropriate for students who wish to pursue research in machine learning.

The following is a tentative grading scheme:

homework (30%): There will be 2-3 homework assignments, including some combination of written theory questions and coding questions. The written portion will be submitted via GradeScope; the coding portion will be submitted both via GradeScope and programmatically to an autograder. Detailed instructions will be provided in the assignment.

midterm exams (40%): There will be a two midterm exams, each covering each half of the course content and worth 20%.

project (30%): We will provide more details on the final project shortly.