IFT 6168, Winter 2025
Causal Inference and ML
Announcements
- We will use Piazza for all communication in this course. All students will receive a link to join Piazza either by Studium or via email (for auditing) by end of day on Monday, Jan 13.
- Until all students have joined Piazza, I'll post all key announcements here so please check this webpage frequently.
- For Jan 15, there is a short assigned reading; see the reading list at the bottom of this page. You'll turn in a very lightweight reading report, with just your questions or 1-2 bullet points or sentences about what you learned from the reading.
Summary
This course combines lectures and seminar-style discussions to understand causality and its role in modern machine learning (ML). Causal inference -- the field of statistics tasked with inferring the effects of actions and policies (i.e., causes) -- has long propelled fields like epidemiology, medicine, and economics, even leading to Nobel prizes (for example, in 2021).
More recently, ML and causality are increasingly connected: causal models parameterized by neural networks offer attractive robust generalization guarantees; ideas from causality such as interventions enable interpreting large deep learning models; causal models over latent variables give rise to interpretable and controllable deep generative models.
The main goal of this course is to give students the tools to think critically about causal modeling and draw connections to research, in ML as well as in other scientific disciplines. As such, we'll begin with fundamentals on causal models, formalizing causal graphical models, interventions, structural causal models, counterfactuals, and inference in these models. Next, we'll cover key results and methods for causal discovery and causal representation learning, challenging problems in causal inference that are addressed with non-iid data. Then, we'll explore more recent advances in the field of causal ML, covering causal techniques in interpretability research, amortized inference for causal modeling, and more.
The structure of course mixes lectures, student-led paper discussions, reading reports, a small coding assignment and a final project.
More recently, ML and causality are increasingly connected: causal models parameterized by neural networks offer attractive robust generalization guarantees; ideas from causality such as interventions enable interpreting large deep learning models; causal models over latent variables give rise to interpretable and controllable deep generative models.
The main goal of this course is to give students the tools to think critically about causal modeling and draw connections to research, in ML as well as in other scientific disciplines. As such, we'll begin with fundamentals on causal models, formalizing causal graphical models, interventions, structural causal models, counterfactuals, and inference in these models. Next, we'll cover key results and methods for causal discovery and causal representation learning, challenging problems in causal inference that are addressed with non-iid data. Then, we'll explore more recent advances in the field of causal ML, covering causal techniques in interpretability research, amortized inference for causal modeling, and more.
The structure of course mixes lectures, student-led paper discussions, reading reports, a small coding assignment and a final project.
Class information
- Start date: January 13, 2025
- End Date: April 9, 2025
- When: Mondays and Wednesdays, 12:30 to 2:30 PM
- Where: Auditorium 1, 2nd floor, 6650 Rue Saint-Urbain (Mila)
- Auditing policy: Auditors will be allowed on a first-come, first-serve basis depending on the enrollment and their commitment to the course work. Auditing requests must be sent by email.
- Mode: This course will be in-person only.
Prerequisites
This is not an introductory course to deep learning. This is an advanced, PhD-level class that covers technical topics like probabilistic graphical models, identifiability results especially in representation learning, etc. You must have taken Fundamentals of Machine learning or an equivalent course, and have some background with probability and statistics, deep learning and unsupervised machine learning. Some familiarity with probabilistic graphical models will be very useful.