NETS 7976 - Reasoning Under Uncertainty with Probabilistic Machine Learning - Fall 2023
Thursdays: 12:00 – 1:30pm
September 7 – December 14, 2022
177 Huntington, room 1005
Summary
This class is an introduction to the field of Probabilistic Machine Learning. The course will build the mathematical foundation for understanding inference, learning, and reasoning through the lens of probability theory. This is especially relevant in today's scientific landscape, given that the theoretical backbone of modern machine learning and artificial intelligence techniques is inextricably built on fundamental concepts from probability theory. Students will leave the class with experience in solving theoretical and applied problems related to Probabilistic Machine Learning. This class is designed in collaboration with Moritz Laber, a PhD student in Network Science.
PDF of this syllabus: here.
Coursework, Class Structure, Grading
The course adopts a flipped classroom concept: Students are exposed to the relevant material through pre-recorded video lectures, readings, and exercises that involve both analytic calculation and coding. The in-class hours are dedicated to recapitulation of the most important points, clarification of questions, as well as open discussion. The grading of the course is based on active participation in discussions and continual work on exercises. In a typical class students will summarize the main points of the lectures in their own words. The weekly exercises and questions, that students prepare prior to class, structure the in class discussion. Discussions will remain open and allow to explore topics deeper as needed.
Learning Objectives and Outcomes
By the end of this course students should have a deep familiarity with the central concepts of probabilistic machine learning and reasoning under uncertainty. In particular, they develop a thorough understanding of Gaussian processes (GP) for regression and classification, from mathematical, algorithmic and applied perspectives. Building on this understanding, they learn about the relevance of uncertainty for deep learning and how this problems can be addressed through the lens of GP. The goal of the course is to equip students with the necessary knowledge and tools to apply probabilistic machine learning in their own work, be it theoretical or applied. Furthermore, students sharpen their general understanding of probabilistic and algorithmic reasoning.
Current: While the course starts from well established foundations of exponential family distributions and GP regression, it progresses towards topics of current research interest, such as uncertainty in deep learning and efficient implementation of GPs.
Practical: The course emphasizes algorithmic challenges posed by probabilistic machine learning and exercises guide students through efficient implementation of the relevant techniques.
Novel & Actionable: The open structure of in-class discussion encourages students to explore the application of the course material to their own area of research as well and domain specific questions outside of machine learning research.
Evaluation
The course evaluation is based on two components:
Completion of weekly exercises: Students present their solutions to weekly exercises in class. These solutions should show an honest effort and active engagement with the material.
Active participation in the weekly meetings: This includes active participation in discussions through posing relevant questions, attempts to answer other students questions and connecting the material to real world problems, e.g. from their own field of research.
Materials
This course uses pre-recorded lectures from the courses Probabilistic Machine Learning and Numerics of Machine Learning that were offered by Philipp Hennig in the summer term 2023 and winter term 2022 respectively at the University of Tübingen as part of a graduate program in Machine Learning. The recordings are available here and slides here. These materials are available under a CC BY-NC-SA 4.0 license. Selected chapters from the three volume series Probabilistic Machine Learning by Kevin Murphy as well as Phillip Hennig's textbook Probabilistic Numerics: Computation as Machine Learning will serve as supplementary reading.
Instructor
Brennan Klein is an associate research scientist at the Network Science Institute, with a joint affiliation at the Institute for Experiential AI. He is the director of the Complexity & Society Lab. His research spans two broad topics: 1) Information, emergence, and inference in complex systems — developing tools and theory for characterizing dynamics, structure, and scale in networks, and 2) Public health and public safety — creating and analyzing large scale datasets that reveal inequalities in the United States, from epidemics to mass incarceration. Dr. Klein received a PhD in Network Science in 2020 from Northeastern University and got his BA in Cognitive Science & Psychology from Swarthmore College in 2014. Website: http://brennanklein.com/.
Week 1: Sep. 7
Introduction & Reasoning under Uncertainty
Lectures:
Lecture 1: Introduction
Lecture 2: Reasoning Under Uncertainty
Supplementary Readings:
Murphy 2: 2.1 Probability: Introduction
Murphy 2: 3.2 Bayesian Statistics
Murphy 2: 4.2 Directed Graphical Models
Week 2: Sep. 14
Continuous Variables & Exponential Families I
Lectures:
Lecture 3: Continuous Variables
Lecture 4: Exponential Families
Supplementary Readings:
Murphy 2: 2.4 The Exponential Family
Murphy 2: 2.5 Transformations of Random Variables
Murphy 2: 3.4 Conjugate Priors
Week 3: Sep. 21
Exponential Families II & Gaussian Probability Distribution
Lectures:
Lecture 5: Exponential Families II
Lecture 6: Gaussian Probability Distribution
Supplementary Readings:
Murphy 1: 2.6 Univariate Gaussian (normal) Distribution
Murphy 1: 3.2 The Multivariate Gaussian (normal) Distribution - 3.3 Linear Gaussian Systems
Murphy 2: 2.3 Gaussian Joint Distribution
Week 4: Sep. 28
Parametric Regression & Gaussian Processes (GP)
Lectures:
Lecture 7: Parametric Regression
Lecture 8: Gaussian Processes
Supplementary Readings:
Murphy 1: 12 Generalized Linear Models
Murphy 2: 15.1 GLM: Introduction - 15.2 GLM: Linear Regression
Murphy 1: 17.2 Gaussian Processes
Week 5: Oct. 5
Understanding through an Extensive Example
Lectures:
Lecture 9: Understanding Gaussian Processes
Lecture 10: Gaussian Processes Regression: An Extensive Example
Supplementary Readings:
Murphy 2: 18.1 GP: Introduction - 18.5 GP with non-Gaussian Likelihoods
Week 6: Oct. 12
Understanding GP through Kernels and Linear Algebra
Lectures:
Lecture 11: Understanding Kernels and Gaussian Processes
Lecture 12: The Role of Linear Algebra in Gaussian Processes
Supplementary Readings:
Schölkopf & Smola Learning with Kernels (2002) Chapter 1: A Tutorial Introduction
Murphy 1: 7.6 Other Matrix Decompositions
Week 7: Oct. 19
Computation, Inference & Logistic Regression
Lectures:
Lecture 13: Computation and Inference
Lecture 14: Logistic Regression
Supplementary Readings:
Henning 2022 Chapter III.14 - III.20 Linear Algebra
Murphy 1: 10 Logistic Regression
Murphy 2: 12 Generalized Linear Models
Week 8: Oct. 26
GP Regression & Deep Learning
Lectures:
Lecture 15: Gaussian Process Regression
Lecture 16: Deep Learning
Supplementary Readings:
Murphy 1: 13 Neural Networks for Tabular Data
Murphy 2: 16 Deep Neural Networks
Week 9: Nov. 2
Probabilistic & Uncertain Deep Learning
Lectures:
Lecture 17: Probabilistic Deep Learning
Lecture 18: Uncertainty in Deep Learning
Supplementary Readings:
Murphy 2: 17 Bayesian Neural Networks
Murphy 2: 18.7 GPs and DNNs
Week 10: Nov. 9
Use Cases & Gauss-Markov Models
Lectures:
Lecture 19: Uses of Uncertainty for Deep Learning
Lecture 20: Gauss-Markov Models
Supplementary Readings:
Henning: I.5 Gauss-Markov Processes: Filtering and SDEs
Week 11: Nov. 16
Parameter Inference
Lectures:
Lecture 21: Parameter Inference I
Lecture 22: Parameter Inference II
Supplementary Readings:
Murphy 1: 8.7 Bound Optimization
Murphy 2: 6.5 Bound Optimization
Murphy 2: 10.1 Variational Inference
Week 12: Nov. 23
Thanksgiving - No Class
Week 13: Nov. 30
Variational Inference & Historic Perspective
Lectures:
Lecture 23: Variational Inference
Lecture 24: Historical Perspective
Supplementary Readings:
Murphy 2: 10 Variational Inference
Week 14: Dec. 7
Probabilistic Numerics
Lectures:
Numerics of ML 6: Solving Ordinary Differential Equations
Numerics of ML 7: Probabilistic Numerical ODE Solvers
Supplementary Readings:
Henning: VI Solving Ordinary Differential Equations
Week 15: Dec. 14
Review Week
Recapitulate the most important concepts introduced during the course. Discuss the application of the course material to current and future research projects, as well as their implications for network science in general.