Advanced Machine Learning
What are the state-of-the-art machine learning methods and why do they work? This graduate-level introductory course to machine learning focuses on the foundations of modern machine learning. We will cover selected topics from supervised learning, unsupervised learning, generative learning, and interactive learning. For each topic, we will describe key algorithmic ideas/intuitions and basic theoretical insights. By the end of the course, students will master main machine learning techniques, and apply/optimize/develop them for specific applications.
Syllabus: Syllabus
Who: Professor Ju Sun (Instructor) Email: jusun AT umn.edu (Office Hours: 4–6pm Thur)
Leon Luo Email: luo00042 AT umn.edu (Office Hours: 3–4pm Mon)
When/Where: Tue/Thur 2:30 – 3:45pm @ Mechanical Engineering 108 (in-person only with UNITE option)
Lecture Schedule
Date | Topics | Notes |
---|---|---|
Sep 03 | Overview | [Slides] |
Sep 05 | Review of high-dimensional calculus | [Notes] |
Three Ingredients of Machine Learning; Linear Predictions | ||
Sep 10 | Linear regression and least-squares problem | [Notes] |
Sep 12 | Gradient descent for unconstrained optimization | |
Sep 17 | More on gradient descent; review of subspaces and hyperplanes | |
Sep 19 | Linear classification: Perceptron, SVM, and logistic regression | |
Support Vector Machines and Kernel Methods | ||
Sep 24 | Margin-maximization principle and hard-margin SVM | [Notes] |
Sep 26 | Review of convex analysis; KKT conditions for convex problems | |
Oct 01 | Analysis of hard-margin SVMs; soft-margin SVM | |
Oct 03 | Property of soft-margin SVMs; solving SVM problem via stochastic gradient descent | |
Oct 08 | Kernel methods | |
Oct 10 | Beyond binary classification: multiclass and other learning settings | |
Elements of Statistical Learning Theory | ||
Oct 15 | Warmup: finite hypothesis class, realizable case | [Notes] |
Oct 17 | General learning setup with finite hypothesis class | |
Oct 22 | Infinite hypothesis class with Rademacher complexity | |
Oct 24 | Infinite hypothesis class with growth function and VC dimension | |
Ensemble Methods: from Simple to Powerful | ||
Oct 29 | Decision stumps and their linear combinations; Adaboost | [Notes] |
Oct 31 | Why Adaboost works; Adaboost as greedy method to find a linear combination | |
Nov 05 | Gradient boosting | |
Nov 07 | Decision trees | |
Nov 12 | Random forests | |
Nov 14 | Generalization of SVM and Boosting via margin bounds | |
Beyond Supervised Learning | ||
Nov 19 | Three interpretation of PCA | [Notes] |
Nov 21 | Random projection and compressed sensing | |
Nov 26 | Nonlinear dimension reduction | |
Nov 28 | HAPPY THANKSGIVING! NO LECTURE | |
Dec 03 | Clustering with non-density-based methods | |
Dec 05 | Clustering with density-based methods | |
Dec 10 | Generative models | |
Homework Assignments
HW1 (Due: Oct 06, 2024)
HW2 (Due: Oct 27, 2024)
HW3 (Due: Nov 30, 2024)
HW4 (Due: Dec 11 2024)
HW5 (Due: Dec 23 2024)