CS7545_Sp23_Lecture_16: Online Gradient Descent - mltheory/CS7545 GitHub Wiki
CS 7545: Machine Learning Theory -- Spring 2023
Instructor: Guanghui Wang
Notes for Lecture 16: Online Gradient Descent
March 02, 2023
Scribes: Chaewon Park, Dunstan Becht
Announcements
Updates on Final project
List of 20-40 papers(topics) will be given for you to choose from, but if you find something related but not in the list, that's also fine
Pick 1-2 papers, summarize and do analysis (e.g., How does this class apply to modern researches)
Two (or may be three) people can work together
No presentation
Homework 3
Likely to be due before Spring break but give grace period until the end of Spring break
No office hours over the break!
Exam
Will cover topics in HW1~HW3
Exam will be within the bounds of HWs and it will test the same thing but in different language
As long as you understand HW solutions perfectly, you will be fine on the exam
Online Learning (cont'd)
Recap: Online Convex Optimization
Framework
For t = 1, 2, ..., T do:
Learner picks a decision $w_t$ from a convex set $K \subseteq \mathbb{R}^d$.
Simultaneously, environment chooses a convex loss function $f_t: \mathcal{K} \rightarrow \mathbb{R}^d$.
Learner observes $f_t$, suffers $f_t(w_t)$, and updates $w_t$.
OGD with $\eta_t = \frac{D}{G \sqrt{T}}$ also leads to an $O(DG\sqrt{T})$ regret bound. This algorithm is "timeless" as it does not need to know $T$ in advance.