[Course description
| Schedule and handouts
| Related material]
Date | Topic | Resources |
---|---|---|
Mar 27 | Intro to PAC learning (notes) | Avrim Blum's slides, [KV, chaps 1-2], [SSS, chaps 2-4], [MRT, chap 2], video - Andrew Ng |
Mar 31 | Sample complexity via growth functions (notes) | Avrim Blum's slides and notes, [KV, chap 3], [SSS, chaps 2-6], [MRT, chap 3] |
Apr 3 | VC dimension (notes) | [KV, chap 3], [SSS, chaps 2-6], [MRT, chap 3] |
Apr 7 | Rademacher complexity (notes) | Avrim Blum's notes, [SSS, chap 26], [MRT, chap 3] |
Apr 10 | Intro online learning (notes) | [AHK survey], [Hazan, Ch 1] |
Apr 14 | Applications of experts (notes) | Bobby Kleinberg lecture notes |
Apr 17 + 21 | Applications, cont. + Follow the perturbed leader (notes) | FTPL by Kalai and Vempala |
Apr 21 + 24 | Intro to online convex optimization (notes) | [SSS survey, Ch 2], [Hazan, Ch 5]; Sebastien Bubeck notes |
May 1 | Follow the regularized leader (notes) | See above, and Bubeck's book; Useful facts about relative entropy |
May 5 | Equivalence to online mirror descent, applications; (notes) | See above; other notes on mirror descent here and here |
May 8 | Application to approx Caratheodory (notes); Intro linear bandits (notes>) | Tight bounds for Approx Caratheodory Thm |
May 12 | Multi-armed bandits (notes>) | [SSS survey, Ch 4]; MAB book draft by Slivkins; Bubeck and Cesa-Bianchi; Bubeck lecture notes part 1 and part 2 | May 15 | Linear bandits | see above |
May 19 | Linear and contextual bandits (notes>) | see above |
May 22 | Stochastic multi-armed bandits | Slivkins, chaps 2-3; Shipra Agrawal notes #1 and #2 |
May 26 | Markovian MAB, Gittins index | Richard Weber lecture notes, chap 7, Shipra Agrawal notes #1, #2, #3 |
June 2 | Student presentations | Linear coupling: An ultimate unification of gradient and mirror descent
Towards minimax policies for online linear optimization with bandit feedback |
June 5 | Sebastien Bubeck guest lecture | Mirror descent and self-concordant barriers for the linear bandit problem |
June 7 | Student presentations |
Katyusha: The first direct acceleration of stochastic gradient methods
Analysis of Thompson Sampling for multi-armed bandits |
In the above list:
Instructors: Nikhil Devanur and Anna Karlin
Time: Mondays and Fridays
in CSE 403, 11:00am -- 12:20pm
Course content: Inspired by the recent and current special semesters at the Simons Institute for Theoretical Computer Science, we will explore some of the key themes and approaches to handling uncertainty in algorithm design and analysis, with particular emphasis on basics of learning theory and online learning. Topics to be covered will be selected from:
Course evaluation: 2-4 homeworks and a small project.
Background expected: Mathematical maturity, basics of probability and undergraduate level algorithms.