The standard approach to machine learning uses a training set of labeled examples to learn a prediction rule that will predict the labels of new examples. Collecting such training sets can be expensive and time-consuming. This course will explore methods that leverage already-collected data to guide future measurements, in a closed loop, to best serve the task at hand. We focus on two paradigms: i) in pure-exploration we desire algorithms that identify or learn a good model using as few measurements as possible (e.g., classification, drug discovery, science), and ii) in regret minimization we desire algorithms that balance taking measurements to learn a model with taking measurements to exploit the model to obtain high reward outcomes (e.g., content recommendation, medical treatment design, ad-serving).
The literature on adaptive methods for machine learning has exploded in the past few years and can be overwhelming. This course will classify different adaptive machine learning problems by characteristics such as the hypothesis space, the available actions, the measurement model, and the available side information. We will identify general adaptive strategies and cover common proof techniques.
List of topics:
A version of this course was previously offered two years ago. Its course website may be a useful resource for related literature and course notes.
Prerequisites: The course will make frequent references to introductory concepts of machine learning (e.g., CSE 546) but it is not a prerequisite. However, fluency in basic concepts from linear algebra, statistics, and calculus will be assumed (see HW0). Some review materials:
There will not be a textbook for the course. Our discussion will be guided by papers, monographs, and lecture notes that are available online. An incomplete list that will grow:
There will be a Slack channel (first day of class). This is your first resource for questions. For private or confidential questions email the instructor. You may also get messages to the instructor through anonymous course feedback.
There will be 3 homeworks (each worth 20%) and one take-home cumulative final exam (worth 40%).
Each homework assignment will be submitted as a single PDF to gradescope. Any code for a programming problem should come at the end of the problem, after any requested figures for the problem. We expect all assignments to be typeset (i.e., no photos or scans of written work). This can be done in an editor like Microsoft Word or Latex (highly recommended). There exist convenient packages for listing Python code in Latex.
Homeworks must be done individually: each student must hand in their own answers. In addition, each student must write their own code in a programming part of the assignment. It is acceptable, however, for students to collaborate in figuring out answers and helping each other solve the problems. You also must indicate on each homework with whom you collaborated.
The homework problems have been carefully chosen for their pedagogical value and hence might be similar or identical to those given out in past offerings of this course at UW, or similar courses at other schools. Using any pre-existing solutions from these sources, from the Web or other textbooks constitues a violation of the academic integrity expected of you and is strictly prohibited.
You will be given 24 hours for the take home exam that is expected to take no more than two hours to complete. It cannot be turned in late. Plan accordingly.
All requests for regrading should be submitted to Gradescope directly. Office hours and in person discussions are limited solely to asking knowledge related questions, not grade related questions. If you feel that we have made an error in grading your homework, please let us know with a written explanation, and we will consider the request. Please note that regrading of a homework means the entire assignment may be regraded which may cause your grade on the entire homework set to go up or down. Regrade requests must be submtted within 7 days (24*7 hours) of the time in which grades are released.