News
Registration for the first projectWritten on 27.11.23 by Tatjana Chavdarova Dear all, |
Zoom exercise session - Thursday Nov 23, 13.30hWritten on 20.11.23 by Sebastian Stich Dear students, We are offering the first online exercise session/office hour this week on Thursday, 13.30h (the link will be added under 'materials'). You are cordially invited to ask any questions you may have, especially about the first part (optimization), but of course also about the more… Read more Dear students, We are offering the first online exercise session/office hour this week on Thursday, 13.30h (the link will be added under 'materials'). You are cordially invited to ask any questions you may have, especially about the first part (optimization), but of course also about the more recent content. (Note that the structure or content will not be different from the 'live' session offered on Tuesdays. It is just an opportunity for people who cannot attend on Tuesdays to ask questions). The zoom link for tomorrow's lecture is again "Zoom link (tatjana)." |
Lecture on Tuesday, Nov. 14th, 2023Written on 14.11.23 by Tatjana Chavdarova Dear all, |
First exercise sessionWritten on 06.11.23 by Sebastian Stich Dear students, Tomorrow we will have our first exercise session. We will discuss questions you might have on the optimization part (a second session will be offered on zoom next week). The zoom link to attend the lecture is the same as last week. |
First lecture on Tuesday, Oct. 24th, 2023Written on 23.10.23 (last change on 24.10.23) by Tatjana Chavdarova Dear students, Place: in CISPA lecture hall. E9.1 (CISPA building), room 0.05 (main lecture room) Dear students, Place: in CISPA lecture hall. E9.1 (CISPA building), room 0.05 (main lecture room)
|
Games in Machine Learning
A typical deep learning (DL) pipeline involves defining a utility or loss function and training a set of parameters to optimize it. This approach has limitations. The specified objective may fail to capture the desired behavior of the model, resulting in high-performing models that struggle to make accurate predictions when faced with slightly modified inputs or that learn irrelevant correlations, ultimately leading to subpar performance. This is solved through a technique called robustification, which employs a min-max objective. In this approach, two models, or players, are trained jointly. Each player has a different real-valued objective that depends on the parameters of both players. Consequently, in the field of machine learning, it becomes necessary to understand how to solve two-player games using iterative, gradient-based methods to find the Nash Equilibrium solution. Furthermore, many problems in machine learning inherently involve multiple players, such as Generative Adversarial Networks, multi-agent reinforcement learning, collaborative robots, and competing drones.
This course will use a framework encompassing all these problems called Variational Inequalities (VIs). Due to its generality, our studies also apply to standard minimization (single-player games). In addition to a brief introduction to game theory needed from a machine learning perspective, this course focuses on:
(i) Variational Inequalities and classes of VIs,
(ii) gradient-based optimization methods to solve VIs,
(iii) applications of VIs in machine learning problems.
Prerequisites: Linear algebra knowledge is required. Machine learning and convex optimization are recommended.
Grading Breakdown:
- 60% theoretical exam,
- 40% two projects in PyTorch (10% first project, 30% second project)
Homework and quizzes are not graded but part of the lecture content and solving them is strongly encouraged.
Only one exam will be offered in the exam session (this means you cannot improve your exam grade in a second attempt).
Date for lecture: Tuesdays, 14-16h.
- Students are encouraged to attend in person, as spontaneous discussions and brief lecture synopses will not be documented.
- We will provide a zoom link for online attendance. (As a courtesy, we will make a best effort attempt to publish recordings of some of the lectures. You should not assume that recordings will be available.)
First lecture: October 24, in CISPA lecture hall. E9.1 (CISPA building), room 0.05 (main lecture room)
Exercise sessions
We offer two slots (that alternate weekly - the detail schedule will be announced):
- Tuesday, 16-17h, in the lecture hall
- Thursday, 14-15h, on zoom
Registered students will find more information under the Lecture Outline tab.
Materials are posted under the materials tab.
About the lecturers:
- Tatjana Chavdarova is an incoming faculty at CISPA, with a research focus on the intersection of game theory and machine learning.
- Sebastian Stich is a faculty at CISPA, with a research focus on optimization for machine learning.