Type | Seminar (7 ECTS) |
Lecturer | Prof. Dr. Jilles Vreeken and Dr. David Kaltenpoth |
vreeken (at) cispa.de | |
Meetings | Thursdays 10:00 till 12:00, room 0.01, E9.1 (CISPA) |
Summary | In this seminar, we will discuss information theoretic approaches to machine learning. We will investigate, among others, the following questions: What is interesting and meaningful structure? How can we identify this from data without overfitting? What is a good model when we don't have a decent prior, target, or even know what we're looking for? What is the ultimate model, and how can we approximate it in practice? We'll explore these in light of Algorithmic Information Theory, and its practical variant, the Minimum Description Length (MDL) principle. We will consider the relevance and application of these to a wide range of problems, from description and prediction to generalization, from neural to symbolic, from associative to causal, and so on. |
Month | Day | Time | Type | Topic | Slides | Req. Reading | Opt. Reading |
---|---|---|---|---|---|---|---|
Oct | 24 | 10:00 | L | Introduction | [1] Ch 1.1–1.10 | [1,2,3] | |
31 | 10:00 | L | Compression | [1] Ch 2.1–2.2, 2.8 | [4,5] | ||
Nov | 7 | 16:00 | L | Information | [2] Ch 2.1-2.8, 3.1, 8.1, 8.3-8.6, 12.1-12.2 | [1] Ch 2.8 (1.11), [6,7] | |
14 | 10:00 | D | Discussion | [8,9] | [10] | ||
19 | 10:00 | L | Estimation | ||||
26 | 10:00 | L | MDL | [3] Ch 1 | |||
Dec | 5 | 10:00 | D | Discussion | |||
12 | yay — no lecture | ||||||
19 | 10:00 | L | Description | ||||
26 | yay holiday — no lecture | ||||||
Jan | 2 | yay holiday — no lecture | |||||
9 | 10:00 | L | Prediction | [11] | |||
16 | 10:00 | D | Discussion | ||||
23 | 10:00 | L | Causation | ||||
30 | 10:00 | L | RL | ||||
Feb | 6 | 10:00 | D | Discussion | |||
20 | S | Student Presentation |
Lecture type key:
We will publish a list of possible student topics on or before December 5th, and ask you to send us your preferences by December 12th. We will let you know the overall assignment by December 19th latest. This means you then have until December 26th to (de)register in LSF.
All required reading will be made available for download. You will need a username and password. These will be given out in the first meeting. If you need a reminder, think of a cat in a box.
The following books provide relevant background to many of the topics of the course, Li & Vitányi [1], Cover & Thomas [2], Jaynes [12], Grünwald [3], Rissanen [13], Rissanen [14], Yamanishi [15], Wallace [16]. Most are also available in a so-called Semesteraparat at the university library.
[1] | An Introduction to Kolmogorov Complexity and its Applications. Springer, 1993. |
[2] | Elements of Information Theory. Wiley-Interscience New York, 2006. |
[3] | The Minimum Description Length Principle. MIT Press, 2007. |
[4] | On Tables of Random Numbers. The Indian Journal of Statistics, Series A, 25(4):369-376, 1963. |
[5] | Three Approaches to the Quantitative Definition of Information. Problemy Peredachi Informatsii, 1(1):3-11, 1965. |
[6] | A Mathematical Theory of Communication. The Bell System Technical Journal, 27:379-423, 623-656, 1948. |
[7] | The bandwagon. IRE transactions on Information Theory, 2(1):3, 1956. |
[8] | The similarity metric. IEEE Transactions on Information Technology, 50(12):3250-3264, 2004. |
[9] | Clustering by Compression. IEEE Transactions on Information Technology, 51(4):1523-1545, 2005. |
[10] | On data mining, compression and Kolmogorov Complexity. Data Mining and Knowledge Discovery, 15(1):3-20, Springer-Verlag, 2007. |
[11] | Minimum Encoding Approaches for Predictive Modeling. In Proceedings of the 14th International Conference on Uncertainty in Artificial Intelligence (UAI), 1998. |
[12] | Probability Theory: The logic of science. Cambridge Press, 2002. |
[13] | Information and complexity in statistical modeling. Springer, 2007. |
[14] | Optimal Estimation of Parameters. Springer, 2012. |
[15] | Learning with the Minimum Description Length Principle. Springer, 2023. |
[16] | Statistical and inductive inference by minimum message length. Springer, 2005. |
In general terms, the course will consist of
The course has one meeting per week, typically on Thursdays 10-12, but with a few exceptions due to room availability. The first part of the course will feature regular lectures covering the basic topics of the course and sessions in which we will discuss material covered in the lectures and scientific articles assigned by the lecturer. During the second part the students will write an essay based on scientific articles assigned to them by the lecturer and will prepare a presentation. All student presentations will be scheduled in a block-seminar style, that is in one or two days near the end of the semester, the exact date is to be announced.
There will be no weekly tutorial group meetings.
Every student will have to give a presentation and write a report on an assigned topic. The presentation should be between 15 and 20 minutes long, and will be followed by a 10 minute discussion. The style of the presentation should be like a lecture; your fellow students can follow it with only the previous lectures of the course as background material, you should only discuss those details that are necessary, and make use of examples where possible. Topic will be assigned by the lecturer.
In addition to the presentation, students will have to hand in a report that discusses the assigned topic. Unlike the presentation, which has to be high-level, the report is where you are allowed to show off what you've learned. Summarize the material you read as clearly as you can in your own words, identifying the key contributions/most interesting or important aspects, relating the topic to any/all previous lectures in the course and/or papers read for the course, all the while correctly referring to the sources you've drawn from. The expected length of the report is 5-10 pages, but you are free to use as many pages as you like.
Students will be graded on the slides, the presentation, your answers in the discussion, and the report. Students are allowed to ask for feedback on their slides once a week, up till latest 1 week before the presentation. The provided reading material are provided as starting points. For some topics the provided materials will be (by far) (way) (more than) enough, for others you may want to gather more material. When in doubt, contact the lecturer.
The deadline for your slides and report is right before the meeting in which you have to give your presentation.
Students should have basic working knowledge of machine learning and statistics, e.g. by successfully having taken courses related to data mining, machine learning, and/or statistics, such as Topics in Algorithmic Data Analysis, Machine Learning, Probabilistic Graphical Models, Elements of Machine Learning, etc.
Discussions will be an essential element of this seminar, and hence there will a maximum of twelve (12) participants.
Students that want to participate should register via the seminar registration system on or before October 16th, and make sure to specify: a) name, b) level (bachelor, n-th year master, PhD student), c) relevant courses taken so far, d) short motivation for why they want to participate.