Student projects: preparatory information

Considering doing a project in MINDS?

This page provides decision-making-helping information for students who think about doing a project in the MINDS group. The information given below is equally relevant for Bachelor thesis projects, Master thesis projects, and individual research projects.

What is “Machine Learning”?

ML is a collection of algorithm design methods which is firmly rooted in probability theory, statistics, linear algebra and some calculus. ML is a particularly math-heavy area of Computer Science. The input to an ML algorithm is (large quantities of) data - for example images, video, biosignals, geo-surveillance data, scientific measurements, financial records, web-crawling harvests. An ML algorithm tries to find regularities (structure, redundancies, symmetries...) in the data and use these to create a formal model of the data, in a format which should be much more compact than the original data. After having thus learnt a model from the data, the model can be exploited for a variety of purposes, for instance data interpolation or prediction, pattern classification, control, fault monitoring, robot action selection, or pattern generation.

Should you go for it?

If you are contemplating choosing ML for your project, the first thing for you is to find out whether you really like this kind of stuff and whether it is accessible to you. The standard way to find out is to take my 2nd year undergraduate course Neural Networks (AI) (KIB.NNKI03) or my 2nd year Master course Machine Learning (KIM.ML09) and emerge from that course with enthusiasm. If you haven’t taken one of those courses (or related ones offered in the CS department), you face a rather steep working-in challenge because you will have to self-study the basics of machine learning. My lecture notes for KIB.NNKI03 or my lecture notes for KIM.ML09 may be good fast-self-study readings if you are determined and dedicated. - Students of mathematics usually have little difficulties absorbing the requisite background knowledge and are by default welcome to engage in a machine learning project.

Prerequisites and admission procedure

Required qualification. Since embarking on a ML project really only makes sense when you have made friends with elementary concepts of this field, we will supervise you only if both you and I can be sure that you muster the basics of the field. We will take this for granted if you have passed the 2nd year Bachelor course Neural Networks, or the 2nd year Master course Machine Learning with a grade of 7.00 or better. If you did not take one of these courses, but have acquired reasonably substantial ML background otherwise, we will carry out a qualification exam.

Admission. Since the supervision capacity in MINDS is limited and ML is a popular subject area, there will be more qualified students wanting to do a project than we can host. We accept students on a first-come, first-serve basis. Herbert Jaeger maintains a waiting list. You can request to be put on that list before you have finished one of the qualifying courses or done the qualification exam. Experience tells that this list starts filling a year in advance, but experience also tells that a significant number of candidates later withdraws (because they changed their minds) or don’t pass the qualification hurdles. Toward the beginning of a semester, the students at the head positions in the list will get an acceptance notification and the ones at the tail positions will get a decline note. Apart from these final notifications, Herbert Jaeger will try to keep everybody on the list continually informed about the predicted chances of getting in, in order to make your planning life easier.

Inside MINDS: which subject options?

Student projects in MINDS are usually related to the active research areas of our group, though we can also negotiate very individual project themes in cases where the student knows what s/he is up to and is qualified to work without close guidance. Our group’s three default thematic areas are

  • Echo State Networks, a learning paradigm for recurrent neural networks - this is likely the most accessible kind of subject and has been the typical choice of CS, IMS, and DE students in the past,
  • Observable Operator Models, the most “mathy” and arguably most elegant and foundational kind of research we can offer - this has been opted for by most mathematics students in the past,
  • Conceptors, the most complex topic because it builds on echo state networks, maybe a choice for the daring ones with a very solid linear algebra background and an interest in cognitive neural dynamics.

We don’t have ready-made project topics for you to pick from. Rather, we’ll be having a cup of coffee together and define a project topic tailored to your interests and background.

Below you find startup reading suggestions for each of these three fields.

Startup reading for Echo State Networks

To get a clearer picture of this subject, please consult the following papers in the given order:

  1. A Scholarpedia article for a first overview.
  2. A short highlight paper: H. Jaeger and H. Haas, Harnessing Nonlinearity: Predicting Chaotic Systems and Saving Energy in Wireless Communication. Science 304, 2 April 2004, pp. 78-80 (preprint pdf)
  3. An easy overview paper of the status of echo state network research in the current landscape of ML: M. Lukoševičius, H. Jaeger, B. Schrauwen (2012): Reservoir Computing Trends. KI - Künstliche Intelligenz, 1-7 (Preprint pdf)
  4. An introductory, more detailed technical report which has a large number of examples in it: H. Jaeger (2001): Short term memory in echo state networks. GMD Report 152, German National Research Center for Information Technology, 2001 (60 pp.) (pdf)

You might also wish to check out a (very good) BSc thesis written on an ESN theme (by Valentin Vasiliu, 2016).

Startup reading for Observable Operator Models

There is unfortunately no really easy introduction paper for OOMs. The most accessible exposition is given in the first part in the 2-part paper

  1. H. Jaeger, M. Zhao, K. Kretzschmar, T. Oberstein, D. Popovici, A. Kolling (2006): Learning observable operator models via the ES algorithm. In: S. Haykin, J. Principe, T. Sejnowski, J. McWhirter (eds.), New Directions in Statistical Signal Processing: from Systems to Brain. MIT Press, Cambridge, MA., 417-464 (draft version, pdf)

Note: OOM theses written in the MINDS group seem to be good for carving an academic carreer path: Cristian Danescu-Mizil (Master thesis 2007) is now an assistant professor at Cornell, Anca Dragan (BSc thesis 2009) is assistant professor at Berkeley, and Josip Djolonga (BSc thesis 2011) is a PhD student at ETH Zurich with a Google European Doctoral Fellowship. You can check out Josip's BSc thesis if you want to get an OOM feeling.

Startup reading for Conceptors

Finally, research on conceptors is young and there is not much literature yet. Before you embark on this expedition, please familiarize yourself a little with echo state networks, because conceptor theory is framed within that context. Here is what can be offered:

  1. A short teaser paper is H. Jaeger (2014): Conceptors: an easy introduction. (arXiv)
  2. The "official coming-out" paper: H. Jaeger (2017): Using Conceptors to Manage Neural Long-Term Memories for Temporal Patterns. Journal of Machine Learning Research 18, 1-43 (pdf at JMRL)

So far there have only been two conceptor-based BSc thesis - and fine ones, too - by Alina Dima (2014)(pdf) and by Rubin Deliallisi (2017)(pdf).