Teemu Roos

Ph.D., Postdoctoral Researcher

Helsinki Institute for Information Technology HIIT
Exactum building, A322
Department of Computer Science
PO Box 68
FI-00014 University of Helsinki
Finland

teemu.roos at cs.helsinki.fi

What's up?

The University of Helsinki has granted EUR 120,000 to project STAM (Algorithmic Methods in Stemmatology) for the years 2009–2011. "Computer programs can do wonders"

As of August 2008, I have been appointed as post-doctoral researcher at HIIT for three years.

Conference and workshop involvement (program committee): WITMSE-08, ECAI-08, PGM-08, UAI-08/09.

Recent Breakthroughs in MDL Learning workshop at ICML/UAI/COLT-08, July 9, Helsinki. Workshop page

In January–April 2008, I visited UC Berkeley (Prof. Bin Yu's group) and ICSI.

Computer-assisted Stemmatology Challenge funded by the Pascal network. Challenge page

Since March 1st 2007, we have two boys (see below)!

Teaching

I am the tutor (FM-HOPS) of the Intelligent Systems sub-programme at the Department of Computer Science.

No courses during the academic year 2008–2009. In Fall 2009, I will teach the new course "Information-Theoretic Modeling".

Research

``Your act was unwise,'' I exclaimed ``as you see by the outcome.''
He solemnly eyed me. ``When choosing the course of my action,''
said he, ``I had not the outcome to guide me.''
[Ambrose Bierce]

I'm a member of the CoSCo group (Complex Systems Computation Group).

I am also the Programme Manager of the PAS (Probabilistic Adaptive Systems) programme of HIIT.

Topics of my interest include

machine learning (distribution-free, game-theoretic);
data analysis (information-theoretic, Bayesian).

Selected publications (more, Google Scholar, DBLP):

T.Roos and T.Heikkilä, (2009). Evaluating Methods for Computer-Assisted Stemmatology using Artificial Benchmark Data-Sets, to appear in Literary and Linguistic Computing. abstract | data-sets
Given a collection of imperfect copies of a textual document, the aim of stemmatology is to reconstruct the history of the text, indicating for each variant the source text from which it was copied. We describe an experiment involving three artificial benchmark data-sets to which a number of computer-assisted stemmatology methods were applied. Contrary to earlier similar experiments, we propose and use a numerical criterion to evaluate all the solutions. Moreover, our primary data-set is significantly larger than used before. The results suggest the superiority of two computer-assisted methods amongst those tested: the maximum parsimony method implemented in the PAUP* software package and a related compression-based method we have proposed in earlier work.
T.Silander, T.Roos, and P. Myllymäki, (2009). Locally Minimax Optimal Predictive Modeling with Bayesian Networks, to appear in AISTATS-09, Clearwater Beach FL, USA, April 16–18, 2009. abstract
We propose an information-theoretic approach for predictive modeling with Bayesian networks. Our approach is based on the minimax optimal Normalized Maximum Likelihood (NML) distribution, motivated by the MDL principle. In particular, we present a parameter learning method which, together with a previously introduced NML-based model selection criterion, provides a way to construct highly predictive Bayesian network models from data. The method is parameter-free and robust, unlike the currently popular Bayesian marginal likelihood approach which has been shown to be sensitive to the choice of prior hyperparameters. Empirical tests show that the proposed method compares favorably with the Bayesian approach in predictive tasks.
T.Roos and B.Yu, (2009). Estimating Sparse Models from Multivariate Discrete Data via Transformed Lasso, Information Theory and Applications Workshop (ITA-09), San Diego CA, USA, February 8–13, 2009.
T.Silander, T.Roos, P.Kontkanen, and P.Myllymäki, (2008). Factorized NML Criterion for Learning Bayesian Network Structures, 4th European Workshop on Probabilistic Graphical Models (PGM-08), September 17–19, Hirtshals, Denmark. slides
T.Roos and J.Rissanen, (2008). On Sequentially Normalized Maximum Likelihood Models, Workshop on Information Theoretic Methods in Science and Engineering (WITMSE-08), Tampere, Finland, August 18–20, 2008. R code
T.Roos, (2008). Monte Carlo Estimation of Minimax Regret with an Application to MDL Model Selection, IEEE Information Theory Workshop 2008 (ITW-08), Porto, Portugal, May 5–9.
J.Rissanen, P.Grünwald, J.Heikkonen, P.Myllymäki, T.Roos, and J.Rousu, (2007). Editorial: Information Theoretic Methods for Bioinformatics, EURASIP Journal on Bioinformatics and Systems Biology. papers
J.Rissanen, and T.Roos, (2007). Conditional NML Universal Models, in Proc. 2007 Information Theory and Applications Workshop (ITA-07), pp. 337–341.
T.Roos, P.Myllymäki, and J.Rissanen, (2006). MDL Denoising Revisited, preprint arXiv cs.IT/0609138. supplementary material | C code
T.Roos, P.Grünwald, P.Myllymäki, and H.Tirri, (2006). Generalization to Unseen Cases, in Advances in Neural Information Processing Systems 18 (NIPS-05), pp. 1129-1136. (An earlier version won the best paper award at BNAIC-05.) NIPS poster
T.Roos, T.Heikkilä, and P.Myllymäki, (2006). A Compression-Based Method for Stemmatic Analysis, in Proc. ECAI-06, pp. 805-806. extended version | challenge
T.Roos, P.Myllymäki, and H.Tirri, (2005). On the Behavior of MDL Denoising, in Proc. AISTATS-05, pp. 309-316. Erratum: Caption of Fig.4 should have sigma=5.0 instead of sigma=10.0.
T.Roos, H.Wettig, P.Grünwald, P.Myllymäki, and H.Tirri, (2005). On Discriminative Bayesian Network Classifiers and Logistic Regression, Machine Learning 59:3, pp. 267-296.
T.Roos, P.Myllymäki, and H.Tirri, (2002). A Statistical Modeling Approach to Location Estimation, IEEE Trans. Mobile Computing 1:1, pp. 59-69.
T.Roos, P.Myllymäki, H.Tirri, P.Misikangas, and J.Sievänen, (2002). A Probabilistic Approach to WLAN User Location Estimation, Int. Journal of Wireless Information Networks 9:3, pp. 155-164. Ekahau Inc.

Past Events

During the Fall term 2007 I lectured the Three Concepts: Information course.

Teaching (TA) in Spring 2007: Three concepts: Utility, Three concepts: Probability, Scientific writing, Personal study plan and teacher tutoring.

During the Spring term 2006, I was a teaching assistant on the Universal AI course by Marcus Hutter (see homework instructions). During the Fall term 2003, I was a teaching assistant on the Graphical Models course by Wray Buntine and Petri Myllymäki. During the Spring term 2002, I was a teaching assistant on the Three Concepts: Information course.

I defended my Ph.D. thesis "Statistical and Information-Theoretic Methods for Data Analysis" on June 9, 2007. The opponent was Prof. Alon Orlitsky (UCSD). Pre-examiners were Prof. Ioan Tabus (Tampere UT) and Prof. Tommi Jaakkola (MIT). electronic version (summary part).

I received a Ph.D. degree (in Computer Science) from the University of Helsinki in 2007. I was supported by HeCSE (Helsinki Graduate School in Computer Science and Engineering). My supervisors were prof. Henry Tirri (on industrial leave), and prof. Petri Myllymäki. In addition to computer science I have minors in mathematics and philosophy (see a list of finished courses).

Disclaimer: If you are a student seeking an exchange position, please do not send me e-mail, but contact the Department of Computer Science for information about the application process. Unfortunately I cannot reply to all e-mail inquiries.

Reason

I am married to the loveliest girl in the world, the light of my life, Eira. ''You are the reason I am. You are all my reasons.''

Since July 31st 2003, the universe revolves around a boy. Since March 1st 2007, we have two boys!

Other

Take a look at (old) Cosco papers visualized using the Similarity Metric of Vitányi and Cilibrasi.

Play a game of Rock, Paper, Scissors: rock beats scissors, paper beats rock, and scissors beat paper. Even such a simple game offers some theoretically interesting problems: Can one predict the other player's choice? What is the best strategy against a good opponent? (Yes, the computer could cheat, but I promise it doesn't.) Such questions were considered by Claude Shannon in the 1950s; see a modern variation of his 'Mind-Reading Machine' based on data-compression (CTW).