A Visit from Marek Smieja

This week I had a nice visit from Marek Smieja, Assitant Professor at the Group of Machine Learning Research at Jagiellonian University in Krakov.

Marek presenting some of his most recent results - you can find the paper on arXiv (and in the upcoming proceedings of this year's edition of NIPS!)

I first met Marek in 2012 on the terrace of Lausanne Palace Hotel - and if that doesn't make a good start for a story then I don't know what does. Drinks in our hands, we were discussing the exciting talks we have heard during the first few days of the IEEE Information Theory Workshop held at EPF Lausanne. Back then Marek and I were working both on information dimension, but we already had the idea that we will soon move towards machine learning.

Fast forward to 2015, when Marek invited me to visit Poland to attend the first edition of Theoretical Foundations of Machine Learning. A year or so later, Marek and I started working on a method for semi-supervised clustering, mixing ideas from model-based clustering and the information bottleneck principle. We have stayed in contact ever since, but this week was the first time we met again since raising glasses on this terrace in 2012. (Marek could not attend the conference in 2015.) Needless to say, I was really happy when he announced that he will be able to visit me at the Know-Center.

Marek in front of the Uhrturm, one of Graz' main attractions.

During two intensive days in Know-Centers "Small Meeting Room" we brought each other up to date regarding our current research projects, found a topic on which we may collaborate in the near future, and defined a student project with the aim to combine Markov aggregation and semi-supervised clustering. I'm looking forward to continuing this year-long academic relationship with Marek!

PS: If you have some hot results on the theory of machine learning ready, you may want to submit it to TFML 2019. Deadline is November 30th.

Santa Fe Institute Spring 2018 Complexity Challenge

At the end of my Schroedinger fellowship I decided to treat myself with participating in the Santa Fe Institute Spring 2018 Complexity Challenge. I've been a fan of the SFI as long as I've known what it does, and I've seen the most impressive papers published by authors affiliated with SFI.

The challenge itself was an agent-based model with rewards, i.e., it was closely connected to partially-observable Markov decision processes and game theory. In my solution, (which earned me an honorable mention) I took neither approach but just played around a bit with strategies I chose ad hoc. Despite this, I was amazed by the rich diversity of behavior I observed. It took me some 40 work hours but it was totally worth the effort. Thanks heaps to the organizers of the challenge, and congrats to the winners -- I hope I get the chance to participate again at some later time!