Suisse Muhurtham

Hola, Señor Mercurial Dhanraj Pillai.  Greetings from the site of the recently concluded World Cup of field hockey. 

You really strengthened your membership in the information theory fraternity during that summer of 2006 in Lausanne, didn’t you?  One result of the connections you developed then was that Sonia and I were able to enjoy the wedding of Dinkar Vasudevan last week in Delhi.  It was a good event.  Even though Dinkar and I had never met before, he knew exactly who I was when he received us – thanks to you. 

That summer of 2006 was good for me as well.  Getting to know Nikos Paragios, getting going with variational level set methods, and doing a bit of work on medical imaging turned out well.  I ended up using variational level set methods in my recently completed Ph.D. thesis: to develop a new algorithm for the machine learning problem of supervised classification, which I termed the geometric level set (GLS) classifier. 

Typically, level set methods are employed when the data is on a pixel or voxel grid.  With the GLS classifier, I broke away from that mold and allowed the data to be in any general Euclidean feature space.  In the final section of my thesis, I suggest that variational level set methods and other similar geometric partial differential equation-based techniques be applied not in pixelized spaces, nor even in general Euclidean spaces, but in the space of probability distributions with associated information geometry.  Any thoughts?

Suisse Sudoku

Yes, Kush, I did notice the Cornell references on Day 2 of the LIDS Student Conference.  In fact, just after that, I had gone to San Diego for the Information Theory and Its Applications Workshop where Yedidia gave another version of the same talk and then Jon Kleinberg from Cornell gave the keynote on Cascading Behavior in Complex Networks which was pretty interesting.  A particularly nice philosophical point that Kleinberg made later during the panel discussion was on the difference between networking (in the engineering sense) and the emerging field of network science.  In the first, everything is designed from the bottom-up and is to be understood deductively whereas in the second, even if things are designed, they get out of hand pretty quickly and therefore must be understood in a top-down inductive style governed by the scientific method ideal of making hypotheses and testing them with experimental and perhaps generative modeling approaches.

Coming back to your reference to the puzzle Sudoku, as I briefly mentioned in an article about information theory and Pingala’s Meru Prastara, there are strong connections between coding and Sudoku.  In fact the connection between puzzles and information theory is very strong, dating back to the seminal work of Shannon, where in Section 7 he made a comment about the feasibility of making crossword puzzles in various languages (see Section VIII of this paper by Immink, Siegel, and Wolf for more on crossword puzzles).  As you know, Shannon was big on toys in addition to puzzles.  For example, he created the ultimate machine which we referenced in the penultimate machine we made for the ITA Workshop with John Sun and Baris Nakiboglu at the request of Rüdiger Urbanke who was in charge of the entertainment portion of the workshop.

I don’t know if you know this, but the only time I’ve actually done a Sudoku puzzle is on the train from Paris back to Lausanne after I had visited you during the summer of 2006, where I was spending the summer with Rüdiger and Emre Telatar working on some things related to low-density parity-check (LDPC) codes: suisse sudoku if you will.