Monday, June 20, 2016

The “I” in IT

This article has been published in the June issue of the Information Theory Society Newsletter, as this month's student column (students are always encouraged to send us their thoughts and ideas!). I gratefully acknowledge Gerhard Kramer's help in improving this article.

Have you ever asked yourself whether IT is the right field of science for you? I never did, but I guess IT wondered more than once whether I am the right person to work on its problems. Here’s my story.

I’m currently a postdoc at the Institute for Communications Engineering, Technical University of Munich, Germany. In this regard, my story has a happy ending, something I wouldn’t have dreamed of years ago. When I started my PhD at Graz University of Technology, Austria, I hadn’t heard a lecture on IT yet. My supervisor and I nevertheless decided to attempt developing a “theory for deterministic information processing”, or what we called it back then. I knew entropy from communications engineering and thought that I could get all the necessary information from Chapter 2 of the “Elements”. I later read Chapter 4 for stochastic processes and even resorted to more detailed results, such as those from Kolmogorov and Pinsker, when I tried taking limits of random variables. Nevertheless, phrases like “a sequence of codes” never appeared in my work. Probably the rest of my remaining scientific credibility would be lost if I told you how many of my papers got rejected - so I won’t. I will tell you, though, what most reviewers agreed on: That the information-theoretic quantities I introduced to characterize deterministic systems lack an operational characterization. That was a valid criticism, and I learned what an operational characterization is only a few months before I obtained my PhD. Relating to IEEE Spectrum’s article on Claude Shannon, I despaired that I had “jumped on the bandwagon without really understanding” (Slepian) and contributed to the “widespread abuse” (Pierce) of IT by other fields.

However, since then I have discovered that this kind of “abuse” is successful in scientific fields outside IT. For example, nonlinear adaptive filters are trained using an error entropy, Rényi and Cauchy-Schwarz divergences are used for learning and classification, and the information-bottleneck method enjoys widespread use in clustering. To the best of my knowledge, only few of these works accompany their objective functions with an operational characterization - mutual information is “just” a nonlinear measure of statistical dependence, and entropy is “just” a statistic that captures more than the error signal’s variance. Remarkably, these heuristic methods often show superior performance. Hence, at least in these other fields, the works are scientifically justified by experiment. (OK, I admit that this is a feeble attempt to justify my work a posteriori. As more self-justification, the Aims & Scope of the IEEE Transactions on Information Theory does state that the journal aims to publish theoretical and experimental papers.)

Should I, therefore, try to publish in other fields and transactions? Probably. My supervisor, suggested more than once that I should spread the word, trying to convince scientists in signal processing to use higher-order statistics, i.e., IT quantities. I haven’t listened, though, because I feel at home in the IT community, and I would not want to miss meeting my IT friends regularly at our annual events. Even in hindsight, I would rather submit to orthodox pressure and provide operational characterizations rather than to publish in a different community. In the future, I hope that I can do both.

That is my decision. As the IT society, we all must decide: what are our traditions (really) and how strong shall we hold on to them? Especially when dealing with future directions of IT, as mentioned in arXiv:1507.05941, how should we contribute to make our quantities be used in fields such as signal processing, control theory, and machine learning? Or should we not? For example, should we propose clustering algorithms using IT quantities, or should we rather focus on deriving asymptotic limits for the clustering problem? You, as a PhD student, must make your own decision: Are you sufficiently “in IT” so as to be accepted by the community? If your research topic is on the border between IT and another field, in which direction do you want to go?