So far this conference has been great, with a good showing of some heavy hitters in the field of signal processing.
For about an hour, my Mallat number was near zero as I watched from near the front row Stéphane Mallat give an excellent presentation of recent work by one of his PhD students. Of course I didn’t understand all of it, but he made me feel as if I did! My take on the work is that it is the creation of a feature space that is invariant to many of the deformations present in, e.g., databases of written digits. His focus here is no longer on perfect reconstruction of a signal, but perfect classification no matter the deformation. The method even has some parallels with quantum mechanics; and the automatic digit recognition rates were convincing, even when random textures were applied to the digits — which destroy any notion of all digits being efficiently representable as points on some low-dimensional manifold. Since by the end of the talk we had not coauthored a paper, my Mallat number jumped back to around 4 or 5.
Another great talk was by Michaël Unser, on generalizations of 2D steerable wavelets to multiple dimensions for performing denoising of biomedical images. The results were pretty amazing, and demonstrations can be found here.
My course this coming semester is a graduate seminar in interactive systems programming; and so the talk by Roderick Murray-Smith gives me essentially all the material I need for my first lecture to inspire my students to do something exciting with interaction. (Murray-Smith’s colleague Stephen Brewster will be at HAID2010 at AAUK in just a few weeks time.)
Martin Vetterli gave the most humorous talk, and at the same time trumpeted a very important message: Reproducible Research. If we as scientists put in a little extra effort to make the results and algorithms described in our publications clearly available, either by sufficient description, or by downloadable software, and if we make meaningful and sufficient comparisons with other algorithms where appropriate, then we will simultaneously improve the quality of our work, and increase our achievements. As an added bonus, the evidence suggests that research made reproducible is cited more, and more quickly than research that is not; and that once you have started making your research reproducible, the time it takes you the next time decreases. The idea isn’t all roses though, as the vast majority of reproducible research is in the form of MATLAB code. I am lucky to have an institution license, but many others are not. There is also the fact that one day the various websites with code that we point to in our papers will not exist any longer.
Finally, Pierre Comon gave the pre-lunch lecture today on tensors in signal processing. An intriguing property we learned is that tensors can have a higher rank than they have dimensions. Tensors requires me to think in a way that feels to me like thinking in four dimensions. (That sentence also had more rank than dimensions.) I get the impression that someday, everything will be tensors, and linear algebra will be taught in elementary school.
Finally finally, I learned today of Hedy Lamarr: Hollywood starlet and inventor of torpedo guidance systems (actually spread spectrum communications). Her co-inventor on the patent was American composer George Antheil, who composed the riotous composition Ballet Mecanique. Now THAT is an excellent confluence of historical personalities. I wonder what my Lamarr number is…