I seem to have gotten all behind on wrapping up the ISIT blogging, so the remainder may be more compressed takes on things. This is not in the compressed sensing world, where the signals are sparse and my comments are meant to reconstruct, but more like lossy compression where (for the Gaussian case).
Abbas El Gamal gave a very nice plenary on “Coding for Noisy Networks” in which he really brought together a lot of different eras and streams of work on network information theory and tried to tie them together in a conceptual framework. There was a nice mix of older and newer results. The thing I liked best about it was that he was very optimistic about making progress in understanding how to communicate in networks from an information-theory perspective, which counteracts the sentiment that I heard that “well, it’s just too messy.”
Te Sun Han gave the Shannon Lecture, of course, and he used his time to give a tutorial on the information spectrum method. I had tried to read the book earlier, and honestly found it a little impenetrable (or rather, I wasn’t sure what I was supposed to use from it). The talk was more like reading the papers — concisely stated, but with a clear line of intuition. I know some people are not a big fan of Shannon Lectures as tutorials, but I think there is also a case to be made that most people are unfamiliar with the information spectrum method. A nice example he gave was to show when the output of an optimal source coder looks “completely random.” Maybe this has been done already, but is there a connection between existing theories of pseudorandomness and the information spectrum method?