University of Wisconsin, Madison just published a video of my talk on SMART.
After looking through the video, I noticed that my talk took 40 minutes, not 30 as I originally mentioned in a previous post.
Thanks again to the people at Madison for hosting me!
I just returned from University of Wisconsin, Madison, where I spent the last few days talking with faculty, students, and postdocs about their research interests and the SMART algorithm. During that trip, I presented SMART to the SILO seminar, which has a varied audience of mathematicians, engineers, and computer scientists. I’ve uploaded the set of slides here, and probably its most notable feature is its length: just 30 slides, which I was able to present in roughly 30-35 minutes. In technical talks, paying attention for more than 20 minutes at a time can be difficult, so lately I’ve resolved to limit all my talks to 30 minutes, even if I’m allotted 50. The length of my talk went over well with everyone whom I spoke to after the talk.
Some good questions came up during the talk, mostly about the trigger graph, which is a minibatching interface built into SMART, and how its spectral properties affect SMART’s convergence rates. Only a coarse answer to this problem appears in the SMART paper, in particular, in the second half of Section 5.5. But clearly more minibatching results in more speed, as long as the per iteration cost does not rise too high.