Christian Queinnec made the most persuasive argument in favor of computer-assisted teaching I have seen to date. A surprising result is that it turns out that teaching Scheme to undergraduates helps more women succeed in computer science. As this issue is right now in a pretty catastrophic state (based on hard enrollment numbers, an outside observer could only conclude that women are being actively excluded from CS education), why not accomplish two great things at once and teach more Scheme?
Hannes and Andreas's talk on Dylan macros for destructuring and manipulating binary data was good (too much code-reading for my tastes, though). The system they built (Network Night Vision) looks great (if you are interested in what you can do with Dylan and DWIM user interfaces, it is also very much worth looking at).
On the topic of user interfaces, Peter Herth's talk on LTk had some really neat demos.
In terms of the quality of presentation Charlotte Herzeel's talk on the HALO temporal-logic-selected-pointcuts AOP tool was my favorite.
One of the themes on Monday was computer vision. Chris Connolly gave a talk about SRI's FREEDIUS system, which is used for geospatial (surveillance) computer vision, and in the afternoon Cyrus Harmon talked about the image processing system he developed to categorize Drosophila melanogaster (a kind of fly) imaginal discs (undeveloped limbs/protruding squishy things) where different genes in different discs at different stages were colored and photographed, producing a whole bunch of images that needed to be accurately categorized. Cyrus developed the ch-image image processing library to help in this regard.
In both FREEDIUS and Harmon's PhD system, all of the low-level image manipulation and feature detection code is written entirely in Lisp (Connolly said that in FREEDIUS Gaussian smoothing is done in either C or Lisp and that they are probably going to drop the C function and do it from Lisp because the performance difference is not significant). So there you go - if you did not want to believe Didier Verna's benchmarks, now you have proof that there is absolutely no reason not to do all of your performance-intensive array processing tasks in Lisp.
Michael Sperber gave a talk on the R6RS standardization effort. R6RS will have many great new features. Combined with Olin Shiver's unveiling of the Scheme loop macro at Danfest, there will be a lot of new reasons to use Scheme.