from Lawrence Saul

I hardly know where to start. Sam was a wonderful colleague and a dear friend. I had the privilege to work closely with him over a period of several years. I am terrible with details: I cannot remember exactly where or when we met, or even how we started collaborating. We were never in the same department and rarely in the same time zone. Indeed, at the peak of our collaboration, we were not only in different countries, but on different continents. Those unfavorable conditions were more than compensated by Sam's boundless energy and infectious enthusiasm. We used to joke that together we worked twenty-four hour days -- because when one of us was eating or sleeping, the other was working, and vice versa. Those were also the days before skype, and it was fortunate that I worked for the phone company because we talked almost as often as we emailed.

I learned so much from Sam. His mind was always racing with new ideas. He had high standards which I later inflicted on my PhD students. In papers and talks, every detail mattered; in research, he never settled for less than all-out effort.

Of course, Sam was not only a consummate researcher. He was also a remarkable person. He had a warm, broad smile and a constant sparkle in his eyes. He spoke alternately with passion and self-effacing humor. He could make a room of PhDs burst into laughter; he could also play gently with small children.

During our work on high dimensional data analysis, Sam and I experienced all the highs and lows that are part of any sustained research effort. In the spring of 2000, after months of flailing (which culminated in missing a critical conference deadline), we finally had our "Eureka!" moment. I have attached Sam's email on that day. I think Sam would approve: he loved to share the excitement of our field. I will always cherish the memories of our friendship.

From: Roweis@gatsby.ucl.ac.uk
Subject: *too* good
Date: May 17, 2000 4:55:40 PM PDT
To: lsaul@research.att.com

This algorithm is absolutely the business. You must call me if you are still checking email.

*Everything* I tried it on worked. I am running out of data for this algorithm to chew up! I almost want it to fail, just to be sure it isn't some kind of trick...

1) The 1D manifolds, of course.
2) I ran it on the translated faces, and it nailed it.
3) I ran it on Josh's swissroll, boom!
4) I ran it on the 8x8 digits and it gave good results, in fact just as good as the best runs I ever had with anything else.
5) I ran it on the 16x16 digits and it gave great results, although hard to display, so now I am working on the geodesic code. That way I'll be able to interpolate between any two items.
6) I ran it on a database of many poses and expressions of a single person's face, again great. Waiting for interpolation on that too.
7) I ran it on a short movie, with the frames shuffled randomly. I told it to learn a 1d manifold. What do you think it did? Bingo, it sorted the frames into the correct movie order (although sometimes it sorts them backwards!)
8) It is running on text now.

Too bad this didn't happen a week ago. It is *perfect* for NIPS...