Day 2: Edinburgh, Scotland (UAI Conference)
The view from my window, evening of 26 Jul 2005.
05:00: Woken up by the cry of a gull; roll over and go back to sleep.
07:22: Awakened again by a gull. Funny how gulls wake me, but the entry of our hostel-mate around 03:00 didn't! The Banafolks, who heard it, say it was rather noisy.
07:22 - 08:00: Brush, pack, eat three cookies (sugar, chocolate, and lemon). Yes! Woohoo! I'm ready to go!!
08:00 - 08:10: Check my maps.
08:10 - 08:40: Trot off to UAI 2005, the 21st Conference on Uncertainty in Artificial Intelligence. Met Changhe Yuan, a student of Marek Druzdzel at the University of Pittsburgh, and got caught up on doings at Marek's Decision Systems Lab (DSL). I last saw Marek at AAAI 1998 and missed him at UAI 2004 in Banff, Canada.
08:40 - 08:45: Arrive at UAI; greet Carmel Domshlak and Judy Goldsmith. Carmel presented a paper at my workshop on Real-Time Decision Support and Diagnosis Systems at the 2002 multi-conference in Edmonton, Alberta, Canada. Co-organized with hpguo, Gene Santos, and Eric Horvitz, this workshop was jointly hosted by KDD, AAAI, and UAI. As for Judy, I've seen her at almost every UAI; she always asks good questions.
08:45 - 08:54: Go over my map once more with my folks; confirm a 17:40 meeting time with them.
08:54 - 09:01: Get signed on via wireless. The University of Edinburgh has temporary passwords and guest accounts printed on sheets that one has to sign for. Fire off a couple of e-mails home to the USA (KDD, FOB, and TEUNC). Say hi to Kathy Laskey and inform my first colleague from another university that I'm now tenured. Fun. Sit up and shake hands with Prakash Shenoy.
09:01 - 10:45: Ben Taskar's tutorial on large margin methods.
- I took my seat around one minute after nine o'clock; people were still trickling in, though, so Ben waited until 09:11 to start.
- Another object lesson in projection problems; the projector cut off the right-hand side of Ben's slides. (Edit, 17:25 CST Tue 09 Aug 2005 - It turns out that using the auto-size option on the projector would have fixed this, but this was not discovered until after two or three tutorials.)
- What an interesting talk. This one is relevant to: machine translation, analytical methods, bioinformatics (proteomics), SLAM, IE (WebKB), pattern recognition, SVM, and the GM Wiki. Nice! I'll be posting notes from this tutorial to rizanab, comptranslation, massforge, geckies, and bayesnets.
- A state of cortical hyperactivity, and no mistake.
10:45 - 11:10: Break - "be back by 11:05", we are told.
- 10:46: Say hello to Irina Rish, a colleague of many of my former classmates who went to IBM's T.J. Watson Research Center in Yorktown Heights, NY, and to Bozhena Bidyuk, a Ph.D. student of Rina Dechter at UC Irvine. Irina and I catch up briefly on IBM folks who came out of chambana, including Ricardo Vilalta, Dan Oblinger, and Mark Brodie.
- 10:48: Kathryn Janeway, captain of the Federation starship Voyager, once called coffee "the finest organic suspension ever devised". Years later, in the finale episode ("Endgame"), her older self declared, "We drink tea now." Word. If she can switch, so can I!
- 10:52: Talked to Sean Guarino of Charles River Analytics about BNet. Ann Feltz, who once contacted me about BNJ, is no longer with the company; the point of contact is now Jonathan Pfantz.
- 10:55: Say hello to Petri Myllymäki, a friend from the University of Helsinki in Finland whose work I've followed since 1996. I've met him at every AAAI or UAI I've been too, starting with AAAI-1997 in Providence, RI.
- 10:56: Greet Fahiem Bacchus and remind him that I sent him e-mail about BNJ. Note to self: Fahiem requests a resend. Judy Goldsmith expresses interest in BNJ, so I promise her a demo later using zengeneral's latest release, v3.1.
- 11:00: Talk to John Mark Agosta of Intel Research about the NSF FIBR and NSF ITR projects. We discuss some of his new research and John Mark refers me to Michael Black regarding a question I had about intrusion detection systems (IDS). Slowly, my brain starts cranking on distributed data fusion in information security, IDS, and virus detection. I mention that I once did a project on virus detection (it was an independent study project I completed at Hopkins in 1990 under Steve Salzberg, and led to my first paper in a programming journal called MacTutor (now MacTech). The paper is available here and presents an implementation of k-approximate dynamic programming and randomized string matching algorithms with application to detection of hand-modified evolutionary viruses.
- 11:05: Still need coffee(ine)!
- 11:06: Ahh, much better. I spot Haiqin Wang of Boeing, whom I last saw at the Edmonton conferences in 2002, when I was there with hpguo, Ben Perry, and Julie Thornton, my students at the time. Haiqin is here for a Bayesian Modeling Applications workshop that we really have to submit a paper to next year. (KDDers, please take note!) We catch up about hpguo, who has gotten his Ph.D., written a textbook, and gone from postdoc to faculty member in the nearly three years since he and I last saw Haiqin.
11:10 - 12:45: Zoubin Ghahramani's tutorial on Gaussian process-based regression.
- Good stuff. Zoubin is at the Gatsby Computational Neuroscience Unit at University College London. I actually talked with Geoff Hinton back in 1998 about applying to the Gatsby Unit for a postdoc, but the timing of my graduation was a bit off, so I ended up going to the National Center for Supercomputing Applications (NCSA) instead.
- This time the presentation started a little behind schedule again, at 11:10. Usually we make up the creeping lag during the lunch break, which has some slack. They always say we are going to shorten the coffee and tea breaks, but we seldom do!
- Be prepared: know the experts. David MacKay of the University of Cambridge is going to give the banquet talk later in the conference, about an applied project that is based on some of the technical methodology in Zoubin's tutorial.
- Halfway through the talk, Zoubin had to switch off the overhead lights for about 15 minutes. Alas that I didn't bring my phial! (It's actually a single-LED hands-free Energizer flashlight, but I can't find a picture online.) I had to rely on my night vision. Halflings have 30' infravision, right?
- Excellent talk - relevant to: FIBR, ITR, linear regression, Gaussian processes, prediction, (nonlinear) time series analysis, nonparametric methods, Dirichlet processes, and applied probability. I'll be posting notes from this tutorial to rizanab, geckies, bayesnets, tessier_ashpool, and stat_geeks.
12:45 - 13:59: Went to Kalpna, a vegetarian Indian buffet, with Zoubin Ghahramani (UCL, Gatsby Unit) and Kevin Leyton-Brown (University of British Columbia, Computer Science department). Discussed Markov chain Monte Carlo (MCMC) methods, hierarchical mixtures of experts, and analytical versus numerical (optimization versus sampling) methods for hyperparameter learning with Zoubin on the way there; listened to Kevin explain some basics of auction theory during lunch. Kevin tells some stories of how he ran demo auctions in his class and felt obligated to enforce the outcomes, even when it meant taking money from students, though he felt bad about doing so and will be donating it to charity. (Here's a great example of classical conditioning!)
14:00 - 14:25: Quick break to put out some fires back home by e-mail.
14:25 - 15:40: Carlos Guestrin's tutorial on sensor networks.
- Carlos went through 200 slides in 1.5 hours. With references. ("I'm glad Zoubin discussed Gaussian processes, so we can go on from there...")
- For comparison, even counting 50 slides as having been skimmed or skipped, 100 slides per hour is about double the speed of David Heckerman and 3-4 times the speed of banazir. All of you who say I go too fast should come listen to one of these gentlemen, and you'll stop complaining.
- Another great tutorial. This one is relevant to: sensor networks, robust software engineering, FIBR, EcoGen, wireless networks, networking, and security. Notes are going on rizanab, bayesnets, networking, wireless, diistributed computing, security - plus tessier_ashpool, of course.
- "The goal was to get you all excited." Well, Qapla'
15:40 - 16:10: Another break. We drink tea now, we drink tea now, we drink tea now. (Edit: 00:10 CST Wed 10 Aug 2005 - I posted this to statements and was asked "who blew what up?")
- Chat with Carlos Guestrin (formerly one of the Kolleri who has graduated from Stanford and is now a new faculty member at Carnegie Mellon University) about FIBR and EcoGen data. Carlos is very interested in the possibility of looking at sensor data from our projects, so I make an appointment to talk to him later in the conferences about how we might work together.
- Talke to Fahiem Bacchus of the University of Toronto about teaching graphical models and AI and his visualization priorities for bayesnets (variable elimination, sampling). We also discuss BNJ's potential uses as a research tool.
- Talk to Changhe Yuan from the University of Pittsburgh about my projects. Discuss the ITR project and the crucial issues of representation and generalization (Hamming distance versus spatial distance), the FIBR project and time series, dynamic graphical models, and dynamical sytems models therein. The conversation wends through his career plans, bioinformatics and medical informatics in general, and DBN learning (structure and parameters).
16:10 - 17:42: Satinder Singh's tutorial on mechanism design.
- This one is a bit out of my area, but it's fascinating, especially after Kevin's preliminary tutorial.
- Singh's talk seems very technical to me, but that's probably because I'm so ignorant of computational intelligence in auctions, game theory, and economics in general. This talk is relevant to both economists and comptuational game theoreticians.
17:42 - 17:55: Wait for the folks; check and answer more e-mails.
17:55 - 20:00: Walk the Royal Mile up to Edinburgh Castle, looking for... a Chinese restaurant. Finally find The Orchid on Castle Terrace in the old town, and eat there.
20:00 - 22:05: Guild of Venturers, redux - find out way home.
22:05 - 00:00: Freshen up, kick back, process photos from London.
00:05 - 00:49: Cat nap, kindasorta.
00:49 - 03:40: Read more of Harry Potter and the Sorcerer's Stone. Tell me, Mister Anderson... what good is a rememberall if it's continuously on? It's a fishie paradox, and that's a fact.