Banazîr the Jedi Hobbit (banazir) wrote,
Banazîr the Jedi Hobbit

  • Mood:
  • Music:

Day 4: Get Your bayesnets to Mars

UK Trip 2005: A Tronkie Travellogue
Day 4: Edinburgh, Scotland

Left: My room at the Scottish Youth Hostel.
Center: The hostel kitchen, photographed from just inside the door.
Right: The view from my window, around 04:30 GMT Thu 28 Jul 2005.

08:00: Wake up.
08:10 - 08:45: Get up and dressed, drink some milk and have a quick bite, and pack for the conference.
08:45 - 08:58: Walk up to the Appleton Tower under the Potterow underpass. It's overcast today and drizzling a bit throughout the day.
08:59 - 09:25: Check mails and reply to one from miyeko, and then realize I should be getting up to the talks.
09:25 - 10:50: Session 3 of the UAI technical talks, on Extended Models. Am up to 12 pages of notes from the conference proper! (Generally my blog notes on Edinburgh fit on a single page per day; but of course, the technical content, the real reason I came here, takes up more space. I also have to write faster when listening to a talk and thinking about new research ideas.)

Edit, 09:55 CST Wed 10 Aug 2005: It's rather a good thing that I have the UAI program to refer back to. I stopped taking written notes on non-conference-related material for 3 days (the 28th, 29th, and 30th) and if I had to rely on my memory, the result would be the same as it was for UAI-2003 and IJCAI-2003 in Acapulco: a gaping hole in my LJ that remains unfilled after two years.

Here are the titles and my first impressions. The name of the person who presented is underlined.

  • A. "Evidence with Uncertain Likelihoods", by Joseph Y. Halpern and Riccardo Pucella. A metareasoning talk; I missed most of it but it's high on my reading list.

  • B. "'Say EM' for Selecting Probabilistic Models for Logical Sequences", by Kristian Kersting and Tapani Raiko. Relevant to hybrid models.

  • C. "Belief Updating and Learning in Semi-Qualitative Probabilistic Networks", by
    Cassio P. de Campos and Fabio G. Cozman". Qualitative/quantitative hybrids are very popular with applied research labs: national labs to an extent, U.S. defense organizations (DARPA, ONR, NRL, ARL, ARO, AFOSR, MDA) more so. A lot of research (some rather ad hoc and shoddy) on fuzzy reasoning and Dempster-Shafer theory has gotten through this way. OTOH, Fabio and Linda van der Gaag always turn out good papers on this topic. Unfortunately, I don't think Fabio came this time; I didn't see him at all.

  • D. "Expectation Maximization and Complex Duration Distributions for Continuous Time Bayesian Networks", by Uri Nodelman, Christian R. Shelton and Daphne Koller. An excellent paper and a great talk. Uri was going to present it, but because he couldn't come to Edinburgh, his advisor presented.

Around 09:33 I notice that zengeneral is on and ask him for a copy of BNJ v3.1. He sends it to me and we go back and forth for a while getting it to work on Hirilonde. By 10:30 he sends it to me.

10:50 - 11:15: Break. I revert to coffee because I don't seem to be getting enough caffeine from the tea. I can't figure out how to get Demerara sugar to dissolve properly. Must post about this and see if people know.

I talk with David Poole about the ITR and FIBR projects, and also about the relational learning problem in the currently-shelved STARWARD project and the GRoWE (the Consortium for Global Research on Water-based Economies) Targeted Excellence project. He points me to the page for his project with GeoReference Online, Ltd. and we discuss the vagaries of geospatial annotation and domain knowledge elicitation for a while. I then say hello to Frederick Eberhardt and talk with him a bit about the NSF CCLI (Course, Curriculum, and Lab Improvement) project.

11:15 - 12:40: Session 4, on Message Passing Methods.

  • "On the Optimality of Tree-reweighted Max-product Message-passing", by Vladimir Kolmogorov and Martin J. Wainwright. I wonder whether he is related to "the" Kolmogorov, but I decide that he must get that too much and that I can just look it up. There needs to be a section in the GMWiki about inference by message passing (max-product and sum-product).

  • "Sufficient Conditions for Convergence of Loopy Belief Propagation", Joris M. Mooij and Hilbert J. Kappen. You might think that LBP is yesterday's news, but it's going strong after eight years at UAI, and I'm going to have it in BNJ if it's the last think I do! There needs to be a section in the GMWiki about belief propagation.

  • "Structured Region Graphs: Morphing EP into GBP", M. Welling, T. Minka and Y.W. Teh. (Edit, 11:10 CST Wed 10 Aug 2005: Tom presented this paper instead of Max, even though Max had to stand in for Xing, Yan, and Hauptmann at the second poster session because none of them could get their travel visas straightened out.) This was very good. There needs to be a section in the GMWiki about region graphs.

Meanwhile, I'm testing out the various features of BNJ v3.1, which never ceases to impress me, and planning to give some demos for Fahiem Bacchus and Judy Goldsmith. I suspect the Danish people in the row behind me are getting an eyeful, and it becomes a moot point when zengeneral, hearing that I have Danish folks behind me, IMs:


in a nice 48-point font. Fortunately or unfortunately, no Bayesian network software riot ensues: the people behind me are from some other company and not Hugin A/S.
12:40 - 13:56: Lunch with Haiqin Wang, Changhe Yuan, and Jin Tian. We find a Chinese buffet and I just have to laugh. Changhe gets curious about the tenure process and asks me and Jin some questions. It feels odd to be the technically senior person at the table, especially since everyone there is more extensively published at UAI than I am, but I do my best to give Changhe some tips. Jin has done very well for himself funding-wise, has been to AAAI, and will be staying for IJCAI, but is stranded in the UK due to his not getting a reentry visa into the USA. We commiserate and discuss this for a while. I give them all an update about hpguo over a long dessert (vanilla ice cream with strawberry and chocolate syrup, banana fritters, and what I thought were lychees but turned out to be pickled onions; how Dumbledoric is that?).
13:56 - 14:10: Get into the room. I first show Haiqin, Changhe and Jin a demo of BNJ v3.1 (thanks, zengeneral). All seem duly impressed; Haiqin recommends that I submit a paper to the Bayesian Modeling Applications workshop, which is going to be collocated with UAI-2006 in Boston next summer, and I agree to do so. We take our seats and I check e-mail for a couple of minutes.
14:10 - 15:15: Invited Talk, "A Walk on Mars: Managing Uncertainty Through Model-based Programming", by Brian C. Williams, MIT. NASA, NASA, NASA, NASA, mushroom, mushroom! We like the Mars!
15:15 - 15:50: Poster Session highlights.
15:50 - 17:30: Poster Session 2 of 2. Especially interesting papers for me:

  • "Bayesian Logic for the 23rd Century", Laskey and daCosta, GMU - on multi-entity Bayes nets (MEBN)

  • "Nonparametric Bayesian Logic", Carbonetto, Kisynski, de Freitas and Poole - see also Milch et al.'s paper at IJCAI-2005. I knew this was coming!

  • "Exploiting Evidence in Probabilistic Inference", Chavira, Allen, and Darwiche.

  • "Hybrid Bayesian Networks with Linear Deterministic Variables", Cobb and Shenoy. Good stuff; very interesting theoretical advance, and an important discussion at the poster where Koller and Friedman posed an important question about infinite representation.

  • "Detection of Concept Changes in Time-Varying Data Streams", Ho and Wechsler.

  • "Expectation-Propagation for Continuous-Time Bayesian Networks", Nodelman, Koller and Shelton.

  • "Mining Associated Text and Images with Dual-Wing Harmoniums", Xing, Yan, and Hauptmann. Funny how Max Welling had to present the highlight for this poster on behalf of Eric Xing et al., but didn't present his own. He proceeded to "pull a masaga" to much hilarity.

17:30 - 18:25: Meet up with the parental units and stop into a convenience store to get a couple of umbrellas. Tell them I should be home from the UAI banquet by about 21:20 but may leave as late as 21:00, maybe a little later.
18:25 - 18:55: Meet up with Changhe Yuan and walk over with him and a large group to the Roxburghe Hotel for the UAI-2005 banquet.
18:55 - 19:55: Sparkling champagne and sparkling conversation in a receiving room that is too small for the two hundred people crammed into it. UAI always likes to be at a university, a little apart from IJCAI, but it likes to have its banquet at a nice posh place. In Seattle (2001), it was Pier 66; in Edmonton (2002), it was the ground floor of a very nice hotel; in Acapulco (2003), it was the top floor of the Hyatt overlooking the Bay of Santa Lucia.

I catch up with Vu Ha, whom I last spoke to at UAI-2002, and he tells me and Changhe about doings at Honeywell.
19:55 - 20:55: Dinner is served. I speak with Suzanne Mahoney about the quirks of defense projects and we share some fun stories about DoD-funded research we've worked on. Changhe discovers who my uncle is and inquires a bit about what I learned from him growing up. I sit at a table with Changhe, Suzanne, Judy Goldsmith, Bruce D'Ambrosio, Russ Almond, Peter Norvig, and another attendee whom I don't recognize. I decide to ask Peter whether he will talk a bit about LJ and Orkut in his invited talk tomorrow, and he says he wasn't planning to discuss social networks but can mention them.
20:55 - 21:40: David MacKay's fascinating talk on Dasher, a fantastic system that uses arithmetic coding as a user interface for paraplegics and quadraplegics. Russ asks me some questions about Dasher's possible use in typing Chinese as David demonstrates how it can spell in Hiragana. (Edit, 11:30 CST Wed 10 Aug 2005 - the next morning, I asked David about Dasher's possible applications: for Braille, as a machine translation interface, or as a "learn by typing" language learning tool, and he told me that a Pinyin version was due out the very next week!)
21:40 - 22:20: Walk home through a wind-driven rain. While on Prince's Street between the Sir Walter Scott memorial and the Balmoral with its unmistakable clock tower, I am able to raise Banamum on my GMRS radio walkie-talkie. "Great!" I think. That's about a mile with lots of buildings between us! Except for some times when the signal is interrupted by tall buildings, we maintain communication as I cross the North Bridge down to the Royal Mile.
22:20 - 00:40: Home again, home again... but where's Banadad? Gone looking for me. Oops. We leave a note at the hostel and go up the Royal Mile, north on Lothian, and back down, but don't go all the way to the Roxburghe. By a quarter after midnight, Banamum and I decide to ask the hostelliers to call the city police just to check whether anyone fitting his description got lost or injured. Just after they call, Banadad arrives. It turns out that Banadad went up to the Roxburghe (sans walkie-talkie) looking for me, but nobody there could tell him where the Roxburghe was, and he got lost. It took him two hours to get home in the dark and rain. I'm just glad he's all right. Note to self: if you take communication devices on holiday travel, make sure everyone has one and has it on!
00:40 - 04:40: After all that excitement, it takes me a while to get to sleep. I finish Harry Potter and the Sorcerer's Stone, take some more pre-dawn photos, and think a bit about tomorrow.

Tags: bnj, conferences, edinburgh, research, scotland, teaching, travel, travellogue, uai, uk

  • Post a new comment


    default userpic

    Your reply will be screened

    Your IP address will be recorded 

    When you submit the form an invisible reCAPTCHA check will be performed.
    You must follow the Privacy Policy and Google Terms of use.