SmILE Semantic Information Systems and Language Engineering Group

Immediately after leaving IJCAI'09 (see the respective report for details), I went to Verona in Italy to attend AIME'09, a conference on Artificial Intelligence in Medicine where I was to present a full paper about CORAAL. Before the main conference started, I attended the IDAMAP'09 workshop, the 14th event in the series of workshops on Intelligent Data Analysis in Biomedicine and Pharmacology. I registered for the workshop since I wanted to learn something about generic data mining in the life science context (in order to complement my moderate experience in text mining). And in the end of the day I wasn't disappointed, indeed. Unfortunately I missed most of the invited talk of Hendrik Blockeel about relational data mining due to a late train connection. However, I managed to talk to him a bit after that and he pointed me to some very good introductory resources pertinent to the topic, and sent me his presentation after that (I can forward it to interested people), so no harm done eventually.

The regular workshop presentations were mostly about temporal data mining, machine learning and also about medical image recognition. You can check the proceedings for the particular papers. Apart of getting some good pointers to relevant work in this area, which can very well serve as a starting point to expand the use cases for our CORAAL tool, I've learned about a nice modular Python suite of general-purpose machine learning and data analysis algorithms. It's name is Orange and it's being developed as an open source project by the AI lab at University of Ljubljana, Slovenia. Apparently, not only Josef Stefan Institute does cool AI stuff in Slovenia ;) . Oh, and I almost forgot - at the workshop, I've met a guy from a Dutch company RRD, who has just recently started to look into aggregation of raw sensor data in order to automatically determine activity of people (i.e., patients). It seemed to me quite related to some of the work of DERI's sensor group, so I pointed him to Manfred hoping there could be some interesting cooperation (although I personally don't understand half of the detailed technical things the guy was talking about ;) ).

Going back to the conference itself - 24 years after the first event in the AIME series, it appears to be the most relevant global forum for applications of AI in medicine (after cessation of series of similar events regularly held at Stanford until about middle of nineties). AIME is frequented by a relatively small, but rather determined community doing a lot of interesting stuff. Although I was never involved in the community before, I could at least vaguely recall some names of the attendees I had spotted when browsing through AI magazine or AI and AI in Medicine journals in the past. Besides that, there were at least three people very well known in the Semantic Web community, too - Robert Stevens from Manchester and Annette ten Teije with Frank van Harmelen from Amsterdam, who gave talks at a workshop on medical knowledge representation.

The conference had an interesting format - there were three full days of presentations, which were all organised within one "serial" track in a single big room (actually the biggest I've ever presented in, though it was not entirely full at all). Only relatively small number of papers was accepted as full contributions with regular 20+5 minute talk. The rest were short papers with 5 minute presentation slot and a poster. The short presentation was a sort of pitch for the poster, which I found very good - you could focus directly on the interesting stuff only and have as long chat with an author in front of the poster later on as you wished. This way the organisers were also able to accommodate all the 40+ papers into one session.

The main topics of the conference were (temporal) data mining, probabilistic modeling and reasoning, natural language processing, planning/scheduling, case-based reasoning and decision support systems, mostly everything applied to real experimental biomedical data, clinical guidelines or patient records. Although I could sometimes follow the talks only with abundant consultations of the proceedings and related basic references, it was largely about really interesting research as far as I could see. The presented contributions were very often directly driven by real-world problems, if not directly elaborated in cooperation with domain experts. If I was to highlight every paper I found interesting, it would take much more paragraphs than I intend to write (and you probably want to read ;) ). Nonetheless, feel free to have a look at the proceedings (Springer, so you have to do it from a proxy that has an access to SpringerLink if you want to get the papers), then feel free to ask me what a particular paper is about and I'll try to give you a summary - if you'll be lucky, I might still remember something.

Invited talks I have to highlight at least briefly, though, since they were pretty interesting (not quite that usual "visionary" content-less stuff). First keynote was given by Catherine Garbay and was about computer vision, namely about opposing two takes on the field. First of them is a positivistic view, i.e., computer vision as problem solving, the second is a constructivist view, i.e., computer vision as an opportunistic exploration of a realm of data (freely adapted from the talk itself). Catherine basically argued for the latter, given the complexity of the problem and openness of the environment one has to take into account. What I especially liked (mainly due to consistency with my own mindset, so I'm admittedly biased ;) ) was the nice statement about the nature of visual perception, claiming it's not merely about recognition of scenes and the objects in them, but also about placing the whole "image" into one's cognitive context, i.e., memory, and interpreting it then. The second invited talk was given by Carol Friedman. It was a high-level overview (with a couple of most relevant technical details included, though) of very nice application of NLP techniques (rather well-known to me) to the processing of electronic patient records (a rather novel and exciting area for me). Perhaps not that "visionary" (pun intended) as the first talk, but definitely not less inspirational.

Regarding my talk, I was really glad I had a full paper with a long presentation slot, since I didn't have to struggle that much with trying to cover all the content in the talk. After presenting the slides, I felt definitely much better than after the fiasco (at least from my personal point of view) at IJCAI'09 (see the IJCAI'09 (see the respective report for details). I conveyed everything I wanted, I even managed to make some people laugh (deliberately, so it was good ;) ). A couple of people from the audience saved the chair of the session from laborous inventing of a question for me and asked some questions that proved they were really paying attention. Few were interested how the knowledge extraction works, which was kind of expected, since I don't go into that much detail in the slides. Then there was a very curious question from Blaz Zupan about whether the framework I designed for the emergent knowledge integration could be also used as a kind of meta-level reasoner, integrating various inference methods in a kind of conceptual way. Within a bit awkward answering of the question (I had to admit that, technically speaking, it's not possible right know, unless you provide some additional background knowledge) I maybe stumbled upon a possible solution, but I still have to work on that a bit. In any case, it was a very good question, although neither the questioning, nor the questioned guy presumably didn't feel very well after my answer ;) . Abstracting from sometimes awkward questions, it seems that what we (me, Tudor and Siggi) did within CORAAL interests people - within or after the presentation I was approached by people from groups in France, Italy, UK and Canada, who were explicitly interested in trying our stuff on their data. We'll see what will come out of this, but at least it motivates me to prepare a public version of the CORAAL back-end (EUREEKA, an implementation of my core PhD research) as soon as possible, let the people play with it (and presumably flood me with requests on many necessary improvements).

One final rather off-topic remark - it is definitely great to attend conferences in Italy! I usually don't drink coffee, only when I really want to stay awake (then I preferably use my simple but functional "on-stove" Italian (sic!) espresso gadget). But in Verona I couldn't resist having the usually incredibly good espresso 2-3 times a day. Also, it's not that common (but very very nice) to serve wine at conference lunches (for instance Soave, white (a bit surprising for Italy, but very good) regional wine), not mentioning the local food ;) . And finally, I was simply shocked (in the very positive sense of the word) by the accommodation in my B&B, located in a quiet quarter just few steps from the city center. After the luxurious, but sterile Sheraton at IJCAI'09 in Pasadena (not that I really need it, but sometimes the conference organisers just force you into such shameful ventures by picking official conference hotels), the mixture of the atmosphere of some of the more "rural" movies of Federico Fellini and of the holiday-with-your-grandmother feeling was like a balm for me. Let alone the fact that it was about three-times cheaper than the Sheraton hotel (with a conference discount and without breakfast). La dolce vita, as I said in the title - great conference with a very pleasant and "familiar" feeling regarding both the fellow attendees and the general atmosphere of the conference location...