Machine Learning Summer School 2012 in Kyoto

Freshly back from two weeks in Kyoto where I attended a Machine Learning Summer School organised at Kyoto University. The whole event was impeccably organised and featured some of the biggest names in the field lecturing on fundamental aspects of machine learning: two highly productive weeks leaving barely enough time in the evening to catch up with friends and enjoy life in my former hometown…

A few random thoughts and observations:

Even though I might have been slightly outside of the target bell curve for this event (mainly designed for people at the beginning of their curriculum in ML, or working in a different field and looking to acquire a good overview of ML), the vast majority of the lectures contained items entirely new to me. And most of the rest still had me racking my brain for long-gone memories of Masters coursework or even undergrad algebra. But the coolest had to be seeing some of the seminal papers I studied years ago during my masters, presented and annotated by their own authors.

Being a resolutely international-minded event (and featuring an exceptionally high level panel of lecturers), this Summer School was easily one of the best academic events I ever attended in Japan (regular conferences included). It also gave me an occasion to test a few of my old hypotheses about what might make some of these other domestic conferences so tedious and useless:

Despite the statistically insignificant size of the sample, it was hard not to notice a strong correlation between quality of speakers (as a lecturer, regardless of their research skills) and geographical origin. While most anglo-saxon (or strongly assimilated) lecturers were consistently engaging and managed to keep their audience interested over long hours of presentation, most Europeans or Japanese lecturers not only lost the entire hall by the second slide, but did not seem to care all that much. There might be an issue of native language skills here (I don’t think so), but the presentation style and contents were undeniably different.

On that note, who the hell thinks it is a good idea to present an average 2-3 equations per slide on a four-hour long lecture, with entire proofs of every single mathematical results in their talk (and very little else)?!?1[Spoiler alert] Answer: apparently the French. That’s who. I like proofs just fine, but I am quite content getting the high-level gloss verbally, with an invitation to check the rest on my own later on. There were way too many times where I swear not a single person in the audience even knew why we needed the original theorem, by the time the proof was over.

Conversely, I noticed an interesting trend across many lectures (that did not deal with purely theoretic topics): while they were each quite eager to introduce their shiny new(ish) machine learning model/framework and go in great details over the properties and advantage of said models, a comparatively very tiny fraction of time was spent discussing inference. Considering some of these models had obvious complexity issues for inference that historically stood in the way of their proper use in real-life (*cough* non-Horn first-order logic clauses *cough*), it would have been nice to get a little more than “and we use methods X, Y or Z to do inference, with heuristic W to speed things up… it’s not as slow as you would think… pinky promise”. If you are going to spend 20 minutes on slides full of equations, why not also describe the inner workings of the inference algorithm and convince me why it will indeed converge and converge well.

Last bit of over-entitled whinging (I swear, the whole event was brilliant and these are all but nitpicking details): what is up with Japan-organised academic events and beverages? And by beverages, I mean water. I like ocha/macha/mugicha/olongcha as much as the next guy2Coffee, somewhat less as the next guy, given that I prefer to save my caffeine tolerance build-up for occasions that truly need it., but sometimes, in the middle of a long, hot, sweaty afternoon, all I want is a glass of cold fresh water. Surely that cannot be hard to provide along with the tea and cola? Hell, while we are at it, a few fruit juices would be nice.

And while on the topic of catering: Japanese (and non-Japanese) organisers should probably keep in mind that international events bring international crowds, with a whole bunch of dietary restrictions. Not labelling snack foods with clear indications of meat content: not that great… Not ensuring there would be a single vegetarian option at the school banquet: really not that smart. While myself happily omnivorous nowadays, I felt pretty bad for the handful of Indians, Muslims and odd Western Veggies that had to spend their entire evening on green salad (and little else).

Anyway, kudos to the organisers for a refreshing two weeks of world-class lectures and presentations.