Tuesday, July 03, 2012

ICML 2012

When I gave a plenary talk at SODA in 03, I had to give the talk a day earlier because the slotted speaker could not make it to the time slot. At ICML, where I was slotted to give the plenary talk at 8:40AM on the first day, I could not make it to the location and the hosts (Andrew McCallum and Joelle Pineau and Charles Sutton), very graciously, scheduled me at 7 PM the same day by carving out a new slot. This tale of two slots worked out well for many: the North Americans who had flown in and had sheepishly slept thru my early AM slot could make it to my evening slot; the attendees had the poster session just before my talk and had a chance to grab some beer and snacks; and, I could go over the standard 50 min --1 hour slot because I was *only* standing between the audience and their dinner.

I spoke about streaming, summarizing the by-now standard material on sketches, l_0 samples and their applications, in particular to compressed sensing and graph algorithms resp. For the latter, I highlighted the work of Ahn, Guha and McGregor on graph connectivity problems using l_0 sketches. Then I discussed new directions: (1) new models of learning algorithms in distributed world where the goal is to minimize the communication (I highlighted recent work by Daume+Phillips+Venkatsubramanian, Blum+Balcan+Fine+Mansour, see here). (2) Continual distributed streaming (highlighted Cormode+M+Yi) and (3) Pan-privacy (I highlighted this work of Dwork, Naor, Pitassi, Rothblum, and Yekhanin). Besides gratuituous jokes, puzzles and comments on recommendations, that was the talk. From the questions afterwards, it seemed like the audience took in the hashing part of it and discussed ways to use it, I hope the other parts will percolate in later. Over dinner:
  • Some in ML learn about developments in TCS world (solely?) by reading Michael Mitzenmacher's blog. Michael: you are carrying us, mate!
  • Like all communities, ICML is tinkering with the conf. One of the suggestions concerned making reviews (think reviewoverflow) public. A central suggestion will be to separate/delink the identities of authors from reviewers (which you can do on the web) but make identity persistent. 
  • Someone asked me how many pairs of shoes (2) and glasses (6) I had while traveling. I said only 2 shoes because very few extroverted mathematicians at ICML to look at my shoes, the ref to the joke was lost in some cases.
  • Finally, I saw printed NY Times in French. 


Blogger Michael Mitzenmacher said...

1) People read my blog? (Besides my mother?)
2) Machine learning people read my blog?

8:04 AM  
Anonymous Anonymous said...

(1) Yes!
(2) Yes, very much!
And you are the window into theory CS, as I heard.

-- Metoo

11:45 PM  

Post a Comment

<< Home