The sessions and talks are over at ICML’10. A variety of powerful information-theoretic concepts with high future potential were discussed in five parallel tracks and invited talks. High-dimensional problems and ‘very large’ data sets were popular this year, approached by online learning, embeddings, sparsity and fast approximations. New nonparametric and mixed membership models were presented along with a number of other fascinating topics- see the proceedings for a comprehensive list. Major commercial players who play a key role in transforming theoretical ideas into practice were also present.
Computational biology, social sciences, and web-associated tasks including multi-label classification, rank optimization, and collaborative filtering were popular application topics. I particularly liked Duncan Watts’ keynote on large-scale experimental social science in the web. The historical scale of recorded social activity in the web allows novel experimental setups in social science- such as studying group-level decision-making. As an example consider a case study where web-based music recommendations were turned upside-down for a group of test subjects- the initially least popular songs started to increase in popularity, and ultimately the reverse ranking became a kind of self-realizing prophecy. Alternatively, have a look at a virtual reprise of the Stanley Milgram obedience experiments. It is worth noting, though, that ethical review standards for such studies are lagging behind.
Complex relational data sets are becoming increasingly available in public domain through open data initiatives in science, politics, and computational biology. Efficient approaches are needed to query and characterize such information. I am curiously looking forward to see what role information-theoretic approaches will play there.