Three-Toed Sloth

E Pluribus Unum (This Week at the Complex Systems Colloquium)

Attention conservation notice: This is an attempt to increase the attendance at this semester's complex systems colloquia by blogging about them in advance. Of minimal relevance if you're not in Ann Arbor or don't care about complex systems, statistics, machine learning, improving your data analysis, or collective cognition and decision-making.

For the final colloquium of the semester, we are very happy to have Prof. Ji Zhu from the UM Statistics Department explain to us why many models are better than one.

"Boosting: A Statistical Perspective"

Abstract: Boosting is a method for combining multiple models for predicting or classifying noisy data. The idea was first introduced by Freund & Schapire (1996), with their AdaBoost algorithm. AdaBoost and other boosting algorithms have attracted a lot of attention due to their great success in data modeling tasks. They are now being applied to medical diagnosis, bioinformatics and genetic modeling, chemical process control, shape, handwriting, speech and face recognition, financial modeling, and a wide range of other important practical problems. In this talk, I will describe the essence of boosting from the point of view of a statistician.

Boosting is one of a large family of procedures in statistics and machine learning, with semi-cute names ("bagging", "stacking", "binning", etc., etc.) for combining predictive models or classifiers, which tend to work very, very well, for reasons which have not been at all clear. After all, taking the average of a large number of simple models is equivalent to building a single really big model, which Occam's Razor tells you not to do, because it's a recipe for over-fitting. Yet boosting works --- why? Ji has a profound understanding of boosting and its kin (he was too modest to mention a novel application he's worked on, so I will), but his talk will be quite accessible. (I've seen the slides.)

4pm, Thursday, December 9th, in room 335 West Hall, Central Campus.

Complexity; Enigmas of Chance; The Collective Use and Evolution of Concepts