April 28, 2010 @ 12pm : Brendan Frey

Department of Electrical and Computer Engineering
University of Toronto

Neural networks reveal the splicing code

My research group at the University of Toronto recently discovered a “splicing code” that enables living cells to rearrange parts of genes, turns twenty thousand genes into hundreds of thousands of genetic messages, and generates complex tissue structures such as neural networks. New biology was revealed by defining a machine learning problem and using regularized maximum likelihood estimation with feature selection to solve it. In this talk, I'll describe how the application of good old-fashioned neural with early stopping and Bayesian integration enabled us to discover new biology and make predictions that are useful for medical applications. I'll describe our technique and show that it significantly improves upon the state of the art, which we defined because nodbody's formalized the problem before. Interestingly, nearly half of the models sampled from the posterior perform worse than random guessing, so our method illustrates what can be gained by surfing the sharp edge of Jensen's inequality.

seminars/seminaritems/2010-04-28.txt · Last modified: 2010/04/19 07:22 by koray