How much do you trust your model?

Abstract: 

I suggest considering two different “regimes” when making predictions from data. The first deals with relatively simple phenomena and expensive data. Here, it is sensible to invest in building an accurate model and inferring correctly from it. The second regime deals with more complex phenomena and bigger data. Here one must cope with the fact that a perfect model is unrealistic, and computational issues force the use of approximate inference. Further, it makes sense to include high capacity predictors (e.g. deep learning) into the model to leverage all the available data. I will discuss research in both these regimes, particularly focusing on graphical models.