|
Abstract
|
This thesis aims at applying the scientific method to machine learning along two lines of
research.
The first one builds on recent work applying modern experimental design for algorithm
selection and hyperparameter tuning (Brendel and Schoenauer (2011), Lacoste et al. (2012), Bardenet et al. (2013), Misir and Sebag 2013). We will explore the interaction between methods (and hyperparameters) and data sets to find out whether and to what extent experience can be
generalized across data sets. The output of this project is a toolbox for practitioners and a stockpile of knowledge on what algorithm works on what (kind of) data sets. T
The second one asks the question of why certain methods work on certain data sets. There are several principles, that explain the success of methods and guide algorithmic development; our
goal is to verify these principles and to discover new ones based on an empirical approach.
|