A1202
Title: Automatic selection of hyper-parameters via the use of softened profile likelihood
Authors: Gengyang Chen - University of Waterloo (Canada) [presenting]
Mu Zhu - University of Waterloo (Canada)
Abstract: The purpose is to extend a heuristic method for automatic dimensionality selection, which maximizes a profile likelihood to identify "elbows" in scree plots. The extension enables researchers to make automatic choices of multiple hyperparameters simultaneously. To facilitate the extension to multi-dimensions, a "softened" profile likelihood is proposed. Two distinct parameterizations of the solution are presented, and the approach is demonstrated on elastic nets, support vector machines, and neural networks.
