Maximizing von-Neumann Entropy for Submodular Semi-Supervised Data Augmentation

• Maxwell Morrison, Electrical Engineering, University of Delaware

• Austin Brockmeier, Department of Electrical and Computer Engineering, University of Delaware

The purpose of this project is to enhance the training of a machine learning model by augmenting a labeled training set by selecting additional examples from an unlabeled set per class in order and increasing the overall diversity of the combined training set.  To do this we attempt to maximize the class-specific entropy of the combined data set (as measured by the von-Neumann entropy of kernel similarity matrix of the data instances), subject to a likelihood constraint on a set of unlabeled points selected to augment each class.  Writing the constraint as a penalty, the overall optimization problem remains submodular and thus we propose to use a greedy algorithm to maximize the von-Neumann entropy of the combinational subset while taking into account the probability that the element in question is part of the subset that it is being tested against.