Mathematics Colloquim

Speaker:     Dr. Zerotti Woods, Johns Hopkins University

Abstract:    In this work, we present a new regularization term that penalizes the conditioning of the weight matrices in a deep neural network.  We give a mathematical argument that suggests that in certain situations, the conditioning number of the weight matrices have a direct impact on the error in classification.  Empirical evidence suggests that improving the weight matrix associated with the output layer of a matrix improves generalizability when classifying ECG data from a benchmark data-set, and also a novel malaria infection data-set.