Neural network trained using ordinal loss function

Abstract

Training an ordinal mapping deep neural network (OMDNN) can include receiving multiple samples, each a computer-processable data structure corresponding to a real-world object and including a data element indicating one of n predefined classes to which each sample is linked. Each sample can be mapped by the OMDNN to sample points of a multidimensional space. The OMDNN can predicts the class of each sample based on an ordinal mapping. Parameters of the OMDNN can be iteratively adjusted in response to misclassifying one or more samples. Iteratively adjusting the parameters can be based on an expected loss determined by an ordinal mapping loss function that measures (a) distances between each sample point in the multidimensional space and each other sample point of the same class and (b) overlap between sample points of different classes.