LOSS FUNCTION ADJUSTMENT FOR INCREASED CLASSIFICATION MARGIN

Loss function adjustment for increased classification margin is performed by applying a neural network to a batch of data samples to obtain a plurality of feature vectors, each data sample in the batch corresponding to a feature vector among the plurality of feature vectors, each data sample in the batch including a label indicating a class among a plurality of classes, clustering first feature vectors among the plurality of feature vectors that correspond to data samples including labels indicating a first class among the plurality of classes to separate the first feature vectors into a plurality of clusters, and adjusting parameters of a loss function to increase a margin between a second class and a cluster among the plurality of clusters that is closest to the second class in a feature distribution space.