Why not perform well in identifying areas?
The reason is that the definition of Softmax itself, it is officially softmax function plus cross-entropy loss. The aim is to have the maximum number of all classes of space in the probability likelihood of ensuring that all categories are correctly classified, as well as the revocation and validation tasks needed is a better generalization performance metric space. A metric space, to ensure proper classification and to ensure good generalization. Although the correlation is very strong, but not directly the same.
Softmax able to amplify the difference between the minor category logit
Focal loss
Motivation: difficulty data of different training
methods: reduced easily separated sample weight loss
Center loss
Motivation: reducing the distance class
method: for each sample decreases from the center of the square of the distance class of
the class centers update method: In each of the mini-batch updated cluster center. loss for the class to sample away from the center.
Contrastive center loss
Motivation: reduce the distance between the classes at the same time increasing the distance-based
method: the distance between the small penalty classes, the denominator is the distance between the same distance form the center point of the different classes.
Ring loss
motivation: characteristic normalized or weight normalized network can control the degree of attention to simple or difficult samples to a certain extent. Specific point, if not restrained, so that the network always desirable features of a single length and greater weight of the sample, and the difficulty of the right model and weight of the sample is smaller. When the same feature vector norm, the maximum margin angle.
Method: normalized feature
arcface loss series
L-softmax
Since
softmax equivalent to
two boundary, causes a larger distance between the network-based learning, within-class distance smaller features.
