We introduce a parametric family of entropy regularizers, which includes label smoothing as a special case, and use it to gain a better understanding of the ...

(ii) Label smoothing assigns infinite cost to sparse output distributions, which may be an undesir- able behavior for language generation tasks. (iii) There is ...

May 2, 2020 · We introduce a parametric family of entropy regularizers, which includes label smoothing as a special case, and use it to gain a better understanding of the ...

(i) With tuning of the regularizer's coefficient, any choice of α can yield similar perfor- mance, i.e. there is nothing special about label smoothing. In fact, ...

A parametric family of entropy regularizers is introduced, which includes label smoothing as a special case, and is used to gain a better understanding of ...

Examples of other label encodings include the label-smoothing regularization (LSR), generalized entropy regularization (GER), and so on [12,26]. In LSR, Szegedy ...

Title: Generalized Entropy Regularization: or There's Nothing Special about Label Smoothing. Authors: Clara Meister, Elizabeth Salezky, and Ryan Cotterell.

Generalized Entropy Regularization or: There's Nothing Special about Label Smoothing ... label smoothing is one, has a connection to entropy regularization.

... label smoothingis one, has a connection to entropy regularization ... Generalized Entropy Regularization or: There's Nothing Special about Label Smoothing.

Generalized entropy regularization can be used with any probabilistic model and data set. Just set the --criterion flag to jensen_cross_entropy and specify -- ...