Work

Régularisation entropique des vecteurs de caractéristiques d'un réseau de neurones pour une meilleur transférabilité

Year: 2023
Type: conference-paper
Abstract: This paper discusses the problem of classification when learning on an approximation of the labels. For example, aregression problem can be discretized to obtain a classification problem that is easier to solve. One problem encountered whentraining with the cross-entropy is the overfitting of the features that occurs with coarse labeling. To address this problem, weintroduce an entropic regularization applied to the feature space to increase features’ diversity. These features can be later used tosolve a refined task with a smaller error. The validity of our method is supported by both theoretical analysis and experiments
Source: HAL
Language: French
Open Access status: green
OpenAlex ID: W4367060574
FWCI:
Cites: 14
Cited by:
Related to: 9