Adaptive structured noise injection for shallow and deep neural networks

Abstract : Dropout is a regularisation technique in neural network training where unit activations are randomly set to zero with a given probability independently. In this work, we propose a generalisation of dropout and other multiplicative noise injection schemes for shallow and deep neural networks, where the random noise applied to different units is not independent but follows a joint distribution that is either fixed or estimated during training. We provide theoretical insights on why such adaptive structured noise injection (ASNI) may be relevant, and empirically confirm that it helps boost the accuracy of simple feedforward and convolutional neural networks, disentangles the hidden layer representations, and leads to sparser representations. Our proposed method is a straightforward modification of the classical dropout and does not require additional computational overhead.
Type de document :
Pré-publication, Document de travail
Liste complète des métadonnées

Littérature citée [35 références]  Voir  Masquer  Télécharger
Contributeur : Jean-Philippe Vert <>
Soumis le : mercredi 20 février 2019 - 01:14:51
Dernière modification le : dimanche 24 février 2019 - 01:28:52
Document(s) archivé(s) le : mardi 21 mai 2019 - 12:22:53


Fichiers produits par l'(les) auteur(s)


  • HAL Id : hal-02025929, version 1


Beyrem Khalfaoui, Joseph Boyd, Jean-Philippe Vert. Adaptive structured noise injection for shallow and deep neural networks. 2019. ⟨hal-02025929⟩



Consultations de la notice


Téléchargements de fichiers