Documents disponibles écrits par cet auteur

Seek for Robustness of Deep Neural Networks to Label Noise / Géraldin NANFACK

Public ISBD Titre : Seek for Robustness of Deep Neural Networks to Label Noise Type de document : projet fin études Auteurs : Géraldin NANFACK, Auteur Langues : Français (fre) Catégories : SDBD Index. décimale : mast 88/18 Résumé : Learning from data with noisy labels (Label Noise) has been a major concern in supervised Machine Learning these recent years. This problem has become even more worrying in Deep Learning where the generalization capabilities have been questioned lately. Indeed, deep learning requires a large amount of data that is generally collected by search engines which frequently return data with unreliable labels. In this Master thesis, we investigate the Label Noise in Deep Learning by proposing variational inference approaches to learn the probability distribution of the true label by considering them as discrete latent variables. Our contributions are : firstly to the best of our knowledges, this is the first work in Label Noise, where the true labels, expressed in term of discrete latent variable are learnt via variational inference with reparameterization trick while observed labels are learnt discriminatively, secondly, the noise transition matrix is learnt during the training without any particular process, neither heuristic nor preliminary phases. The theoretical results show how true label distribution can be learned by variational inference in any discriminative neural network and implementations done on MNIST or CIFAR32 show very satisfying results robust to overfitting and memorization. Then, this work can be adapted to any discriminative architecture based on neural networks whose optimization would be to involve maximizing the likelihood probabilities of observed labels. Seek for Robustness of Deep Neural Networks to Label Noise [projet fin études] / Géraldin NANFACK, Auteur . - [s.d.]. Langues : Français (fre) Catégories : SDBD Index. décimale : mast 88/18 Résumé : Learning from data with noisy labels (Label Noise) has been a major concern in supervised Machine Learning these recent years. This problem has become even more worrying in Deep Learning where the generalization capabilities have been questioned lately. Indeed, deep learning requires a large amount of data that is generally collected by search engines which frequently return data with unreliable labels. In this Master thesis, we investigate the Label Noise in Deep Learning by proposing variational inference approaches to learn the probability distribution of the true label by considering them as discrete latent variables. Our contributions are : firstly to the best of our knowledges, this is the first work in Label Noise, where the true labels, expressed in term of discrete latent variable are learnt via variational inference with reparameterization trick while observed labels are learnt discriminatively, secondly, the noise transition matrix is learnt during the training without any particular process, neither heuristic nor preliminary phases. The theoretical results show how true label distribution can be learned by variational inference in any discriminative neural network and implementations done on MNIST or CIFAR32 show very satisfying results robust to overfitting and memorization. Then, this work can be adapted to any discriminative architecture based on neural networks whose optimization would be to involve maximizing the likelihood probabilities of observed labels.

Réservation

Réserver ce document

Exemplaires

Code barre	Cote	Support	Localisation	Section	Disponibilité
mast 88/18	mast 88/18 GER	Texte imprimé	Unité des masters	Mast/18	Disponible