A Deep Generative Model for Semi-Supervised Classification with Noisy Labels

Maxime Langevin; Edouard Mehlman; Jeffrey Regier; Romain Lopez; Michael I Jordan; Nir Yosef

doi:10.48550/arxiv.1809.05957

A Deep Generative Model for Semi-Supervised Classification with Noisy Labels

Maxime Langevin
, Edouard Mehlman
, Jeffrey Regier
, Romain Lopez
, Michael I Jordan
, Nir Yosef

Research output: Contribution to journal › Article

Abstract

Class labels are often imperfectly observed, due to mistakes and to genuine ambiguity among classes. We propose a new semi-supervised deep generative model that explicitly models noisy labels, called the Mislabeled VAE (M-VAE). The M-VAE can perform better than existing deep generative models which do not account for label noise. Additionally, the derivation of M-VAE gives new theoretical insights into the popular M1+M2 semi-supervised model.

Original language	English GB
Number of pages	3
Journal	arxiv.org
DOIs	https://doi.org/10.48550/arxiv.1809.05957
State	In preparation - 16 Sep 2018
Externally published	Yes

Access to Document

10.48550/arxiv.1809.05957

https://arxiv.org/abs/1809.05957License: Other

Cite this

@article{cb891c185dc043febac91f491e164f14,

title = "A Deep Generative Model for Semi-Supervised Classification with Noisy Labels",

abstract = "Class labels are often imperfectly observed, due to mistakes and to genuine ambiguity among classes. We propose a new semi-supervised deep generative model that explicitly models noisy labels, called the Mislabeled VAE (M-VAE). The M-VAE can perform better than existing deep generative models which do not account for label noise. Additionally, the derivation of M-VAE gives new theoretical insights into the popular M1+M2 semi-supervised model. ",

author = "Maxime Langevin and Edouard Mehlman and Jeffrey Regier and Romain Lopez and Jordan, \{Michael I\} and Nir Yosef",

year = "2018",

month = sep,

day = "16",

doi = "10.48550/arxiv.1809.05957",

language = "الإنجليزيّة",

}

TY - JOUR

T1 - A Deep Generative Model for Semi-Supervised Classification with Noisy Labels

AU - Langevin, Maxime

AU - Mehlman, Edouard

AU - Regier, Jeffrey

AU - Lopez, Romain

AU - Jordan, Michael I

AU - Yosef, Nir

PY - 2018/9/16

Y1 - 2018/9/16

N2 - Class labels are often imperfectly observed, due to mistakes and to genuine ambiguity among classes. We propose a new semi-supervised deep generative model that explicitly models noisy labels, called the Mislabeled VAE (M-VAE). The M-VAE can perform better than existing deep generative models which do not account for label noise. Additionally, the derivation of M-VAE gives new theoretical insights into the popular M1+M2 semi-supervised model.

AB - Class labels are often imperfectly observed, due to mistakes and to genuine ambiguity among classes. We propose a new semi-supervised deep generative model that explicitly models noisy labels, called the Mislabeled VAE (M-VAE). The M-VAE can perform better than existing deep generative models which do not account for label noise. Additionally, the derivation of M-VAE gives new theoretical insights into the popular M1+M2 semi-supervised model.

U2 - 10.48550/arxiv.1809.05957

DO - 10.48550/arxiv.1809.05957

M3 - مقالة

SN - 2331-8422

JO - arxiv.org

JF - arxiv.org

ER -

A Deep Generative Model for Semi-Supervised Classification with Noisy Labels

Abstract

Access to Document

Fingerprint

Cite this