Latent Space Factorisation and Manipulation via Matrix Subspace Projection

Xiao Li, Chenghua Lin, Chaozheng Wang, Frank Guerin

Research output: Working paper

Abstract

This paper proposes a novel method for factorising the information in the latent space of an autoencoder (AE), to improve the interpretability of the latent space and facilitate controlled generation. When trained on a dataset with labelled attributes we can produce a latent vector which separates information encoding the attributes from other characteristic information, and also disentangles the attribute information. This then allows us to manipulate each attribute of the latent representation individually without affecting others. Our method, matrix subspace projection, is simpler than the state of the art adversarial network approaches to latent space factorisation. We demonstrate the utility of the method for attribute manipulation tasks on the CelebA image dataset and the E2E text corpus.
Original languageEnglish
PublisherArXiv
Number of pages10
Publication statusSubmitted - 26 Jul 2019

Fingerprint

Factorization

Cite this

Latent Space Factorisation and Manipulation via Matrix Subspace Projection. / Li, Xiao; Lin, Chenghua; Wang, Chaozheng; Guerin, Frank.

ArXiv, 2019.

Research output: Working paper

Li, Xiao ; Lin, Chenghua ; Wang, Chaozheng ; Guerin, Frank. / Latent Space Factorisation and Manipulation via Matrix Subspace Projection. ArXiv, 2019.
@techreport{6da66a06b7684d29a0a5de781147b559,
title = "Latent Space Factorisation and Manipulation via Matrix Subspace Projection",
abstract = "This paper proposes a novel method for factorising the information in the latent space of an autoencoder (AE), to improve the interpretability of the latent space and facilitate controlled generation. When trained on a dataset with labelled attributes we can produce a latent vector which separates information encoding the attributes from other characteristic information, and also disentangles the attribute information. This then allows us to manipulate each attribute of the latent representation individually without affecting others. Our method, matrix subspace projection, is simpler than the state of the art adversarial network approaches to latent space factorisation. We demonstrate the utility of the method for attribute manipulation tasks on the CelebA image dataset and the E2E text corpus.",
author = "Xiao Li and Chenghua Lin and Chaozheng Wang and Frank Guerin",
note = "32nd Conference on Neural Information Processing Systems (NIPS 2018), Montr{\'e}al, Canada",
year = "2019",
month = "7",
day = "26",
language = "English",
publisher = "ArXiv",
type = "WorkingPaper",
institution = "ArXiv",

}

TY - UNPB

T1 - Latent Space Factorisation and Manipulation via Matrix Subspace Projection

AU - Li, Xiao

AU - Lin, Chenghua

AU - Wang, Chaozheng

AU - Guerin, Frank

N1 - 32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada

PY - 2019/7/26

Y1 - 2019/7/26

N2 - This paper proposes a novel method for factorising the information in the latent space of an autoencoder (AE), to improve the interpretability of the latent space and facilitate controlled generation. When trained on a dataset with labelled attributes we can produce a latent vector which separates information encoding the attributes from other characteristic information, and also disentangles the attribute information. This then allows us to manipulate each attribute of the latent representation individually without affecting others. Our method, matrix subspace projection, is simpler than the state of the art adversarial network approaches to latent space factorisation. We demonstrate the utility of the method for attribute manipulation tasks on the CelebA image dataset and the E2E text corpus.

AB - This paper proposes a novel method for factorising the information in the latent space of an autoencoder (AE), to improve the interpretability of the latent space and facilitate controlled generation. When trained on a dataset with labelled attributes we can produce a latent vector which separates information encoding the attributes from other characteristic information, and also disentangles the attribute information. This then allows us to manipulate each attribute of the latent representation individually without affecting others. Our method, matrix subspace projection, is simpler than the state of the art adversarial network approaches to latent space factorisation. We demonstrate the utility of the method for attribute manipulation tasks on the CelebA image dataset and the E2E text corpus.

M3 - Working paper

BT - Latent Space Factorisation and Manipulation via Matrix Subspace Projection

PB - ArXiv

ER -