IntroVAC: Introspective Variational Classifiers for learning interpretable latent subspaces

• Published in: Engineering applications of artificial intelligence Vol. 109; p. 104658
• Main Authors: Maggipinto, Marco, Terzi, Matteo, Susto, Gian Antonio
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2022
• Subjects: Attribute manipulation; Autoencoders; Deep learning; Interpretability; Deep learning; Attribute manipulation; Autoencoders; Interpretability
• ISSN: 0952-1976; 1873-6769
• DOI: 10.1016/j.engappai.2021.104658

Learning useful representations of complex data has been the subject of extensive research for many years. In particular, with the diffusion of complex Deep Learning-based approaches in engineering applications, the possibility to interpret, to a certain degree, model predictions is of fundamental importance for both the model users and developers. In the context of Deep Neural Networks, Variational Autoencoders have gained lots of attention since they provide an explicit model of the data distribution based on an encoder/decoder architecture which is able to both generate images and encode them in a low-dimensional subspace. However, the latent space is not easily interpretable and the generation capabilities show some limitations since images typically look blurry and lack details. In this paper, we propose the Introspective Variational Classifier (IntroVAC), a model that learns interpretable latent subspaces by exploiting information from an additional label and provides improved image quality thanks to an adversarial training strategy. We show that IntroVAC is able to learn meaningful directions in the latent space enabling fine-grained manipulation of image attributes. We validated our approach on the CelebA dataset. When compared with standard Variational Autoencoder Classifiers, the proposed approach outperform them by achieving a Frechét Inception Distance of 25.5 versus a value of 63.9.