Autoencoder with Orthogonal Variant
Álvaro Anzueto-Ríos, Felipe Gomez-Castaneda, Luis Martin Flores-Nava, Jose Antonio Moreno-Cadenas · 2023
Autoencoder is a widely used neural architecture for dimensionality reduction. It can be considered similar to the principal component analysis (PCA) methodology. However, the final distribution of the components between classes does not establish orthogonality between them, which can result in a reduced separation between different classes. To address this issue, we present a modified autoencoder architecture with an orthogonal variant. The error minimization equation has been changed to ensure orthogonality between the final components. Our experiments show that the proposed orthogonal autoencoder architecture generates a final distribution with more separability than the PCA numerical process and a typical autoencoder architecture. This makes it a promising approach for applications that require high separability between different classes.