Learning optimally separated class-specific subspace representations using convolutional autoencoder

TL;DR

This paper introduces a convolutional autoencoder architecture with a novel class-specific self expressiveness layer to generate well-separated class-specific subspace features, improving classification performance on various datasets.

Contribution

The paper proposes a new autoencoder-based method with a class-specific self expressiveness layer to enhance class separation in subspace representations.

Findings

Significant improvement in classification accuracy over existing methods.

Effective separation of class-specific subspaces in feature space.

Robustness to noisy and overlapping subspaces.

Abstract

In this work, we propose a novel convolutional autoencoder based architecture to generate subspace specific feature representations that are best suited for classification task. The class-specific data is assumed to lie in low dimensional linear subspaces, which could be noisy and not well separated, i.e., subspace distance (principal angle) between two classes is very low. The proposed network uses a novel class-specific self expressiveness (CSSE) layer sandwiched between encoder and decoder networks to generate class-wise subspace representations which are well separated. The CSSE layer along with encoder/ decoder are trained in such a way that data still lies in subspaces in the feature space with minimum principal angle much higher than that of the input space. To demonstrate the effectiveness of the proposed approach, several experiments have been carried out on state-of-the-art machine learning datasets and a significant improvement in classification performance is observed over existing subspace based transformation learning methods.