Learning optimally separated class-specific subspace representations using convolutional autoencoder

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