Sparse Feature Extraction Model with Independent Subspace Analysis

摘要

Recent advances in deep learning models have demonstrated remarkable accuracy in object classification. However, the limitations of Convolutional Neural Networks such as the requirement for a large collection of labeled data for training and supervised learning process has called for enhanced feature representation and for unsupervised models. In this paper we propose a novel unsupervised sparsity-based model using Independent Subspace Analysis (ISA) to implement a hierarchical network for feature extraction. The results of our empirical evaluation demonstrates an improved classification accuracy when max pooling is paired with square pooling within each layer. In addition to accuracy, we further show that it also reduces the data dimensions within the layers outperforming known sparsity-based models.