Boosting Sparsity-Induced Autoencoder: A Novel Sparse Feature Ensemble Learning for Image Classification

摘要

As a model of unsupervised learning, autoencoder is often employed to perform the pre-training of the deep neural networks. However, autoencoder and its variants have not taken the statistical characteristics and the domain knowledge of training set into the design of deep neural networks and have abandoned a lot of features learned from different levels at the pre-training process. In this paper, we propose a novel sparse feature ensemble learning method for natural image classification, named boosting sparsity-induced autoencoder, to fully utilize hierarchical and diverse features. Firstly, a sparsity encourage method is introduced by adding an extra sparsity-induced layer to exploit the representative and intrinsic features of the input. And then, the ensemble learning is taken into consideration of the construction of the model to improve and boost the accuracy and stability of a single model. The classification results on three datasets demonstrate the effectiveness of the proposed method.