A Deep-Layer Feature Selection Method Based on Deep Neural Networks
Inspired by the sparse mechanism of the biological nervous system, we propose a novel feature selection algorithm: features back-selection (FBS) method, which is based on the deep learning architecture. Compared with the existing feature selection method, this method is no longer a shallow layer approach, since it is from the global perspective, which traces back step by step to the original key feature sites of the raw data by the abstract features learned from the top of the deep neural networks. For MNIST data, the FBS method has quite well performance on searching for the original important pixels of the digit data. It shows that the FBS method not only can determine the relevant features for learning task with keeping a quite high prediction accuracy, but also can reduce the space of data storage as well as the computational complexity.
KeywordsFeatures back-selection Deep neural networks Deep-layer architecture Key sites
This research was supported by NSFC Nos. 11471006 and 11101327, National Science and Technology Cooperation Program of China (No. 2015DFA81780), and the Fundamental Research Funds for the Central Universities (No. xjj2017126).
- Hinton, G.E.: A practical guide to training restricted Boltzmann machines. Momentum 9(1), 599–619 (2010)Google Scholar
- LeCun, Y., Cortes, C.: MNIST Handwritten Digit Database (2010). http://yann.lecun.com/exdb/mnist
- Jiang, B., Ding, C., Luo, B., et al.: Graph-Laplacian PCA: closed-form solution and robustness. CVPR 9(4), 3492–3498 (2013)Google Scholar