With the development of machine learning, artificial intelligence and other fields, the processing of data mining has become more and more complex. As a data preprocessing step, feature selection is very important in many tasks, like classification, clustering and regression, etc. However, traditional feature selection methods learns similarity matrix from original data to calculate relevant data. What this method learns is the relationships which are linear between data and their labels, and it cannot deal with complex nonlinear data well in real-world applications. In this article, we proposed a feature selection method based on neural network that can select discriminative feature subsets by neural network pruning. And update all weights by gradient descent. Experimental results of our method on several real-world datasets achieve competitive or superior performance compared to three close related feature selection approaches.