Abstract:

Due to the lack of prior information, the group Lasso model is trained based on the group number parameter that groups the units uniformly, continuously and fixedly, which easily leads to biased estimates about the group structure of variables. We propose a feature clustering adaptive sparse group autoencoder, which uses the feature clustering method to change the grouping of the hidden layer unit in the process of iteration so that it can adaptively change with the convergence of the features, achieving better estimation of group structure of the variables. Experiments show that the model can better capture the relevant information of the group structure of the variables during the training process and improve the image classification performance to a certain extent.

Key words: autoencoder, group Lasso, feature clustering, adaptive