Single-layer Unsupervised Feature Learning with l2 regularized sparse filtering

Abstract

Patch-based Single-layer Unsupervised Feature Learning (SUFL) has been successfully applied in several tasks of computer vision. In the feature learning process, the key ingredient is how to learn a good feature mapping that connects patches to feature vectors. Among various feature mapping methods, the sparse filtering is easy to be implemented and hyper-parameter free. However, the standard sparse filtering method only considers the sparsity distribution of the learned features, ignoring the feature mapping matrix itself. This will lead to a random magnitude for mapping matrix and further weaken the generation performance. In this paper we proposed L2 regularized sparse filtering for the feature mapping in SULF. Classification experiments on three different datasets, i.e., CIFAR-10, small Norb, and subsets of CISIA-HWDB1.0 handwritten characters, show that our method has better performance comparing with the standard sparse filtering.