Correlation Projection for Analytic Learning of a Classification Network

In this paper, we propose a correlation projection network (CPNet) that determines its parameters analytically for pattern classification. This network consists of multiple modules with each module containing two layers. We first introduce a label encoding process for each module to facilitate a locally supervised learning. Subsequently, in each module, the first layer conducts what we call the correlation projection process for feature extraction. The second layer determines its parameters analytically through solving a least squares problem. By introducing a corresponding label decoding process, the proposed CPNet achieves a multi-exit structure which is the first of its kind in multilayer analytic learning. Due to the analytic learning technique, the proposed method only needs to visit the dataset once, and is hence significantly faster than the commonly used backpropagation, as verified in the experiments. We also conduct classification tasks on various benchmark datasets which demonstrate competitive results compared with several state-of-the-arts.