Enhanced Deep Learning with Improved Feature Subspace Separation

Abstract

This research proposes a new deep convolutional network architecture that improves the feature subspace separation. In training, the system considers M classes of input sets $\{\mathcal{C}_i \}_{i = 1}^M$ and M deep convolutional networks $\{DN_i \}_{i = 1}^M$ whose filter and other parameters are randomly initialized. For each input class $\mathcal{C}_i$, Convolutional Neural Network generates a set of features ℱ i . Then, a local subspace S i is matched for each set ℱ i . This is followed with a full training of the deep convolutional network DN i based on a decision criteria developed with computation of rejections of all features in $\{\mathcal{F}_i \}_{i = 1}^M$ to S i . Five different deep convolutional network topologies are used to show that the proposed technique works better for small network topologies and has comparable performance to more complex networks.