Consecutive Dimensionality Reduction by Canonical Correlation Analysis for Visualization of Convolutional Neural Networks

抄録

In this paper, we develop a visualization tool suitable for deep neural networks (DNN). Although typical dimensionality reduction methods, such as principal component analysis (PCA), are useful to visualize highdimensional data as 2 or 3 dimensional representations, most of those methods focus their attention on how to create essential subspaces based only on a given unique feature representation. On the other hand, DNN naturally have consecutive multiple feature representations corresponding to their intermediate layers. In order to understand relationships of those consecutive intermediate layers, we utilize canonical correlation analysis (CCA) to visualize them in a unified subspace. Our method (called consecutive CCA) can visualize “feature flow” which represents movement of samples between two consecutive layers of DNN. By using standard benchmark datasets, we show that our visualization results contain much information that typical visualization methods (such as PCA) do not represent.