Convolutional Visual Feature Learning: A Compositional Subspace Representation Perspective

The main contribution of this paper is to provide a new perspective to understand the end-to-end convolutional visual feature learning in a convolutional neural network (ConvNet) using empirical feature map analysis. The analysis is performed through a novel mathod called compositional subspace model using a minimal ConvNet. This method allows us to better understand how a ConvNet learn visual features in a hierarchical manner. A handwritten digit recognition using MNIST dataset is used to experiment the empirical feature map analysis. The experimental results conclude our proposal on using the compositional subspace model to visually understand the convolutional visual feature learning in a ConvNet.