Visualizing and Comparing Convolutional Neural Networks
This work addresses the limited interpretability of CNNs for researchers and practitioners, but it is incremental as it builds on existing visualization techniques.
The authors tackled the problem of understanding the internal mechanisms of Convolutional Neural Networks (CNNs) by visualizing internal representations and comparing architectures, showing that deeper CNNs offer advantages.
Convolutional Neural Networks (CNNs) have achieved comparable error rates to well-trained human on ILSVRC2014 image classification task. To achieve better performance, the complexity of CNNs is continually increasing with deeper and bigger architectures. Though CNNs achieved promising external classification behavior, understanding of their internal work mechanism is still limited. In this work, we attempt to understand the internal work mechanism of CNNs by probing the internal representations in two comprehensive aspects, i.e., visualizing patches in the representation spaces constructed by different layers, and visualizing visual information kept in each layer. We further compare CNNs with different depths and show the advantages brought by deeper architecture.