CNN Visualization

Team Project

Visualization of popular convolutional neural network (CNN) models.

Figure 1. Visualization of popular CNN architectures with saliency maps.

Convolutional Neural Network (CNN) has been proposed by LeCun et al. in 1989 for digit recognition. Recent advancement in hardware, datasets, and novel network architectures have further enhanced the capabilities of CNN. Applications ranging from semantic segmentation, object recognition, image processing, and medical imaging are all benefited from the development of CNN. However, little is known for the inner mechanism of how CNN works, including its mathematical foundation as well as the roles of each interior layer.

In this project, we aim to review the latest visualization methods for CNN and apply them to popular CNN architectures, including AlexNet, VGG, GoogLeNet, and ResNet. We adopt five visualization methods, including activity visualization, Deconvolutional Network (DeConvNet), Saliency Map, Deep Generator Network (DGN), and Plug and Play Generative Networks (PPGN).

For more information, please refer to our poster (in English) or technical report (in Chinese). Our implementation code is also publically available at GitHub.