ZF Net (Visualizing and Understanding Convolutional Networks)

Status: Read

Author: Matthew D. Zeiler, Rob Fergus

Topic: CNNs, CV , Image

Category: Visualization

Conference: ECCV

Year: 2014

Link: https://link.springer.com/chapter/10.1007/978-3-319-10590-1_53

Summary: Visualize CNN Filters / Kernels using De-Convolutions on CNN filter activations.

Questions

What did authors try to accomplish?

Visualize CNN Filters / Kernels.
Trying to get better CNN architecture by understanding the filter activation visualizations.

What were the key elements of the approach?

Use De-Convolutional Neural Networks (DeConvnets) in each layer of CNN after setting all activations to zero other than the one to be visualized.
Understanding what are the key regions that different CNN Kernels look at after training.
Understanding the effects of obscuring different regions of image.

What can you use yourself from this paper?

How to visualize the filters of a trained CNN model.
How to understand what a trained model is looking at to make predictions.

What other references to follow?

N/A