Lesson 3 Notes
Visualizing Convolutional Layers
Let's begin by discussing some of the insight from Matthew Zeiler's Visualizing and Understanding Convolutional Networks and related works.
In Zeiler's paper, we are able to visualize the sorts of image features each filter in a Convolutional layer detects, such as edges, gradients, corners, etc. The Deep Visualization Toolbox is a great tool that allows us to play around with different images, and to see what image features activate certain filters. We can even see what imagenet images activate certain filters. We recommend playing around with this toolbox, and note the increasing complexity of the filters in deeper layers.
What exactly is a convolution?
Let's revisit Convolutions to solidify our understanding of what they are. Recall from Lesson 0 the MNIST dataset, a collection of 28x28 grey-scale images of handwritten digits with known labels.
MNIST DIGIT IMAGE