Advanced Deep Learning with Python

Guided backpropagation

Guided backpropagation (Striving for Simplicity: The All Convolutional Net, https://arxiv.org/abs/1412.6806) allows us to visualize the features that are learned by a single unit of one layer of a CNN. The following diagram shows how the algorithm works:

Guided backpropagation visualization; Inspired by https://arxiv.org/abs/1412.6806.

Here is the step-by-step execution:

1. First, we start with a regular CNN (for example, AlexNet, VGG, and so on) with ReLU activations.
2. Then, we feed the network with a single image f⁽⁰⁾ and propagate it forward until we get to the layer, l, we're interested in. This could be any network ...