D-DL4CV-Lec14d-Intermediate_Features_via_Guided_Backpropagation

Intermediate Features via Guided Backpropagation

Goal

Answer “what are the intermediate features looking for?” by calculating gradients from a single neuron activation to each input pixel.

We choose one specific neuron activation (a single scalar value), not an entire filter.

Steps

1. Choose one neuron: Pick one activation value (e.g., Conv5$[channel=17,x=6,y=8]=0.73$)
2. Backpropagate: Calculate each input pixel’s contribution to that activation
3. Visualize: Display gradient information as images

The “Guided” Modification

Problem: Standard backprop creates noisy visualizations due to negative gradients

Solution: During ReLU backpropagation:

• Set negative gradients to 0 (normal ReLU)
• Also set gradients to 0 where forward pass was negative

This double filtering produces cleaner visualizations. Why it works better isn’t fully understood, but results are significantly clearer.