guided backpropagation

**Guided Backpropagation** is a **visualization technique that modifies the standard backpropagation to produce sharper, more interpretable saliency maps** — by additionally masking out negative gradients at ReLU layers during the backward pass, keeping only features that both activated the neuron and had positive gradient. **How Guided Backpropagation Works** - **Standard Backprop**: Passes gradients through ReLU if the input was positive (forward mask). - **Deconvolution**: Passes gradients through ReLU if the gradient is positive (backward mask). - **Guided Backprop**: Applies BOTH masks — gradient passes only if both input AND gradient are positive. - **Result**: Highlights fine-grained input features that positively contribute to the activation of higher layers. **Why It Matters** - **Sharp Maps**: Produces much sharper, more visually detailed saliency maps than vanilla gradients. - **Feature-Level**: Shows individual edges, textures, and patterns rather than blurry activation regions. - **Limitation**: Not class-discriminative — guided Grad-CAM combines it with Grad-CAM for class-specific, high-resolution maps. **Guided Backpropagation** is **the double-filtered gradient** — keeping only the positive signals in both forward and backward passes for crisp saliency maps.