Guided Backpropagation

Standard backpropagation passes gradients through a ReLU if the input was positive. Guided Backpropagation adds a second gate: it only passes the gradient if the gradient itself is also positive. This combines the forward activation mask with a backward gradient mask.

• Standard ReLU Backward: grad_output * (input > 0)
• Guided ReLU Backward: grad_output * (input > 0) * (grad_output > 0)
• Effect: This strictly prevents negative gradients from flowing backward, eliminating neurons that act as 'inhibitors' of the target class from the visualization.

By zeroing out negative gradients, Guided Backpropagation reconstructs the input features that only have a purely positive influence on the activation of the target neuron. This acts as a strong prior that natural images are composed of features that positively excite higher-level detectors.

• Noise Reduction: Suppresses the grainy, high-frequency noise common in standard saliency maps.
• Object Outlines: Tends to produce crisper outlines of objects rather than filling in solid regions.
• Comparison to DeconvNet: Structurally similar to the Deconvolutional Network approach but applies the gating logic directly within the standard ReLU backward hook rather than a separate deconvolutional architecture.

In frameworks like PyTorch, Guided Backpropagation is implemented by registering a custom backward hook on every ReLU layer in the model. The hook intercepts the gradient tensor during the backward pass and clamps its negative values to zero.

pythonCopy
def relu_hook(module, grad_input, grad_output):
    # grad_output[0] is the gradient from the next layer
    modified_grad = torch.clamp(grad_output[0], min=0.0)
    return (modified_grad,)

• Model Modification: The technique requires temporarily modifying the model's computation graph.
• Inference Only: It is applied post-hoc to a trained model; no retraining is required.

Guided Backpropagation suffers from a significant flaw: it can visualize features that the model is not actually sensitive to. Because it stops negative gradients, it can fail to detect when a small change in an input feature would destroy the activation.

• Saturation Problem: If a neuron is operating in a saturated region where increasing the input no longer increases the output, the gradient is zero. Guided Backpropagation cannot highlight this feature even if it is critical.
• Violation of Sensitivity: This causes the method to violate the Sensitivity axiom, which states that if one input differs from a baseline in one feature and produces a different prediction, that feature must receive a non-zero attribution.

Guided Backpropagation is a pre-axiomatic attribution method. It predates the rigorous mathematical frameworks established by Integrated Gradients and DeepLIFT.

• vs. Integrated Gradients: Guided Backpropagation does not satisfy Completeness (the sum of attributions does not equal the output difference from a baseline).
• vs. Grad-CAM: Guided Backpropagation produces high-resolution, fine-grained maps, while Grad-CAM produces coarse, class-discriminative localization maps. The two are often combined as Guided Grad-CAM by point-wise multiplying the high-resolution Guided Backpropagation map with the upsampled Grad-CAM heatmap.

Despite its theoretical limitations, Guided Backpropagation remains a popular tool for qualitative debugging of convolutional neural networks during the development phase.

• Neuron Inspection: Helps researchers visually inspect what specific high-level neurons in a trained network have learned to detect.
• Failure Mode Analysis: Can quickly reveal if a model is focusing on spurious background correlations rather than the primary object.
• Generative Model Analysis: Applied to generators in GANs to understand the internal feature synthesis process.