Guided Backpropagation: Interpreting CNN Features

• Guided Backpropagation is a backpropagation-based interpretability technique that uses double ReLU gating to highlight positive activations and produce edge-focused saliency maps.
• It restricts gradient flow, filtering out noise by allowing only gradients associated with positive forward activations to pass, thus enhancing visual clarity.
• Extensions such as Feature-Guided Gradient Backpropagation improve class-specific discrimination in applications like face verification while preserving high spatial fidelity.

Guided Backpropagation is a backpropagation-based interpretability technique that modifies the standard gradient computation in deep networks to produce visually sharp, high-resolution saliency maps. It operates by restricting the backward flow of gradients through rectified linear unit (ReLU) activations using an additional "backward ReLU" gating; this yields saliency maps characterized by clear edge-like structures in input space. While Guided Backpropagation originated with Springenberg et al. (2015) as a tool for interpreting convolutional neural networks (CNNs), subsequent theoretical and empirical analysis has revealed that its output—contrary to widespread belief—primarily reflects human-perceptible input structures rather than class-specific discriminative information. Modern variants and applications, such as Feature-Guided Gradient Backpropagation (FGGB) for face verification, offer improved fidelity and discriminative power by combining channel-specific gradient analysis and aggregation strategies.

1. Mathematical Principles and Backward Signal Flow

In a standard ReLU network, the forward pass produces activations $o = \sigma(y) = \max(y, 0)$. During backpropagation, the gradient from upper layers, denoted $g = \frac{\partial f}{\partial o}$, is passed through the same ReLU gate so that the backpropagated gradient with respect to the pre-activation $y$ is $g_+ = g \cdot \mathbb{1}(o > 0)$. Guided Backpropagation (GBP) introduces a second gate in the backward direction: $$g' = g \cdot \mathbb{1}(g > 0) \cdot \mathbb{1}(o > 0)$$, allowing only those gradient components that are both positive and originated from positive forward activations to propagate further. This additional gating mechanism, when propagated through the entire network, substantially attenuates noisy or negative activations and biases the resulting saliency map toward visually interpretable changes that correlate with presence of image edges and texture (Nie et al., 2018).

2. Theoretical Explanation and Behavior in Random Networks

Nie et al. (2018) provide a rigorous theoretical analysis of Guided Backpropagation, characterizing its behavior via a detailed model of a random three-layer CNN. The derived formula shows that the GBP saliency map $s_k^{\mathrm{GBP}}(x)$ under large numbers of filters converges, up to normalization, to the input image $x$ itself:

$s_k^{\mathrm{GBP}}(x) \approx x$

This result is established using the law of large numbers on the sum over random filters and input patches. Theoretical analysis attributes this image recovery property to the combination of local connectivity in convolutional architectures and the double (forward and backward) ReLU gating. The result contrasts sharply with standard saliency maps and DeconvNet (under random weights), which yield zero-mean, isotropic Gaussian noise rather than structured image recovery. Max-pooling introduces a similar image recovery property in DeconvNet by enforcing additional selection over activation locations (Nie et al., 2018).

3. Comparative Experimental Observations

Empirical visualizations confirm that, for a random CNN, only Guided Backpropagation recovers the input image, whereas other methods produce noise. In trained networks (e.g., VGG-16 with max-pooling), GBP recovers edges and textures that correspond to features learned by the earliest convolutional layers. In practical experiments:

• Saliency maps change substantially across class logits, while GBP and DeconvNet maps remain nearly unchanged with different class targets, indicating that GBP lacks class sensitivity.
• Under adversarial attacks, saliency maps shift in accordance with new class predictions, but GBP and DeconvNet outputs are virtually identical for clean and adversarial images.
• Localization of edge and texture information in GBP is attributable to local receptive field topology; in fully connected networks, strong image recovery via GBP would require exponentially large hidden dimensions, rendering GBP ineffective for non-convolutional structural priors.
• Partial reloading of pretrained weights demonstrates that early convolutional layers in trained CNNs filter out background, but the overall image recovery character of GBP is not significantly altered by changes in higher layers (Nie et al., 2018).

4. Limitations and Interpretability Considerations

The critical theoretical and experimental insights from (Nie et al., 2018) undercut the assumption that Guided Backpropagation genuinely attributes model decisions to class-discriminative input evidence. Instead, GBP saliency maps are better interpreted as visualizations of image structure preserved through convolutional layers, filtered to emphasize edges and textures. These maps lack significant class sensitivity—the same map is recovered regardless of which output logit is targeted. From an interpretability standpoint, GBP does not visualize the "evidence" for the model's specific class prediction, but rather provides a diagnostic of layerwise feature selectivity, especially the types of edge patterns the early network layers detect and preserve.

5. Extensions: Feature-Guided Gradient Backpropagation in Face Verification

Feature-Guided Gradient Backpropagation (FGGB) adapts the gradient backpropagation paradigm for face verification tasks by decomposing input saliency generation into feature-channel-level analyses. Let a verification network $f$ map input images $(I_A, I_B)$ to embeddings $g = \frac{\partial f}{\partial o}$0, using cosine similarity to yield an "Accept" or "Reject" decision. Rather than backpropagating from the final similarity score, FGGB:

1. Computes per-channel gradients $g = \frac{\partial f}{\partial o}$1,
2. Normalizes gradients by Frobenius norm and element-wise absolute value,
3. Assigns channel-wise weights derived from weighted cosine similarity,
4. Aggregates weighted, normalized maps into a single input-domain saliency map,
5. Decomposes the result into positive (similarity) and negative (dissimilarity) contributions depending on the decision boundary.

FGGB maintains spatial fidelity and resolves noisy gradient issues noted with standard GBP. Quantitative evaluations on LFW, CPLFW, and CALFW datasets show that FGGB achieves lowest Deletion (compactness) and among the highest Insertion (recovery) scores for both similarity ("Accept") and dissimilarity ("Reject") maps, outperforming or rivaling gradient- and perturbation-based explainability baselines while maintaining orders of magnitude higher runtime efficiency (Lu et al., 2024).

6. Relationship to Other Backpropagation-Based Visualization Methods

The landscape of backpropagation-based interpretability methods includes:

• Standard Saliency Maps: Backpropagate the final class score, provide noisy, less interpretable maps.
• DeconvNet: Modifies only the backward pass through ReLU, does not enforce forward-gate, recovers input only under certain pooling structures.
• Guided Backpropagation: Uses double (forward and backward) ReLU gates, filtering for positive gradients associated with positive activations, yielding sharp input-edge reconstructions but not class-conditional evidence.
• Grad-CAM/Grad-CAM++: Pool gradients globally, reweight convolutional feature maps, yield coarser, class-sensitive but less spatially localized maps.
• Perturbation-Based Methods (e.g., RISE, CorrRISE, MinusPlus): Rely on masking and forward passes to estimate pixel importance, achieving class sensitivity but incurring major computational cost.

Feature-Guided Gradient Backpropagation operates at feature-channel granularity and leverages the cosine similarity-driven task structure, stabilizing and localizing saliency without sacrificing inference efficiency (Lu et al., 2024).

7. Summary Table: Properties of Backpropagation-Based Saliency Methods

Method	Class Sensitivity	Spatial Detail	Computational Cost
Saliency Map	High	Noisy/coarse	Low (1 backward pass)
DeconvNet	Low (random net)	Noisy/varies	Low
Guided Backprop	Low (random net); Edge-like/clean	Low
Grad-CAM	High	Coarse	Low (1 backward pass)
Perturbation-based	High	Varies	High (many forward passes)
FGGB	High (for verification)	Fine	Low (channel-wise backprop)

This table summarizes the empirical and theoretical properties of central methods discussed in (Nie et al., 2018) and (Lu et al., 2024).

8. Implications and Contemporary Validation

Contemporary research validates the central findings that the practical value of GBP lies primarily in providing visual insight into the types of structures encoded by early convolutional layers, but not in isolating class-specific or decision-critical evidence. Modifications such as FGGB, which adapt the signal aggregation and weighting process to the task and embedding structure, can recover selective discriminative power in verification settings without compromising computational efficiency or spatial precision (Lu et al., 2024). A plausible implication is that future interpretability research should distinguish between input-reconstructive saliency (which reveals architectural biases and learned features) and truly decision-relevant attribution methods which align more closely with the model's output variation.