Grad-CAM (Gradient-weighted Class Activation Mapping)

Uses the gradients of a target concept flowing into the final convolutional layer to produce a coarse localization map. Unlike Class Activation Mapping (CAM), Grad-CAM requires no architectural changes or re-training. It computes the gradient of the score for class c with respect to feature map activations A^k of a convolutional layer, then globally average-pools these gradients to obtain neuron importance weights α^c_k. The final heatmap is a weighted combination of forward activation maps followed by a ReLU to highlight only features with a positive influence on the class of interest.

Applicable to any CNN-based architecture without modification, including:

• Fully-convolutional networks
• Networks with attention mechanisms
• VGG, ResNet, and Inception architectures
• Models used for image-based fraud detection, such as analyzing altered checks, forged signatures, or tampered identity documents This flexibility allows compliance teams to audit a wide variety of vision models already deployed in production.

Standard Grad-CAM produces low-resolution heatmaps due to the spatial downsampling in final convolutional layers. Guided Grad-CAM fuses Grad-CAM with Guided Backpropagation to create high-resolution, class-discriminative visualizations. Guided Backpropagation zeroes out negative gradients during the backward pass, capturing fine-grained pixel-level detail. The pointwise multiplication of the upsampled Grad-CAM heatmap with the Guided Backpropagation output yields a sharp saliency map that precisely highlights the pixels driving a fraud classification decision.

Grad-CAM can be extended to generate counterfactual visual explanations by negating the gradient of the target class score. This highlights the regions that, if modified, would decrease the model's confidence in its prediction. For fraud auditing, this answers the question: 'What would need to change in this document image for the model to classify it as legitimate?' This capability is critical for generating adverse action reason codes in regulated financial environments.

Grad-CAM can localize multiple instances of the same class within a single image without bounding box annotations. By thresholding the heatmap and applying connected-component analysis, distinct regions of interest are segmented. In fraud detection, this enables:

• Identifying multiple altered fields on a single scanned form
• Detecting several forged signatures across a multi-page document
• Pinpointing all suspicious regions in a composite identity document image

When combined with multimodal models, Grad-CAM heatmaps serve as visual grounding for generating natural language explanations. The highlighted regions are encoded alongside the original image and fed into a language model to produce human-readable justifications like: 'The model flagged this transaction because the signature in the lower-right quadrant exhibits inconsistent stroke pressure patterns.' This bridges the gap between saliency maps and the reason codes required by regulators for adverse action notifications.

A foundational visualization technique that highlights input features—typically pixels—most influential to a neural network's classification. Computed by taking the gradient of the target class score with respect to the input image.

• Produces fine-grained, noisy importance maps
• No architectural constraints; works on any differentiable model
• Often serves as the baseline against which Grad-CAM is compared
• Key limitation: lacks class-discriminative localization precision

A variant of standard backpropagation that modifies the gradient flow through ReLU non-linearities, zeroing out negative gradients to produce sharper, less noisy visualizations than raw saliency maps.

• Combines with Grad-CAM to create Guided Grad-CAM
• Prevents backward flow of negative gradients, suppressing irrelevant features
• Produces high-resolution, fine-grained importance maps
• Does not provide class-specific localization on its own

A technique that decomposes a neural network's output prediction and redistributes relevance scores backward through layers using local conservation rules until reaching the input.

• Maintains a relevance conservation property: total relevance is preserved across layers
• Produces pixel-level explanations without relying on gradients
• Particularly effective for models with complex, non-linear architectures
• Widely used in domains requiring strict auditability, such as financial fraud imaging

An interpretability method that explains neural network decisions in terms of human-friendly, high-level concepts rather than low-level pixels. Measures how sensitive predictions are to learned concept vectors.

• Uses user-defined example sets to define concepts like 'signature forgery' or 'altered document'
• Produces a quantitative TCAV score indicating concept importance
• Model-agnostic within the class of neural networks
• Bridges the gap between pixel-level heatmaps and semantic understanding

A hybrid visualization that combines the class-discriminative localization of Grad-CAM with the fine-grained resolution of Guided Backpropagation through element-wise multiplication.

• Produces high-resolution, class-specific heatmaps
• Retains the 'what' (class) and 'where' (location) properties
• Standard approach for detailed model debugging in image-based fraud detection
• Commonly used to explain why a specific document region triggered a fraud alert

A model-specific attribution method that computes feature contributions by integrating gradients along a path from a baseline input (e.g., a blank image) to the actual input.

• Satisfies the completeness axiom: attributions sum to the prediction difference
• Provides pixel-level importance scores with theoretical guarantees
• Baseline choice critically affects explanation quality
• Often used alongside Grad-CAM for complementary global and local explanations