Explainable AI (XAI)

Produces a coarse localization heatmap highlighting the regions of an input image that most influenced a CNN's classification decision. For quality inspection, Grad-CAM visually answers: "Which pixels caused the model to flag this component as defective?"

• Uses the gradients of the target class flowing into the final convolutional layer
• Generates a low-resolution heatmap overlaid on the original defect image
• Requires no architectural changes or re-training
• Example: A Grad-CAM overlay reveals the model focused on a micro-crack rather than a benign surface scratch, validating correct behavior

Explains any black-box classifier's individual prediction by perturbing the input and fitting a locally faithful, interpretable surrogate model (e.g., a linear model or decision tree).

• Model-agnostic: works with CNNs, Vision Transformers, or any architecture
• Generates superpixel-based explanations showing which image segments contributed positively or negatively to a defect classification
• Perturbs the input by turning superpixels on/off and observing prediction changes
• Example: LIME identifies that a specific texture patch, not the component's edge geometry, drove a 'scratch' classification

Applies cooperative game theory to assign each input feature an importance value for a particular prediction, ensuring consistent and theoretically grounded attributions.

• Computes Shapley values by averaging marginal contributions across all feature coalitions
• Provides both global (dataset-level) and local (per-image) interpretability
• For vision tasks, partitions images into superpixels and calculates each region's contribution
• Example: SHAP analysis across 10,000 inspected parts reveals that 'edge roughness' features consistently dominate false reject decisions, guiding data augmentation strategy

An axiomatic attribution method that computes feature importance by accumulating gradients along a straight-line path from a baseline (e.g., a black image) to the actual input.

• Satisfies Sensitivity and Implementation Invariance axioms that other methods violate
• Produces fine-grained pixel-level attribution maps
• Particularly effective for identifying the precise boundary pixels that triggered a defect classification
• Example: Integrated Gradients isolates the exact contour of a soldering void, distinguishing it from surrounding thermal discoloration that a human inspector might incorrectly flag

Generates minimally perturbed versions of an input image that would change the model's classification, answering: "What would need to change for this part to pass inspection?"

• Identifies the smallest semantic changes required to flip a 'defect' prediction to 'conforming'
• Uses generative models or optimization in latent space to produce realistic counterfactuals
• Provides actionable guidance for process engineers on acceptable tolerance boundaries
• Example: A counterfactual shows that reducing a surface blemish's contrast by 15% would reclassify the part as acceptable, informing polishing specifications

Synthesizes an idealized input image that maximally activates a specific neuron, channel, or class logit in a trained network, revealing what the model has learned to detect.

• Applies gradient ascent in input space with regularization priors (e.g., total variation, jitter)
• Reveals learned defect prototypes: e.g., what the 'dent' detector neuron is actually looking for
• Critical for debugging spurious correlations: if a 'scratch' neuron activates for lighting gradients, the model has learned a non-robust feature
• Example: Feature visualization exposes that a 'corrosion' classifier is responding to water droplet reflections rather than actual pitting, prompting dataset cleaning

Gradient-weighted Class Activation Mapping generates a coarse localization map highlighting the regions of an input image most influential to a CNN's prediction. It computes the gradient of the target class score with respect to the final convolutional layer's feature maps, producing a heatmap overlay. For defect classification, this reveals whether the model is focusing on the actual defect or spurious background correlations.

SHapley Additive exPlanations is a game-theoretic approach to model interpretability. It assigns each input feature an importance value for a particular prediction, representing its contribution relative to a baseline. In quality inspection, SHAP values can quantify exactly how much a specific pixel region or extracted feature pushed the model toward a 'defect' classification, providing a mathematically rigorous audit trail.

Local Interpretable Model-agnostic Explanations approximates any black-box classifier locally with an interpretable surrogate model. LIME perturbs the input image by segmenting it into superpixels and observes how predictions change, learning which superpixels are most critical. This technique is model-agnostic, meaning it can explain a CNN, Vision Transformer, or any other architecture without internal access.

A saliency map visualizes the gradient of the output class score with respect to each input pixel, indicating which pixels, if changed minimally, would most affect the classification. This provides a raw, fine-grained sensitivity analysis. While computationally efficient, vanilla saliency maps can be noisy; techniques like SmoothGrad reduce visual noise by averaging gradients over multiple noisy copies of the input image.

Concept Bottleneck Models (CBMs) enforce interpretability by design. The architecture first predicts high-level, human-understandable concepts from the input, and then uses only those concept scores to make the final classification. In manufacturing, concepts might include 'scratched surface,' 'discoloration,' or 'misaligned edge.' This allows an operator to intervene by correcting a mispredicted concept directly.

A counterfactual explanation describes the minimal change to an input required to alter a model's prediction to a predefined alternative outcome. For a rejected component, it answers: 'What would need to be different for this part to pass inspection?' This is generated through optimization, finding the closest input that crosses the decision boundary, providing actionable diagnostic feedback for process engineers.