Layer-wise Relevance Propagation

Layer-wise Relevance Propagation (LRP) is an explainable AI technique that attributes a neural network's prediction back to its input features by propagating relevance scores layer by layer from the output toward the input. Unlike gradient-based saliency methods, LRP redistributes the prediction score using conservation rules, producing a heatmap showing which inputs contributed most to a decision. In accessibility-related AI work, LRP is used to validate that models making decisions about EEG signals, medical images, or assistive predictions rely on plausible features, which matters for trust calibration and for catching shortcut learning that could disadvantage disabled users. Alongside Grad-CAM and SHAP, LRP is one of the standard tools used in brain-computer interface research to check the neurophysiological basis of model predictions.