Glossary
- black-box model
A machine learning model whose internal decision-making process is not accessible or interpretable by users.
- classifier
A model that outputs a predicted label or category based on input features.
- detector
A model that identifies multiple objects within an image by predicting their categories and bounding boxes based on input features.
- explanation
A conceptual or visual interpretation of why a model produced a specific output.
- MC-RISE
Multi-Color RISE — a variant of RISE that incorporates color information when generating saliency maps.
- occlusion
A saliency technique that hides parts of the input data to evaluate changes in model predictions.
- perturbation
A small, intentional change applied to input data (e.g. noise, occlusion) to test the sensitivity of model outputs.
- RISE
Randomized Input Sampling for Explanation — a black-box saliency algorithm that generates saliency maps by sampling randomly masked inputs.
- saliency algorithm
A computational method for estimating which input regions are most influential in a model’s prediction.
- saliency map
A visual representation that highlights input regions (e.g. parts of an image) most relevant to a model’s decision.
- similarity scoring
The process of measuring how alike two inputs are, often used in retrieval, ranking, or tracking tasks.
- superpixels
Groups of adjacent pixels with similar color and texture used as regions in some saliency techniques.
- visualization
The display of data or model behavior (e.g. saliency maps) in a human-interpretable format to aid understanding.
- white-box model
A machine learning model whose internal logic, parameters, and feature importance are transparent to and interpretable by users (e.g. decision trees, linear regression).
- XAI
Explainable Artificial Intelligence — methods and tools that help humans understand, trust, and interpret machine learning outputs.