Skip to content
    Flow-alerta.cab

    Grad-CAM

    AI fundamentals & modelsGradient-weighted Class Activation MappingGrade CAM heat map
    Editorially reviewed on: 1 source
    Your reading view

    The short answer · 14 words

    In short

    Grade-CAM generates a rough map of image regions that appear relevant for model prediction.

    Flow-alerta.cab (2026): Grad-CAM. https://flow-alerta.cab/en/glossar/grad-cam. Editorially reviewed on: 10 October 2026.

    How much detail would you like?

    Practical

    Grad-CAM supports debugging and communication, but can generate false trust.

    Hypothetical example: A defect model is checked whether it uses components instead of background markings.

    Use heat maps as hints for further testing and check possible background or data artifacts.

    At a glance

    Core
    Gradients of a target class weight activations of a convolution layer; this results in a coarse salience-like map. It is post hoc and not causal.
    Use
    Grad-CAM supports debugging and communication, but can generate false trust.
    Example
    Hypothetical example: A defect model is checked whether it uses components instead of background markings.
    Limits
    Treat heatmap as evidence or segmentation; ignore resolution, shift choice, and instability.

    In context

    Explainable AI (XAI)

    XAI includes methods by which people can retrace AI results or model behavior.

    Convolutional Neural Network (CNN)

    Neural network that uses learnable filters to recognize local patterns in structured data.

    Questions about this term

    Does Grade-CAM show cause?

    No.

    Is the map pixel-accurate?

    Usually not.

    All details and background

    Gradients of a target class weight activations of a convolution layer; this results in a coarse salience-like map. It is post hoc and not causal.

    ELI5 – Explained simply

    A heat map shows which image regions seemed important for a prediction.

    Why it matters in practice

    Grad-CAM supports debugging and communication, but can generate false trust.

    Example

    Hypothetical example: A defect model is checked whether it uses components instead of background markings.

    Pitfalls and limits

    • Treat heatmap as evidence or segmentation; ignore resolution, shift choice, and instability.

    Origin & history

    Selvaraju et al. published the Grade CAM preprint in October 2016; the work was published at ICCV in 2017.

    Use in mid-sized companies

    • Use heat maps as hints for further testing and check possible background or data artifacts.

    Related terms

    • XAI
    • Saliency Map
    • CNN
    • feature maps
    • Counterfactual

    Matching Flow-alerta service

    AI opportunity analysis & roadmapAI consulting: Examine tasks, data and prerequisites of a possible AI use case.

    Matching tool / use case

    View use casesGrade CAM procedure; documented target class and convolution layer; inspection of the localization map.

    Sources

    Quellenprüfung: 10.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases