Skip to content
    Flow-alerta.cab

    Dropout

    AI fundamentals & modelsDropout regularization
    Editorially reviewed on: 2 sources
    Your reading view

    The short answer · 14 words

    In short

    Regularization method for neural networks, in which randomly selected neurons are deactivated during training.

    Flow-alerta.cab (2026): Dropout. https://flow-alerta.cab/en/glossar/dropout. Editorially reviewed on: 09 October 2026.

    How much detail would you like?

    Practical

    Relevant as a regularization technique; test effect based on independent validation data.

    Possible example: A team compares a model with and without dropout on withheld data. Robustness against altered lighting is tested separately.

    A possible regularization in the training of own models; not every architecture or task benefits equally.

    At a glance

    Core
    During training, randomly selected activations are set to zero. This can reduce overfitting; the benefits depend on the model and settings. Many implementations are already scaling in training, so that dropout does not make any further changes in the normal evaluation mode.
    Use
    Relevant as a regularization technique; test effect based on independent validation data.
    Example
    Possible example: A team compares a model with and without dropout on withheld data. Robustness against altered lighting is tested separately.
    Limits
    Leave dropout active during inference (except specifically for Monte-Carlo dropout for uncertainty estimation); too high dropout rate prevents learning.

    In context

    Backpropagation

    Efficient method for calculating gradients in multilayer neural networks.

    Deep Learning

    Deep learning learns multi-level representations with neural networks.

    Questions about this term

    Is Dropout Active in Prediction?

    Usually not. Monte-Carlo Dropout is a targeted exception for investigating uncertainty.

    When is scaled?

    It depends on the implementation. PyTorch uses scaling in training; in evaluation mode, the usual dropout layer acts as an identity.

    All details and background

    During training, randomly selected activations are set to zero. This can reduce overfitting; the benefits depend on the model and settings. Many implementations are already scaling in training, so that dropout does not make any further changes in the normal evaluation mode.

    ELI5 – Explained simply

    Like a sports team, where players randomly have to sit on the bench in training, so that not everyone relies on a star.

    Why it matters in practice

    Relevant as a regularization technique; test effect based on independent validation data.

    Example

    Possible example: A team compares a model with and without dropout on withheld data. Robustness against altered lighting is tested separately.

    Pitfalls and limits

    • Leave dropout active during inference (except specifically for Monte-Carlo dropout for uncertainty estimation); too high dropout rate prevents learning.

    Origin & history

    Srivastava, Hinton et al. published the basic JMLR article “Dropout: A Simple Way to Prevent Neural Networks from Overfitting” in 2014. This is not a claim of the first invention in 2014.

    Use in mid-sized companies

    • A possible regularization in the training of own models; not every architecture or task benefits equally.

    Related terms

    Matching Flow-alerta service

    TrainingAI Potential Analysis & Roadmap – Basics and Quality Control of Model Training

    Matching tool / use case

    View use casesTool: PyTorch / torch.nn.Dropout. Use Case: Investigate overfitting with appropriate dropout setting.

    Sources

    • jmlr.orgSrivastava et al. (2014), Dropout: A Simple Way to Prevent Neural Networks from Overfitting
    • docs.pytorch.orgPyTorch, torch.nn.Dropout

    Quellen geprüft am 09.10.2026.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases