Noise addition

As explained in Machine Learning System Design

The deliberate perturbation of data values with random noise to assess or improve the stability of model decisions.

Noise addition perturbs sample values to test how stable threshold-dependent metrics are. Small random changes can shift precision and recall, revealing a narrow, fragile decision-boundary margin. section 7.4

Treat noise level as a hyperparameter and combine perturbation with Poisson bootstrap to estimate confidence bounds for recall at a given precision or specificity. section 7.4

Distribution of precision and recall with sample size of 200;000; every point is a bootstrapped original dataset, no noise added.

Compare the bootstrapped precision and recall distributions without added noise against the perturbed results. section 7.4

Distribution of precision and recall with a sample size of 200,000; every point is a bootstrapped original dataset, noise added.

See how adding noise changes the bootstrapped precision and recall distributions. section 7.4

How Machine Learning System Design covers it

How deep it goes

Demonstrates effects on precision and recall, then combines noise as a hyperparameter with Poisson bootstrap confidence bounds.

Examples

Compares bootstrapped precision and recall distributions with and without minor normally distributed noise.

What you can do after

Evaluate model robustness using bootstrapped lower confidence bounds for recall at a given precision or specificity.

How it connects

Connects threshold selection, precision, recall, specificity, decision-boundary margins, and Poisson bootstrap.

liveBookmanning.com

Read Machine Learning System Design on liveBook Get the book at manning.com

Related concepts

• mode collapseaddresses · Mode collapse is a generative-model failure in which outputs concentrate on a narrow range of patterns and fail to represent the diversity of the target distribution.
• anonymizationkind of · Anonymization is the alteration or removal of data that reduces the ability to identify individuals or link them to records.
• privacyused by · Privacy is the protection and controlled handling of information or activities to prevent inappropriate access, collection, use, or disclosure.

This explainer was generated by Manning's knowledge graph from the text of Machine Learning System Design. The section links point to the parts of the book it draws on.