What question did this study set out to answer?

To measure and characterize how neural language models memorize and reproduce verbatim training sequences based on model capacity, data duplication, and context length.

February 15, 2022Open Access

Quantifying Memorization Across Neural Language Models

View Full Paper

Ask AI

Bookmark

Authors

NCNicholas CarliniGoogle (United States)DIDaphne IppolitoCarnegie Mellon University MJMatthew JagielskiNortheastern University

Discussion

Member takes

Implication

Quantitative analysis demonstrates log-linear increases in verbatim data memorization in neural language models, highlighting escalating privacy risks as systems scale.

Key Points

To measure and characterize how neural language models memorize and reproduce verbatim training sequences based on model capacity, data duplication, and context length.
Assessed verbatim training data emission across varying model capacities and diverse language model families.
Quantified memorization rates as functions of example duplication frequency and prompt token context size.
Verbatim memorization scales log-linearly with model capacity, duplicate training frequency, and prompt context length.
Cross-family model comparisons reveal more complex dynamics, indicating that memorization is more widespread than previously recognized and worsens with scale without mitigations.

Cite This Study

Carlini et al. (2022) studied this question.

synapsesocial.com/papers/6a0eab4ffca5c6c9f447a032 https://doi.org/10.48550/arxiv.2202.07646

View Full Paper

Ask AI

Bookmark

Implication

Quantitative analysis demonstrates log-linear increases in verbatim data memorization in neural language models, highlighting escalating privacy risks as systems scale.

Quantifying Memorization Across Neural Language Models

Discussion

Member takes

Implication

Key Points

Cite This Study

Quantifying Memorization Across Neural Language Models

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider