Counterfactual Memorization in Neural Language Models

TL;DR

This paper introduces the concept of counterfactual memorization in neural language models, analyzing how individual training data points influence model predictions and memorization, which is crucial for understanding privacy risks.

Contribution

It formulates a new notion of counterfactual memorization, enabling identification and analysis of specific training examples that influence model outputs.

Findings

Counterfactual memorization can be estimated for training examples.

Influence of memorized data affects validation and generated texts.

Method helps trace the source of memorization in language models.

Abstract

Modern neural language models that are widely used in various NLP tasks risk memorizing sensitive information from their training data. Understanding this memorization is important in real world applications and also from a learning-theoretical perspective. An open question in previous studies of language model memorization is how to filter out "common" memorization. In fact, most memorization criteria strongly correlate with the number of occurrences in the training set, capturing memorized familiar phrases, public knowledge, templated texts, or other repeated data. We formulate a notion of counterfactual memorization which characterizes how a model's predictions change if a particular document is omitted during training. We identify and study counterfactually-memorized training examples in standard text datasets. We estimate the influence of each memorized training example on the validation set and on generated texts, showing how this can provide direct evidence of the source of memorization at test time.