Training-Level Memorization Detection

What It Is

Memorization detection monitors the training process for indicators that specific examples are being encoded into model parameters at a level that constitutes memorization rather than generalization. The detection operates on gradient patterns, parameter sensitivity, and output fidelity metrics that distinguish memorization from appropriate learning.

Why It Matters

Memorization creates legal, ethical, and technical risks. Memorized copyrighted content can be reproduced verbatim, creating rights violations. Memorized personal data can be extracted, creating privacy violations. Memorized training examples produce brittle behavior that fails on novel inputs. Detection during training prevents these risks from materializing.

How It Works

The detection system tracks per-example gradient norms, parameter sensitivity patterns, and periodic output fidelity tests. When a training example produces gradient patterns indicating it is being memorized beyond the depth profile's permitted integration level, the system flags it for intervention. Intervention options include reducing the example's learning rate, increasing regularization, or excluding the example from further training.

The memorization thresholds are defined by the training governance policy and may vary by content class and entropy band.

What It Enables

Memorization detection enables training that produces models with appropriate generalization properties. Models trained under memorization governance can be deployed with greater confidence that they will not reproduce specific training examples. This is particularly valuable for models trained on rights-governed content, where any memorization constitutes a potential rights violation.