D\'ej\`a Vu Memorization

Déjà vu memorization refers to the unintended storage and retrieval of specific training data by large machine learning models, impacting generalization and raising privacy concerns. Current research focuses on quantifying this memorization in various architectures, including vision-language models and large language models, and developing methods to mitigate it, such as employing reinforcement learning to discourage verbatim or approximate reproduction of training examples. Understanding and controlling déjà vu memorization is crucial for improving model robustness, ensuring data privacy, and advancing the trustworthiness of AI systems.