Collaborative Research: MFAI: Uncovering the Mathematical, Statistical, and Contextual Mechanisms behind Memorization in Artificial Intelligence

The study of memorization in Artificial Intelligence (AI) models has become increasingly important as AI has become more widely adopted in society. This phenomenon presents significant risks, including potential privacy violations, copyright infringement, and failure to generalize beyond the limited training data. Notwithstanding its critical importance, there is a lack of understanding of the factors underlying memorization in AI, and without such understanding, the memorization problem cannot be systematically addressed, managed or mitigated. This project will develop foundational theory that will elucidate the mathematical, statistical, and contextual principles that affect memorization in AI, culminating in the development of reliable, robust, and privacy-preserving AI models. A unique aspect of the project's research framework is that it combines optimization theory, dynamical systems theory, and information theory to develop useful insights into the interplay between memorization, generalization, privacy, reproducibility, and model robustness. These insights will lead to the development of AI models that are less susceptible to privacy violations, model overfitting, and adversarial attacks, thereby enhancing the trustworthiness and applicability of AI across many domains. The investigators will integrate the research into the curriculum, engage undergraduate students in research, and hold workshops to foster collaboration and share our findings with the wider academic and professional communities. The research in this project centers around several interconnected themes that provide a systematic and innovative framework for studying the memorization phenomenon in AI. The first theme is the development of computationally efficient and principled metrics for quantifying memorization in AI models. The project will leverage harnessing memorization proxies for various learning objectives, including machine unlearning and deduplication. The second theme is the mathematical characterization of performance tradeoffs and dependencies between memorization, generalization, and the replicability of large AI models in diverse settings. Based on such findings and characterizations, the research will lead to the development of mechanisms to prevent overfitting while ensuring efficient learning and generalization to practical scenarios where the training and test data contain frequent outliers governed by long-tailed data distributions. The third theme is validation of the project's theory and methods on important practical AI applications, including multitask inference of dynamical systems, heterogeneous and federated learning problems, and understanding incentives in human decision making. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.