Collaborative Research: SCALE MoDL: Representation Theoretic Foundations of Deep Learning

点击翻译按钮获取中文摘要

In the past decade, deep learning has had transformative impacts across society. However, progress has often relied on heuristic methods, massive data, and great computing power. This comes with limited theoretical understanding and has at times given rise to failures of generalization and vulnerable performance in extreme scenarios. This project will address these limitations by developing strong theoretical foundations for deep learning using representation theory, which is the mathematical study of symmetry. Symmetry plays a key role in human reasoning. Greater understanding of the role symmetry plays in deep learning will unlock a variety of improved models. These include models that can learn from scientific knowledge and not just raw data, models with trustable, guaranteed performance, and models that can recombine patterns they have already learned — as humans do easily — to generalize to new situations more rapidly. An explicit goal of this project is to broaden research into why deep learning works. To this end, the investigators will integrate the research into education and establish a mentorship program for high school students from groups underrepresented in science.The goal of the research is to understand the role of representation theory in enabling efficient optimization and improved generalization of deep learning even in domains with approximate or unknown symmetry. This project pursues three lines of research that will broaden the impact of representation theory in deep learning beyond strict inductive biases. The first is the trade-off between the degree of symmetry in the model and the degree of symmetry in the domain. This line of research will study networks that combine equivariant and non-equivariant features. The second line of research will examine learning symmetry directly from data to improve generalization in domains without known symmetries. The third aim is to develop a theoretical basis for deep learning using quiver representations. This perspective reveals the symmetry of the structure of deep-learning models themselves, through their parameter spaces, even when the domains have no obvious symmetry.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.