Collaborative Research: RI: Medium: Lie group representation learning for vision

点击翻译按钮获取中文摘要

The quest to build intelligent machines capable of sensing, understanding and acting in their environment presents one of the great scientific challenges of our time. Despite recent advances in artificial intelligence (AI), the realization of robust, autonomous vision systems that understand and interact with the physical world remains elusive. Mathematically, vision requires understanding the relationships among an immense variety of object shapes, each subject to an immense variety of geometric and lighting transformations, leading to an explosion of possible visual scenes. This project aims to break through this barrier by developing a mathematically grounded computational theory of vision that will enable a new class of neural network learning algorithms to parse visual scenes into their constituent objects and transformations, thereby enabling computers to better represent the world around them. The results and computational tools arising from this research will be disseminated to the scientific community and general public through courses, seminars, hackathons, and open-source software contributed to the Geomstats library.The premise of this project is that the current limitations of AI and computer vision can be addressed with an appropriate mathematical framework, Lie theory, that models the hierarchical structure of natural transformations in the visual world. The investigators will develop generalizations of foundational signal processing transforms through explicit Lie group operations encoded in learnable G-Modules (Group-Modules). These modules directly tackle the combinatoric explosion in vision by factorizing images into shapes and their underlying transformations. Specifically, the team will develop G-modules that learn group-equivariant representations of the transformations contained in natural images (Aim 1), robust representations of shape by collapsing group orbits only with respect to specific transformations (Aim 2), and disentangling of transformation and shape via factorization (Aim 3). The modules are assembled into hierarchical architectures that can learn complex representations of transformations and shapes (Aim 4). Together, these aims provide a new paradigm that grounds existing models of vision and gives a set of guiding principles for the design of future deep learning architectures with enhanced abilities to sense and understand the world.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.