RI: Small: Visual Cortical Recurrent Circuits for Manifold Learning and Memory Attention

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

The human brain consumes millions of times less energy than artificial intelligence (AI) systems, yet it remains more flexible, versatile, and effective at solving complex problems. A key reason lies in the brain’s ability to learn abstract concepts and their relationships, and to construct internal models of the world--rather than simply memorizing and retrieving patterns from massive datasets. This project aims to investigate the computational mechanisms underlying a recently discovered neural process in the brain: the ability of neurons in the early visual cortex to rapidly form local recurrent circuits, enabling them to quickly encode relationships between concepts. Although these neurons primarily act as local feature detectors, they can dynamically adjust their responses based on global context through these recurrent connections. However, the computational rationale behind this mechanism remains poorly understood. The project seeks to develop a machine learning framework to formalize this rapid recurrent plasticity. The central hypothesis is that these recurrent circuits do more than encode global context as attractors--they also perform manifold transformations that bring semantically related concepts closer together in the neural activity latent space for encoding dependency among concepts, while compressing irrelevant dimensions of variation. The investigator will test this theory in tasks such as associative learning, de-noising, and pattern completion, combining computational modeling with neurophysiological experiments. The project will evaluate the predictive power, limitations, and practical utility of these biologically grounded models in real-world computer vision applications. This research will offer a new conceptual framework for understanding cortical recurrent circuits and pave the way for a new generation of biologically inspired AI systems--systems that are more efficient, generalizable, and aligned with the flexible intelligence of the brain, with broad potential impact on neuroscience, technology, and national security. 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.