Symmetry is an important visual feature that is omnipresent in nature, science, and artwork. The perception of symmetry is relevant for understanding a visual system’s ability to encode long-distance spatial relationships, rather than purely local ones. Though an easy visual feature for humans to learn, symmetry presents challenges for deep neural networks (DNNs). This project will investigate learned representations in DNNs for symmetry. We will explore how training DNNs on simplified inputs may lead them to learn a solution that captures long-range relationships, and identify symmetry in image classes not seen during training. The results will also be instructive for other visual tasks that require the use of long-range relationships, which are crucial for robust image understanding.

I’m participating in SuperUROP to gain experience with the full life cycle of a research project — from idea conception, to formulating a research plan, to presenting my results. I’m excited to build on my UROP from last year and delve deeper into improving the visual capabilities of deep neural networks, while learning from my mentors and exploring a future in research.