Traditional computer vision benchmarks primarily evaluate models on static images, often failing to reflect real-world performance. In this project, we propose test-time world simplification, where models actively modify their visual environment to make recognition easier. Inspired by models of human attention, we will develop a framework that manipulates images through cropping, magnification, and semantic editing while performing classification. We will evaluate this approach on challenging image datasets for current vision systems and explore how we can transform vision models from passive observers into active perceivers.

I am participating in SuperUROP because I want to explore advances in machine learning that I wouldn’t see in a traditional classroom. Coming from a background in biology, I’ve had limited exposure to computer science. However, after taking classes at MIT, I’ve become deeply curious about computer science and even switched my major to Course 6-7 to add-on computer science to molecular biology. I also completed an internship in deep learning last summer and absolutely loved it. I want to continue to experience the thrill of innovation in technology through this research experience.