Current machine-learning models for action recognition operate either by first extracting features such as keypoints and bounding boxes and then classifying the action, or by performing full-frame action classification. Recent work exploring how humans interpret images and recognize actions show that we are capable of identifying and localizing a high number of semantic features and parts in images. Our research plans to use the best of both worlds: extract semantically-relevant features from pertinent bounding boxes (full interpretation in a temporal context). The focus will be on a subset of actions performed by hand, which poses the challenge of obstruction and occupying a small number of pixels in real- life images.

I am participating in SuperUROP because I am very interested in artificial intelligence (AI) and want to be involved in furthering the state of the art. I believe AI will be a big part of the solutions to humanity’s problems in the 21st century and I want to be part of that effort. I have done research in natural-language processing (NLP) and interned at an AI startup. I hope to learn a lot through this program and publish an impactful paper.