Automatic Discovery of Visual Circuits
Achyuta Rajaram developed an automatic method to discover which parts of a computer model that analyzes images are involved in decision-making. This knowledge sheds light on what these algorithms are 鈥渢hinking,鈥� which can help make them more effective, fair, and safe.
Achyuta Rajaram, 17, of Exeter, improved automatic discovery of visual circuits for the computer science project that he submitted to the Regeneron Science Talent Search. In machine learning, computer algorithms find patterns in data to answer important, practical questions. Achyuta鈥檚 research improved our ability to discover what computer models, that find patterns in images, are 鈥榯hinking鈥� when they analyze a photo and which parts of their 鈥榤echanical brains鈥� are contributing to the decision making. For example, when a model identifies a car in a photo, does it first identify wheels and use this to identify 鈥榗ar-ness,鈥� or does it look for something else?
Achyuta鈥檚 key contribution to this effort was to develop an automatic method for recognizing which parts of the algorithms identify what. This knowledge sheds light on what these algorithms are 鈥榯hinking,鈥� which can help make them more effective, fair, and safe.
Achyuta, the son of Nivedita Chevvakula and Rajaram Ramaswamy Kumaraswamy, attends Phillips Exeter Academy, where he is co-head of the physics, chemistry, and chess clubs. His passion for jazz drumming has led him to play in groups ranging from small combos to symphonia orchestra.
Beyond the Project
Achyuta grew the school鈥檚 chess club from three players to more than 30. He teaches chess strategy, analyzes games for participants, and organizes student trips and in-house tournaments.
FUN FACTS: Achyuta loves cats but doesn鈥檛 own any at home, so he used machine learning to index 2,300 cat memes for text search. He says it’s a nice way to take a break鈥� just to stare at some cute cats.
