One of the main reasons behind the success of convolutional neural networks (CNNs) is the fact that they exploit the structure of image data. Designing algorithms and architectures that make use of more general symmetries in data has been a very active area of research. Despite the great progress, a common limitation is the reliance on having the symmetry structure of data be predefined.

In this project, we focus on learning symmetries of the input data directly from the data itself, as opposed to relying on a priori specified symmetries. The end goal of the project is to design learning algorithms that can make efficient use of structured data and achieve good sample complexity.

I am interested in SuperUROP because it gives me the opportunity to focus for a longer period of time on a problem I deeply care about under the mentorship of truly amazing researchers. I hope during the program I’ll be able to apply the knowledge I have gathered throughout my years at MIT and contribute something valuable to the field.