Tara Akhound-Sadegh ( Université McGill )


Neural networks have proven very successful in a variety of applications, such as image recognition and translation. However, they still struggle when modeling complex physical systems. Recent developments in machine learning, and specifically in deep learning have drawn inspirations from concepts in physics to achieve a better performance on deep models. For example, some of these models have been inspired by conservation laws in physics, building models that are equivariant to some specific symmetry transformations (such as rotations), meaning that a symmetry transformation of the input results in a predictable transformation of the output. The goal of this project is to use mathematical concepts, such as group theory and representation theory, to develop a deep model that is able to learn the underlying symmetry of the observation space using only a series of observations of the environment as its input.