Miles Stoudenmire
Flatiron institute
Abstract:
Tensor networks are a tool to store and compute with tensors having many indices, which would otherwise be exponentially costly. While in physics the prototypical use of tensor networks is for approximating many-body wavefunctions, it has been appreciated over the last decade that tensor networks such as matrix product states (MPS) are much more general tools. One emerging application of tensor networks is for machine learning, where they can be used for a variety of tasks such as supervised learning, generative modeling, and compression of neural network parameters. Compared to other machine learning approaches, tensor networks offer advantages such as scalability, more sophisticated training algorithms, and theoretical clarity. After reviewing the basics of tensor networks, I will introduce the setting of machine learning and explain how tensor networks can be used for machine learning tasks, showing exemplary results for both classical (non-physics) applications and physics applications.