Miles Stoudenmire
Flatiron Institute
Abstract
In this second talk, I will focus more on promising future directions for tensor network based machine learning, such as the use of other tensor networks besides matrix product states. One particularly promising direction is developing more sophisticated algorithms for training compared to basic gradient-descent algorithms. I will show how applying a density-matrix based, direct training algorithm to a synthetic data set of even-parity bit strings not only leads to state-of-the-art results, but also supports a theoretical explanation of how the model learns and an accurate prediction of how well it generalizes.