For many latent variable models, learning can be done by tensor decomposition on the third-order moment matrix. However, many previous works either focus on the undercomplete case (when the number of hidden components is smaller than the dimension), or require higher order moments. In this talk I will show simple tensor power method still works even for overcomplete case.

Based on joint works with Anima Anandkumar and Majid Janzamin.

Video Recording