Understanding Over-parametrization Through Matrix Sensing

Опубликовано 25 января 2018, 17:50

We study the problem of recovering a low-rank matrix from linear measurements using an over-parameterized model. We show that the gradient descent process on the square loss function, starting from a small initialization, can converge to the ground truth matrix. Although the total number of observations is much smaller than the total number of parameters.

See more at microsoft.com/en-us/research/v...