Understanding Over-parametrization Through Matrix Sensing

  • Yuanzhi Li | Princeton University

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.

Watch Next