Implicit Regularization in Matrix Sensing: A Geometric View Leads to Stronger Results
We may think of low-rank matrix sensing as a learning problem with infinitely many possible outcomes where the desired learning outcome is the unique low-rank matrix that agrees with the available data. To find this desired low-rank matrix, for example, nuclear norm explicitly biases the learning outcome towards being low-rank. In contrast, we establish here that simple gradient flow of the feasibility gap is implicitly biased towards low-rank outcomes and successfully solves the matrix sensing problem, provided that the initialization is low-rank and sufficiently close to the manifold of feasible outcomes. Compared to the state of the art, the geometric perspective adopted here leads to substantially stronger results.
READ FULL TEXT