NeurIPS 2020

On Uniform Convergence and Low-Norm Interpolation Learning

Meta Review

The paper considers over-parameteried linear regression, explores uniform convergence to see if consistent minimum norm solution can be recovered, illustrates that UC does not help here, but focusing on a restricted setting (interpolation) holds promise. The reviewers appreciated the main results. It is unclear if the analysis and results has implications beyond the linear setting considered.