NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:4545
Title:Gradient Dynamics of Shallow Univariate ReLU Networks


		
The paper studies the training dynamics of 1D ReLU networks with one hidden layer in the kernel regime and the adaptive regime, and shows connections with spline interpolation. The reviewers all appreciated the importance of the contribution in the kernel regime, but found that the analysis of the adaptive regime is not as clear/rigorous. Still, there is an overall consensus that the paper is above the bar for acceptance.