2 Preliminaries Weconsidertheproblemoftrainingofatwo-layerReLUneuralnetworkwith mscalarinputsanda singlescalaroutputusingtheleast-squaresloss: min

Feb-11-2026, 16:03:01 GMT–Neural Information Processing Systems 

We show that learning in the kernel regime yields smooth interpolants, minimizing curvature, and reduces tocubic splines for uniform initializations.