Economists keep getting overparameterization wrong: SGD's implicit regularization is the point

Afinetheorem · x · 2026-09-07

Economist Jesús Fernández-Villegas pushes back on the common economist complaint that deep neural networks 'require regularization,' noting economics regularizes all the time—control variables, Bayesian priors, splines, kernels, sieves—just implicitly and without cross-validation. Large neural networks make regularization implicit: SGD picks the minimum-norm solution among those fitting the training data. The thread also highlights Mahdi Kahou's joint work 'The Blessings of Overparameterization,' showing heavy overparameterization reduces dependence on initialization.

Original post →

More from Research

Research channel →