Why do deep neural networks generalize so well even when they have far more parameters than training examples? This question has long puzzled researchers because traditional statistical learning theory predicts that such over-parameterized models should overfit badly. A new paper on arXiv tackles this puzzle head-on, proposing a unified explanation it calls “neural expansion.”

The abstract describes neural expansion as a single mechanism that can account for the generalization behavior of deep networks. It positions this mechanism against a backdrop of existing theoretical approaches—such as margin-based bounds—which have offered partial explanations but no comprehensive picture. The paper suggests that these disparate paradigms may be reconciled under one framework.

Because the announcement only includes the abstract, the technical details of neural expansion are not yet visible from the summary alone. Still, the promise of a unified theory for one of deep learning’s most striking empirical properties is notable. The full paper is available on arXiv for those who want to dig into the mathematics.