hidden width

New research challenges standard neural scaling exponent derivation

A new research paper proposes an alternative to the standard geometric derivation of neural scaling exponents, which typically relies on the intrinsic dimension of a data manifold. The authors demonstrate that for modul…