Tsinghua Team Proposes 'Effective Degree' to Accurately Predict Neural Network Generalization

机器之心 · wechat · 2026-08-06

Qianjue Tech and Tsinghua University proposed a new metric called 'Effective Degree' (ED), turning the abstract 'simplicity bias' of neural networks into a computable and optimizable tool. This research predates a similar world model theory recently highlighted by Turing Award winner Yann LeCun by about a year.

Solving the Pain Point: Existing complexity metrics (like parameter norm or sharpness) often depend on parameterization and are hard to compute stably on real models. ED constructs interpolation paths between data points to 'slice' the high-dimensional neural function, fitting it with orthogonal polynomials to quantify high-order components.

Core Advantages:

Original post →

More from Research

Research channel →