撞上LeCun新研究！清华团队让神经网络“简单性”可测可优化

TL;DR - A Tsinghua team proposes Effective Degree (ED), a differentiable measure of neural-network simplicity based on polynomial approximations along data interpolation paths. ED strongly tracks generalization and can regularize training across vision, language, vision-language, and reinforcement-learning tasks.

• ED measures function complexity by weighting Chebyshev polynomial components according to their degree and magnitude.
• On CIFAR-10 ResNet18 experiments, ED correlated with the generalization gap at 0.99, outperforming parameter norms and sharpness metrics.
• Because polynomial fitting is differentiable, ED can be added directly to training objectives to discourage unnecessary high-order behavior.
• ED does not ensure causal or robust features: models may still favor simple but unreliable shortcuts.