千诀科技与清华团队提出有效阶指标，让神经网络简单性可测可优化
神经网络“简单性”有了可测可优化的量化指标，比 LeCun 类似研究早一年。
千诀科技联合清华团队提出“有效阶”（Effective Degree，ED）指标，沿数据点插值路径用正交多项式拟合神经网络，将“简单性偏置”转化为可计算、可优化的量。实验显示 ED 能预测泛化差距，并可在训练中直接优化，提升图像、文本、强化学习等任务性能。
正文摘录
.38.05.png) Qianjue Technology, in collaboration with Tsinghua University, has proposed a new "complexity ruler" for neural networks. As model training data grows, we have increasingly more benchmarks to evaluate whether a model answers correctly, yet it remains difficult to judge whether it has truly learned transferable regularities or merely "memorized" training data in a complex way. The team's proposed "Effective Degree" (ED) is the first to transform the abstract "simplicity bias" in this research line into a metric that can be computed, compared, and directly used in training on real-scale models. This research not only provides a new tool for understanding why…