S2-18 · 进阶S2-18 · Advanced

欠拟合 · 连训练题里的规律都没学会Underfitting · Missing Even the Training Pattern

capacity too low → training error high + test error high

欠拟合是模型过于简单,连训练数据中的稳定规律也没学到。它的典型证据不是“测试差”三个字,而是训练误差和测试误差都偏高。Underfitting means the model is too simple to learn stable structure even in training data. Its signature is not merely poor testing: both training and test error remain high.

01 · 用直尺描弯路01 · Trace a bend with a ruler

一条只能轻微倾斜的直线,能同时贴近这组明显弯曲的红点和蓝点吗?Can a nearly straight line follow a visibly curved pattern across both red and blue points?

02调低曲线能力,看两种误差一起升高Lower Curve Capacity and Raise Both Errors

唯一变量:曲线复杂度One variable: curve complexity
训练误差Training error
MSE × 1000
测试误差Test error
MSE × 1000
诊断Diagnosis
欠拟合UNDERFIT
能力不足insufficient capacity
S2-18 → S2-19 → S2-20同一数据集 · 同一坐标系 · 红训练 / 蓝测试 / 黑模型same dataset · same axes · red train / blue test / black model

03欠拟合的证据链The Evidence Chain for Underfitting

simple model · missed pattern · high train · high test
CAPACITY

表达能力低Low capacity

黑线能弯曲的程度不足。The black line cannot bend enough.

PATTERN

稳定规律漏掉Pattern missed

不是噪声,而是主趋势也没跟上。It misses the main trend, not just noise.

TRAIN

训练也差Training is poor

红点普遍离黑线较远。Red points remain far from the line.

TEST

测试同样差Testing is poor

蓝点也没有被稳定规律解释。Blue points are not explained either.

04最小机制:能力上限压住了可达到的误差Minimal Mechanism: A Capacity Ceiling Limits the Best Fit

low capacity → high bias
Etrain ↑ Etest
训练误差高是关键线索:模型连已见样本都解释不好。提高到足够复杂度时,两种误差会一起下降。High training error is the key clue: the model cannot explain examples it already saw. Raising capacity enough lowers both errors.
限制复杂度Limit complexity可表达曲线变少fewer shapes available
主规律学不全Pattern is missed红蓝点都偏离red and blue both miss
双误差偏高Both errors high诊断为欠拟合diagnose underfitting

05边界实验:把曲线能力降到 0Boundary Test: Reduce Curve Capacity to Zero

主动制造失败Create a failure

用近乎平直的线解释弯曲规律Use an almost flat line for a curved pattern

训练点和测试点会同时远离模型。这个失败告诉你:此时不该继续“防过拟合”,而要先增加表达能力。Both train and test points move away from the model. The remedy is more expressive capacity, not stronger anti-overfitting constraints.

尚未执行。Not run yet.
测试差就一定是过拟合。Poor testing always means overfitting. → ✓ 先看训练误差;训练也差更像欠拟合。Check training error first; poor training suggests underfitting.
模型越简单越会泛化。Simpler models always generalize better. → ✓ 简单过头会漏掉稳定规律。Too simple misses stable structure.
继续加约束能修复欠拟合。More constraints fix underfitting. → ✓ 通常需要增能力、增特征或更好优化。Usually add capacity, features, or better optimization.
06 · 一句话带走06 · One line to keep

欠拟合不是只在新数据上差,而是模型太简单,导致训练与测试误差都偏高。Underfitting is not merely poor new-data performance; insufficient capacity keeps both training and test error high.