🦜 过拟合：背题 vs 懂规律🦜 Overfitting: Memorizing vs Understanding

上一片里，我们把学习率的步子调得刚好，训练误差一路往下。可是——练得越熟，就学得越好吗？把练习册答案全背下来的小机器人，考试一换题就傻眼。这一片说说这种“背题”毛病。 Last leaf we tuned the learning rate just right, and the training error slid steadily down. But does “better on the practice book” really mean “learned more”? A robot that memorizes every answer in the workbook freezes on a fresh exam question. This leaf is about that “memorizer’s” problem.

第 1 步：两个学生，两种学法Step 1: Two students, two ways to study

先看同一个班里的两个学生——他们的学法完全不一样。Meet two students in the same class — their study styles could not be more different.

背题机器 vs 懂规律的机器 — 小明把练习册做了一遍又一遍，连答案在第几页都背下来：练习册次次满分，一换新题就发懵。小红边做边想规律，练习册偶尔错一两道；可新题一来，她照样会做。模型也一样：练习册＝训练数据，新题＝没见过的数据。把练习册背下来，不等于学会。 Ming drills the workbook again and again until he even remembers which page each answer sits on: perfect on the workbook, clueless on a fresh question. Hong thinks about the pattern as she practices and misses a question or two — yet new questions don’t scare her. Models are the same: the workbook is the training data, fresh questions are data it has never seen. Memorizing the workbook is not learning.

怎么知道模型是在“学”还是在“背”？把两条误差曲线画出来就清楚了。How can you tell whether a model is learning or memorizing? Plot two error curves and it becomes obvious.

第 2 步：两条曲线的故事Step 2: A tale of two curves

训练误差一路降，测试误差先降后升 — 蓝绿线是训练误差——练习册上的出错率，一路向下，越练越熟。橙线是测试误差——新题的出错率，先跟着降，随后却抬头向上。抬头那一刻，模型从“学规律”滑向“背题”。虚线和小旗标出了翻转点：过拟合从这里开始。 The blue-green line is training error — the mistake rate on the workbook; down and down it goes. The orange line is test error — the mistake rate on fresh questions; it falls along at first, then turns and climbs. The moment it turns, the model slips from learning patterns into memorizing answers. The dashed line and little flag mark the flip: overfitting starts here.

把两种学法画在同一张散点图上，一眼就能分出来。Put both study styles on one scatter plot and the difference jumps out.

第 3 步：穿过每个点，还是抓住规律Step 3: Touch every dot, or catch the pattern

歪扭线穿过每个点 vs 平滑线抓住规律 — 歪扭的线硬要精确穿过每一个练习点，连小抖动都背了下来——遇到新题却东倒西歪。平滑的线允许一两个点不在线上，却抓住了整体趋势——新考点落下来，正好接住。所以聪明模型不追求练习满分，它追求的是：换新题，也稳。 The wiggly line insists on touching every practice dot, memorizing even the tiny jitters — then flails on new questions. The smooth line lets a dot or two sit off-curve but keeps the overall trend — a new test point lands right on it. A smart model doesn’t chase a perfect practice score; it chases staying steady on new questions.

🎮 你来挑模型（40 秒）🎮 Your turn: pick the model (40 seconds)

下面 4 轮，每轮同一组练习点、两条线。点出“考试型选手”——新考点来了也稳的那条。Four rounds below. Each shows the same practice dots and two lines: tap the “exam-ready” one — the line that stays steady when a new test point appears.

一句话记住它：训练误差很低、测试误差却抬头，就是模型在“背题”——这叫过拟合。 Remember it in one line: training error keeps dropping while test error turns upward — the model is memorizing. That’s overfitting.

训练误差：练习册上的出错率，只会一路降Training error: mistakes on the workbook — it only goes down 测试误差：新题上的出错率，先降后升＝过拟合来了Test error: mistakes on new questions — down then up means overfitting 考试型模型：允许小错，抓的是规律Exam-ready model: allows small misses, keeps the pattern

← 浏览全部 302 个知识点← Browse all 302 topics

内容参考 OpenStax 等公开教材，多来源核对 · AI 生成、人工审核 · 发现错误欢迎指正，帮这片叶子长得更好。 Based on OpenStax and other open textbooks, cross-checked across sources · AI-generated, human-reviewed · Spotted a mistake? Tell us — help this leaf grow.