S2-02 · 基础S2-02 · Foundation

标签 · 样本旁边的标准答案Label · The Answer Beside an Example

sample + label → training signal

标签是训练样本附带的标准答案或类别标记。但历史记录里的标签不自动等于可信真值:错标也会被模型认真学进去。A label is the answer or class marker attached to a training example. A historical label is not automatically trusted truth: the model will learn a wrong label just as earnestly.

01 · 答案也可能写错01 · Answer keys can be wrong

如果练习册把两道答案印错,认真照答案学习的人会不会反而学歪?If an answer key misprints two answers, can a diligent learner be trained in the wrong direction?

02错标影响实验台Mislabel Impact Bench

标签翻转型Label flip
标签总数Labels
8
训练样本training examples
错标Wrong labels
0
对比可信真值vs reviewed truth
边界偏移Boundary shift
相对复核边界from reviewed boundary

虚线 = 可信标签边界;实线 = 当前标签边界。橙圈标出被错贴的样本。Dashed = reviewed-label boundary; solid = current-label boundary. Orange rings mark mislabeled examples.

03标签链条里的 4 个角色Four Roles in a Label Pipeline

证据边界Evidence boundary
PART 01

样本Example

散点位置描述被判断的对象。The point position describes the item being judged.

PART 02

历史标签Recorded label

系统已有的答案,可能来自旧流程或人工录入。An existing answer may come from an old process or manual entry.

PART 03

可信真值Trusted truth

经过定义和复核、可用于评价的答案。An answer reviewed against a definition and fit for evaluation.

PART 04

训练信号Training signal

模型只看到当前标签,不知道它是否录错。The model sees the supplied label, not whether it was entered correctly.

04最小机制:标签决定误差方向Minimal Mechanism: Labels Set the Error Direction

(x, y)
example = (x, y)
x 是样本特征,y 是标签。y 写错时,模型仍会把错误 y 当成应逼近的目标。x is the example's feature input and y is its label. When y is wrong, the model still treats that wrong value as the target.
翻转标签Flip a label标准答案改变The answer changes
误差方向反转Error reverses正确点被当成错误A correct point looks wrong
边界被拉歪Boundary bends模型服从错误信号The model obeys the bad signal

05边界实验:把记录当真值Boundary Test: Treat Records as Truth

主动制造失败Create a failure

让两个错标进入训练Train on two wrong labels

点击后把历史记录原样交给模型,观察边界偏离可信基线多少。Pass the historical record directly into training and inspect how far its boundary moves from the reviewed baseline.

尚未执行。先看虚线与实线重合。Not run yet. The dashed and solid boundaries currently overlap.
数据库里已有的标签就是真值。An existing database label is ground truth. → ✓ 历史标签要经过定义、抽检与复核。Historical labels need definitions, sampling, and review.
只错两条,影响可以忽略。Two wrong labels cannot matter. → ✓ 边界附近的少量错标也可能改变方向。A few boundary cases can still alter the direction.
模型能自动识别哪个标签写错。The model automatically knows which label is wrong. → ✓ 没有额外证据时,错标也是训练信号。Without extra evidence, a wrong label is still a training signal.
06 · 一句话带走06 · One line to keep

标签是训练样本附带的答案,但记录不自动等于真值;错贴两个点再复核,就能检验错误标签怎样拉歪模型边界。A label is the answer attached to an example, but a record is not automatically truth. Flip two labels and review them to verify how label errors distort the boundary.