先看這個問題為什麼值得解Why this problem is worth solving
情境先行Start with the situation
訓練影像裡每隻牛都站在草地上。模型可能學到動物形狀，也可能只記住綠色背景；單看準確率無法分辨。 Every training image of a cow shows grass. A model may learn animal shape or green backgrounds; the same accuracy cannot tell which.
如果表徵保留的是偶然相關，在原資料分布上的探針分數可能很好看，換到部署環境卻立即失效。 If a representation preserves accidental correlations, an in-distribution probe can look strong and fail immediately under deployment shift.
先備知識Prerequisite
- 理解訓練集、測試集的切分，以及特徵的基本概念 Understand train/test splits and the idea of features
讀完你能做到After this path
- 區分表徵品質與探針模型的容量 Separate representation quality from probe capacity
- 針對干擾因素設計介入測試與分布外切分 Design nuisance interventions and out-of-distribution splits
- 判斷不變性何時有幫助、何時會丟失任務資訊 Judge when invariance helps and when it discards task information
閱讀框架Concept frame
- 核心問題Question
- 一個表徵在線性探針（linear probe）上得分較高，是否就代表它更能泛化？ Does a better linear-probe score mean a representation generalizes better?
- 對照軸Comparison axis
- nuisance 與目標在訓練資料中的相關程度 Training correlation between nuisance and target
- 判讀方式Evidence to inspect
- 並排閱讀三種相對情境，再回到論文檢查消融實驗、反事實測試與適用範圍。Read three relative scenarios side by side, then inspect the paper's ablations, counterfactual tests, and scope statements.
回答後查看解析Answer, then view the explanation
你的預測Make a prediction
哪個實驗最能檢查表徵是否依賴背景捷徑？ Which experiment best checks whether a representation relies on a background shortcut?
看現象，不急著講故事Observe before narrating
三個診斷訊號Three diagnostic signals
這些是待檢查的診斷假說。下一步用條件比較、消融實驗或反事實測試排除其他解釋。These are diagnostic hypotheses to test. Use condition comparisons, ablations, or counterfactual tests to rule out alternatives.
一次看完三種情境Read three scenarios side by side
三種情境對照Three-scenario comparison
當 nuisance–target 相關提高，哪種 split 才能維持診斷力？ As nuisance–target correlation rises, what split preserves diagnostic power?
對照軸：Comparison axis: nuisance 與目標在訓練資料中的相關程度 Training correlation between nuisance and target
捷徑誘因較弱，可優先比較 sample efficiency 與表示可讀性。 Shortcut pressure is weaker; prioritize sample efficiency and representation accessibility.
需要環境分層、counterfactual augmentation 與 group metrics。 Use environment strata, counterfactual augmentation, and group metrics.
僅靠 IID split 幾乎無法辨識目標特徵與捷徑。 An IID split alone can barely distinguish target features from shortcuts.
從直覺到機制From intuition to mechanism
把機制講清楚Explain the mechanism
好的表徵像行李箱：保留任務需要的物品，也不該把每間旅館的地毯一起帶走。 A useful representation is luggage: it keeps task-relevant items without carrying every hotel's carpet.
模型從觀測資料、資料增強方式與環境中學出下游任務可讀取的特徵；目標是保留任務所需資訊，同時降低對無關干擾因素的敏感度。 Input: observations plus augmentations or environments. Output: features readable by downstream tasks. Objective: retain task information while controlling nuisance sensitivity.
- 01 資料與 augmentation 指定哪些變化被視為相同。 Data and augmentations specify which transformations count as equivalent.
- 02 objective 把相似、預測或壓縮約束寫入表示空間。 The objective writes similarity, prediction, or compression constraints into representation space.
- 03 probe 與 transfer protocol 測試哪些資訊可被取用。 Probe and transfer protocols test which information is accessible.
把單一 probe、單一資料集或漂亮的 embedding 圖當成普遍 representation quality。 Treating one probe, one dataset, or an attractive embedding plot as universal representation quality.
表徵比較要同時固定 probe capacity，改變 nuisance，並在多個 shift 下報告。 Representation comparisons should fix probe capacity, intervene on nuisances, and report across multiple shifts.
不用回看頁面的自我檢查Check your recall without looking back
三個回想題Three recall questions
先口頭回答，再展開解析。Answer aloud before opening each explanation.
01 probe score 同時量到哪兩件事？ What two things can a probe score mix together?
表示中的資訊，以及 probe 從有限資料讀出資訊的能力。 Information in the representation and the probe's ability to extract it from finite data.
02 為何 invariance 不一定好？ Why is invariance not always beneficial?
若被消除的變化與目標相關，表示會丟失完成任務所需資訊。 If the removed variation is target-relevant, the representation loses information needed for the task.
03 OOD split 應改變什麼？ What should an OOD split change?
改變預期部署會變動的環境因素，同時維持清楚的目標定義。 It should vary environmental factors expected to shift in deployment while keeping the target definition clear.
把工具帶回你的研究Take the tool back to your research
最後一個任務One final task
從一篇 representation 論文的 metadata／abstract 寫出『想保留的訊息、想忽略的 nuisance、使用的 probe、尚未測的 shift』四格；未知項保持未知。 From a representation paper's metadata or abstract, fill four boxes—information to retain, nuisance to ignore, probe used, and untested shift—leaving unknowns explicit.