過学習と過小学習のクイズ

クイズに答えると、答えと解説が表示されます。

1. Under comparable train and held-out conditions, an illustrative model scores 98% on training and 71% on held-out data. Which fitting concern is consistent with that gap?
   • Underfitting because training performance is necessarily poor
   • A guarantee that the held-out set has no information
   • Overfitting to training-specific patterns
   • Proof that every larger model will fix the gap
2. Which performance pattern can be consistent with a model underfitting useful structure?
   • Poor performance on both its training and held-out data
   • Excellent training performance with poor held-out performance only
   • Guaranteed perfect performance on both datasets
   • A guarantee that the objective is minimized globally
3. In a bias-variance analysis, what does high estimator variance concern?
   • An unavoidable lack of any measurement noise
   • Sensitivity of fitted predictions to which training sample is used
   • A guarantee that the model is a linear function
   • The wall-clock time of a training epoch
4. A house-price model has too limited a representation of relevant inputs and fits poorly. Which candidate change should be evaluated?
   • Removing all relevant inputs without evaluation
   • Assuming a bigger model must always be correct
   • Using the final test set repeatedly as training data
   • Adding useful features or comparing a more suitable model representation
5. 過剰適合の検出にホールドアウト検証またはテスト セットが不可欠なのはなぜですか?
   • モデルが学習していないデータのパフォーマンスを測定し、一般化ギャップを明らかにします
   • トレーニングをより速く実行できるようになります
   • モデルが決してエラーを起こさないことを保証します
   • 損失関数の必要性を置き換えます
6. What pattern is called model-size double descent in the studied examples?
   • Every dataset must produce exactly the same error curve
   • Training loss is halved exactly twice on every epoch
   • Test error can rise near an interpolation regime and then fall as capacity increases
   • Two models are always trained and averaged
7. Why should the final test set be kept separate from repeated model-selection choices?
   • It must be used for every learning-rate decision
   • It guarantees zero future prediction error
   • It removes the need for validation data
   • Repeated selection on it can make the final reported estimate optimistic
8. In a squared-error bias-variance decomposition with fixed information and observation conditions, what can model-capacity changes alone fail to remove?
   • The definition of the model title
   • The conditional noise component
   • Every possible systematic bias by definition
   • The requirement to calculate predictions