Overfitting and the Bias-Variance Trade-off · seed 1 · A4, ink-friendly. The answer key prints on its own page for grown-ups.

Too stiff, too wobbly, or just right

Computing · Machine Learning · ages 19-21
Name ______________________   Date ____________

  1. A straight line fits curving data and misses the same way no matter how much data arrives. What is this?

    • Bias
    • Variance
    • Good luck

  2. A model always predicts 0.4 while the true share is 0.5. With bias as average prediction minus truth, what is the bias?

    Answer: ______________

  3. A model with high variance changes its answer a lot when the training sample changes slightly.

    Circle one:   True   False

  4. True or false: if a model's fitted curve changes a lot every time the training sample changes slightly, the model has high variance.

    Circle one:   True   False

  5. Which remedy best cuts variance?

    • Averaging many models over resampled data
    • Freezing one stiff line forever
    • Collecting zero new samples

  6. A team fits a straight line to 10 data points. They then add one point far away from the rest, and the line tilts sharply toward it, so the line now predicts the other data worse. What is the main source of the failure, and why?

    • Variance, the fit moves with the sample
    • Bias, because a straight line is too simple
    • It is just noise, so no model could do better
    • Overfitting is impossible with a straight line

  7. A model always predicts 0.4 for the true proportion of visitors who buy something, no matter what data it is shown. The true proportion is 0.5. Bias is the average prediction minus the true value. What is the model's bias?

    Answer: ______________

  8. Training error falls but held out error climbs as you add flexibility. What is happening?

    • Perfect learning
    • Underfitting
    • Overfitting

  9. Two models err: one steady and stiff, one swinging with samples. Which pairing is right?

    • Both are bias
    • Stiff is variance, swinging is bias
    • Stiff is bias, swinging is variance

  10. A team pours in ten times more data yet a line on curved data still misses identically. Why no cure?

    • The extra data was too clean
    • The fault is bias, which more data barely moves
    • Variance eats all data
LightMySky · lightmysky.comW1-mt_xznyNxeNNw-s1

Answer key

For grown-ups. Fold this page away before handing over the rest.

Too stiff, too wobbly, or just right W1-mt_xznyNxeNNw-s1

  1. Bias · Steady misses from a too stiff model mean bias.
  2. -0.1 · 0.4 minus 0.5 is minus 0.1, a steady underestimate.
  3. True · Variance is sensitivity to the particular sample seen.
  4. True · Variance measures how much a model's predictions would change if it were trained on a different sample. Large swings from small sample changes are exactly what high variance means.
  5. Averaging many models over resampled data · Bagging averages away sample quirks while keeping the pattern.
  6. Variance, the fit moves with the sample · The failure described is the fitted line moving a lot when one point changes. That is sensitivity to the particular sample, which is variance. The line may be simple, but the stated failure is its instability, not its simplicity.
  7. -0.1 · The average prediction is 0.4 because the model always says 0.4. Bias = 0.4 - 0.5 = -0.1. The negative sign means it underestimates every time, which is bias: a steady miss that more of the same kind of data will not fix.
  8. Overfitting · The split, falling training with rising held out, marks overfitting.
  9. Stiff is bias, swinging is variance · Steady means stiffness, swinging means sample hunger.
  10. The fault is bias, which more data barely moves · Stiffness misses stay steady; data calms swings, not stiffness.
Worksheet · LightMySky