Theory test: Overfitting and regularisation

6 questions. Pass with 5 correct.

1. 01 What is the validation set for?

   Fitting the model's weightsChoosing hyperparameters and comparing modelsReporting the final, unbiased resultReplacing the test set when it's too small
2. 02 Training accuracy 99%, test accuracy 71%. The most likely diagnosis is:

   High bias (underfitting)High variance (overfitting)Too much regularisationThe test set is too large
3. 03 Collecting much more training data mainly reduces:

   BiasVarianceThe irreducible noiseAll three equally
4. 04 Which penalty tends to set some weights exactly to zero?

   L2 (ridge)L1 (lasso)Both equallyNeither
5. 05 Increasing the ridge strength $\lambda$ typically:

   Increases variance, decreases biasIncreases bias, decreases varianceLeaves bias and variance unchangedMakes the model more flexible
6. 06 You're forecasting daily sales. How should you split the data?

   Random shuffle, then splitChronologically: train on earlier days, test on later daysUse all data for training and testingSplit by day of the week