AI0-001 Machine Learning and Deep Learning Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Remove all pixels with zero variance across the dataset

  Why it's wrong here

  Constant pixels carry no discriminative information, but removing them is a minor cleanup, not a step that improves class separation or generalization for an SVM. It does not address feature scaling or margin tuning, and digit images rarely have many truly constant pixels. On its own this action leaves the dominant scaling and regularization issues untouched, so it is not one of the two most appropriate steps here.

• ✗

  Increase the number of support vectors by loosening the margin

  Why it's wrong here

  The number of support vectors is an outcome of training, not a preprocessing knob. Deliberately enlarging it by loosening the margin usually means accepting more misclassifications and a less discriminative boundary. It does not make the linear kernel separate digit classes better and can degrade generalization. Treating support vector count as a tunable goal reflects a misunderstanding of how the margin and C interact.

• ✗

  Apply a nonlinear kernel such as RBF while claiming to keep the linear kernel

  Why it's wrong here

  Switching to an RBF kernel changes the model class and contradicts the stated goal of helping the linear kernel. Kernel choice is a modeling decision, not preprocessing, and the RBF kernel introduces its own gamma parameter that must be tuned. Selecting it does not prepare the linear SVM for better separation; it replaces the linear SVM entirely, so it is not an appropriate preprocessing step in this scenario.

• ✓

  Tune the C regularization parameter with cross-validation

  Why this is correct

  C controls the trade-off between maximizing the margin and penalizing misclassification. Too large a C produces a narrow margin that fits noise and overfits; too small a C underfits. Because the appropriate value depends on the dataset, tuning C with cross-validation on the scaled features selects a model that generalizes better. This is a core step in preparing an SVM for digit classification and directly targets generalization.

• ✓

  Scale each pixel feature to a common range such as 0 to 1

  Why this is correct

  SVMs rely on distances and dot products, so features with larger numeric ranges dominate the margin computation. Raw pixel intensities may vary in scale across samples and channels, so normalizing them to 0 to 1 or standardizing to zero mean and unit variance puts all features on equal footing. This improves convergence and margin optimization, and it is a standard prerequisite before fitting an SVM with a linear kernel.