AI-900 Practice Question: Describe fundamental principles of machine learning on Azure

Answer analysis

Option-by-option breakdown

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

• ✗

  Choosing between model accuracy and computational cost

  Why it's wrong here

  Choosing between model accuracy and computational cost is a practical engineering decision about resources, hardware, and operational budgets, not a statistical property of estimators. You can often improve accuracy by adding parameters or ensembling, but that changes bias and variance in ways unrelated to an accuracy-cost curve. The bias-variance tradeoff describes how model error changes with complexity, irrespective of price, energy, or run-time costs.

• ✓

  The balance between model simplicity (underfitting) and model complexity (overfitting)

  Why this is correct

  The bias-variance tradeoff decomposes total expected prediction error into bias, variance, and irreducible noise, where high bias signals underfitting from a model too simple to capture patterns, and high variance signals overfitting from a model too sensitive to training noise. As model capacity increases, bias falls but variance rises, so the optimal model complexity minimizes expected error, not training accuracy. This central concept drives practices like train/validation splits and regularization (e.g., L2 penalty) to navigate the tradeoff.

• ✗

  Deciding whether to use biased training data or unbiased test data

  Why it's wrong here

  This statement conflates data bias/fairness with the statistical bias-variance tradeoff. In ML, bias-variance decomposition concerns model error due to simplifying assumptions (bias) and sensitivity to training sample fluctuations (variance). Biased training data introducing demographic skew is a dataset/fairness problem, while unbiased test data concerns representative evaluation, not model complexity or generalization error sources.

• ✗

  The tradeoff between training speed and model size

  Why it's wrong here

  Training speed versus model size is a computational resource tradeoff involving memory footprint, throughput, and infrastructure cost, not generalization error. Although a larger model may have lower bias or higher variance, the speed/size tension addresses engineering limits such as latency and deployment constraints. Bias-variance is specifically about how model capacity affects error on unseen data, independent of wall-clock training time.

Go deeper

Related to this question