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.

• ✗

  Training multiple different models and comparing their performance

  Why it's wrong here

  Comparing multiple different models — for example, training a decision tree, logistic regression, and a neural network, and then comparing accuracy — is model selection or benchmarking, not cross-validation itself. Cross-validation is a specific evaluation protocol that repeatedly splits a single dataset to estimate the generalization performance of one model. While cross-validation is often used within a model comparison workflow, the technique itself does not involve training multiple algorithms; it rotates training and validation data for a given model to obtain reliable performance metrics.

• ✓

  Repeatedly training and evaluating the model on different data splits for reliable performance estimates

  Why this is correct

  In k-fold cross-validation, the data is randomly partitioned into k equal-sized folds; the model is trained on k−1 folds and evaluated on the remaining fold, and this process is repeated k times so each example serves as validation data exactly once. The evaluation results from all folds are averaged to produce a more stable and less biased estimate of model performance than a single train/test split. This repeated resampling reduces the variance of the performance estimate and makes better use of limited labeled data.

• ✗

  Checking if a model works correctly by running it backward

  Why it's wrong here

  Running a model backward is not a validation technique; in deep learning the backward pass is backpropagation, which computes gradients for weight updates, not a way to measure predictive quality. Cross-validation is a forward-only resampling procedure: data is split into train and test folds, the model is trained in the normal direction, and the predictions are compared with ground truth on the held-out fold. There is no mechanism by which reverse execution could expose generalization error.

• ✗

  Training a model on two different datasets simultaneously

  Why it's wrong here

  Training concurrently on two different datasets, such as combining a primary dataset with an auxiliary one for multi-task learning, is not cross-validation. Cross-validation requires a single dataset to be partitioned into complementary training and validation folds, with the model repeatedly retrained and evaluated on rotated subsets. Simultaneous training on two datasets simply fits one model on more data; it produces no held-out evaluation and therefore cannot estimate generalization performance.

Go deeper

Related to this question