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.

• ✗

  Collecting more labelled data from external sources to supplement training

  Why it's wrong here

  Collecting more labelled data from external sources is data acquisition, which adds entirely new, independent instances to the training set. Data augmentation instead generates synthetic variants from the already-collected examples, often preserving their original labels while applying transformations like flips or synonym replacement. External collection costs time and annotation effort and may introduce distribution shift; augmentation multiplies the effective size of your existing pool without new external information.

• ✓

  Creating synthetic training variants (flips, rotations, synonyms) to expand small datasets

  Why this is correct

  Creating synthetic training variants — such as image flips/rotations or text synonym substitution — is the core of data augmentation. These transformations generate new, plausible examples from existing ones, multiplying effective dataset size while teaching the model invariances (e.g., a cat is still a cat after horizontal flip) and reducing overfitting. This is exactly the technique used to expand small datasets in computer vision and NLP.

• ✗

  Increasing the number of compute nodes to process large training datasets faster

  Why it's wrong here

  Increasing compute nodes scales the distributed training infrastructure, which reduces wall-clock time by parallelizing gradient computation and batch processing. However, this does not alter the training dataset itself — data augmentation is a data-space technique that synthesizes new examples from existing samples. In short, compute scaling addresses throughput, not data diversity or the root cause of a small, overfit-prone dataset.

• ✗

  Adding more evaluation metrics to get a richer view of model performance

  Why it's wrong here

  Adding evaluation metrics such as precision, recall, or F1-score provides a richer post-training view of model performance, but it does not touch the training data. Data augmentation, by contrast, operates before or during training to expand the dataset with transformed copies (e.g., rotated images, paraphrased text). Metrics measure outcomes; augmentation shapes the learning signal. These are fundamentally different stages of the ML workflow.

Go deeper

Related to this question