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.

• ✓

  Upload tagged images, Train, Evaluate, Publish

  Why this is correct

  This is the correct order for Azure Custom Vision. You must first upload and tag images so the model has labeled examples to learn from. Training then creates an iteration by finding patterns in those tagged images. Evaluation follows training and uses a held-out test set to compute metrics like precision, recall, and mean average precision. Finally, publishing exposes the chosen iteration as a prediction endpoint, which is required before the model can be used by applications.

• ✗

  Train, Upload tagged images, Evaluate, Publish

  Why it's wrong here

  Training cannot be the first step because Azure Custom Vision requires a dataset of tagged images to create any model iteration. Without uploaded and tagged images, there is no training data to split into training and validation sets, and the training job has no input. The service will reject the attempt and there is no model to evaluate or publish. Tagged images must be added to the project before any training run can start.

• ✗

  Upload tagged images, Evaluate, Train, Publish

  Why it's wrong here

  Evaluation cannot occur before training because evaluation measures how well a trained model's predictions match the tagged test images. If no model iteration exists, there are no predictions to compare against ground truth labels, and metrics such as precision, recall, and mean average precision cannot be computed. Azure Custom Vision only generates performance metrics as part of the training process on a withheld portion of the tagged images. Therefore, training must happen before evaluation, not after.

• ✗

  Upload tagged images, Train, Publish, Evaluate

  Why it's wrong here

  Publishing before evaluation is risky and incorrect for building a reliable custom vision model. While Azure Custom Vision technically allows you to publish an iteration to a prediction endpoint at any time, skipping evaluation means you have not examined the model's precision, recall, or accuracy on a test set. This can lead to exposing an underperforming or biased model to production traffic. The required quality gate is to evaluate the iteration first, then publish to the endpoint only after confirming the metrics meet acceptance criteria.

Go deeper

Related to this question