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.

• ✗

  A statistical method for updating model confidence as new training data arrives

  Why it's wrong here

  This describes Bayesian inference or online Bayesian model updating, not Bayesian optimization. Bayesian optimization is a global optimization strategy for expensive black-box functions, typically the validation performance as a function of hyperparameters. It uses a probabilistic surrogate model, such as a Gaussian process, to estimate where the optimum is likely to lie, and then chooses the next configuration to evaluate based on an acquisition function. It does not update 'confidence' in model predictions; rather, it maintains a belief about the location of the best hyperparameters.

• ✓

  A smart hyperparameter search that uses past trial results to select promising configurations

  Why this is correct

  Bayesian optimization is indeed a smart hyperparameter search technique: it builds a probabilistic surrogate model, such as a Gaussian process, from all previously evaluated hyperparameter configurations and the resulting performance scores. It then uses an acquisition function, like expected improvement, to balance exploration and exploitation and select the next candidate configuration that is most likely to improve upon the current best result. This approach minimizes the number of expensive training runs needed, making it highly effective when each hyperparameter evaluation is costly, such as when training large neural networks.

• ✗

  An automatic method for adjusting learning rate during training based on gradient information

  Why it's wrong here

  This is a description of adaptive gradient optimizers like Adam, Adagrad, or RMSprop, which modify the learning rate for each parameter based on past gradient magnitudes. Those are parameter update rules used within a single training run, whereas Bayesian optimization operates at a higher level: it searches the hyperparameter space across multiple training runs. Bayesian optimization does not use gradient information from the loss surface; it treats the relationship between hyperparameters and model performance as a black-box function and builds a probabilistic surrogate model instead. Learning rate is just one of many hyperparameters that Bayesian optimization might tune, but the method itself is not adjusting it during training.

• ✗

  A probabilistic approach to labelling uncertain training examples

  Why it's wrong here

  This option confuses Bayesian optimization with active learning or uncertainty-based sample selection. Active learning uses a model's probability outputs to identify unlabeled examples that are most uncertain, so they can be labeled by a human annotator to improve the model efficiently. Bayesian optimization, in contrast, does not label data or select training examples; its objective is to locate the optimal hyperparameter configuration by probabilistically modeling the objective function over the hyperparameter space. While it is a probabilistic approach, it is applied to search, not to labelling tasks.

Go deeper

Related to this question