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.

• ✗

  Adjusting the physical voltage supplied to GPU hardware during training

  Why it's wrong here

  Changing GPU voltage (e.g., undervolting or overclocking) is an infrastructure/hardware-level action that affects power consumption, heat, and clock speed, not the algorithmic configuration of the model. Hyperparameters are parameters of the training procedure itself, such as learning rate schedule, batch size, or momentum, which directly control how gradients update the weights. Hardware adjustments may alter training speed or stability but don't change the model's learning dynamics or final configuration space. Thus it's completely orthogonal to hyperparameter tuning.

• ✓

  Searching for the optimal algorithm settings (learning rate, batch size) that maximise model performance

  Why this is correct

  Hyperparameter tuning is the process of systematically searching a defined configuration space for the combination of algorithm settings, such as learning rate, batch size, number of hidden units, or regularization coefficient, that minimizes the loss on validation data. In Azure Machine Learning, HyperDrive performs this search using random sampling, grid sampling, or Bayesian optimization, and can use early termination policies to prune poor-performing runs. The goal is to maximize a specified primary metric, often accuracy or F1 score, on the validation set.

• ✗

  Training the model to predict hyper-specific rare events in the data

  Why it's wrong here

  Rare event prediction involves handling class imbalance, anomaly detection, or oversampling techniques, which are data-level or model-architecture decisions. Hyperparameter tuning, however, is the systematic search over scalar settings like learning rate, batch size, regularization strength that govern the training algorithm's behavior. It doesn't change the target distribution or task definition; it optimizes how the chosen algorithm learns from a fixed dataset.

• ✗

  Compressing model weights to reduce inference latency

  Why it's wrong here

  Model compression techniques like pruning, quantization, and knowledge distillation are applied after training to shrink the model and speed up inference. They don't affect the training process or search over algorithm configurations. Hyperparameter tuning, in contrast, occurs before/during training by varying learning rate, batch size, etc., to optimize the loss landscape.

Go deeper

Related to this question