AI-900 Practice Question: Describe features of generative AI workloads 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 training approach using multiple human experts to annotate data for different domains

  Why it's wrong here

  This confuses data labeling with model architecture. Mixture-of-Experts does not involve people annotating training data by domain; rather, a learned gating network routes each input token to a sparse subset of trainable 'expert' components. Those experts are not pre-assigned to human-defined domains—their specialization emerges automatically from the data and task during end-to-end training. Human domain experts may label data for supervised learning, but that is unrelated to how MoE layers compute.

• ✓

  An architecture with many specialised sub-networks that only activates a few per token — enabling efficient large models

  Why this is correct

  Correct: MoE is a model architecture in which each token is processed by only a small subset of many specialized neural-network modules, selected by a learned router. Since only the top-k experts are activated per forward pass, models can have billions more parameters while keeping per-inference FLOPs comparable to a much smaller dense model. This sparse activation gives large model capacity and improved performance without the proportional compute cost that a dense model of the same total parameter count would require.

• ✗

  Combining predictions from multiple separately trained AI models at inference time

  Why it's wrong here

  Combining outputs from multiple independently trained models is ensemble learning (for example, bagging or boosting), not Mixture-of-Experts. An MoE model is a single network whose expert sub-networks are trained jointly and share a common layer stack; at inference, a router sends each token to a few experts whose outputs are combined inside the model, rather than averaging full predictions from separate models. Ensembles also do not route inputs to specialized components—they aggregate complete model decisions.

• ✗

  A training technique where multiple ML experts review and validate model outputs

  Why it's wrong here

  This describes a human-in-the-loop quality-assurance workflow, not Mixture-of-Experts. In MoE, the 'experts' are neural-network sub-modules (typically feed-forward layers) whose parameters are learned, and a router chooses a few per token; no human reviews or validates model outputs during training or inference. While human feedback can be used for fine-tuning or evaluation, that is a separate process and is not the defining architectural mechanism of MoE.

Go deeper

Related to this question