AI-900 Practice Test 37 — 15 Questions

Question 1

A data scientist trains a regression model to predict house prices using features like bedrooms, square footage, and location. The model achieves an R-squared of 0.95 on the test set. However, when deployed to predict prices in a new city with different property characteristics, the predictions are very inaccurate. Which concept best explains this poor performance?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Overfitting. The model achieved an R-squared of 0.95 on the test set, indicating excellent performance on data from the same distribution. However, when deployed to a new city with different property characteristics, the predictions were very inaccurate. This is a classic symptom of overfitting, where the model has learned noise and patterns specific to the training data (e.g., city-specific price trends) that do not generalize to unseen data from a different distribution.

Answer

Underfitting

Answer

High bias

Answer

Data drift

Question 2

A quality control team uses computer vision to inspect manufactured parts. They need to detect whether a part has any defects and also identify the type of defect (e.g., scratch, crack, dent) from an image. Which Azure Computer Vision capability should they use?

Accepted Answer

B: Object detection. Object detection is the correct capability because it not only identifies the presence of defects in an image but also localizes each defect with a bounding box and classifies it into specific types (e.g., scratch, crack, dent). This meets both requirements: detecting whether a part has defects and identifying the type of each defect.

Answer

A: Image classification

Answer

C: Semantic segmentation

Answer

D: Optical character recognition (OCR)

Question 3

A data scientist trains a binary classification model to predict loan defaults. The dataset contains 98% non-default cases and only 2% default cases. The model predicts 'non-default' for every instance, achieving 98% accuracy on the test set. Which metric would best reveal that the model fails to identify any actual defaults?

Accepted Answer

Recall for the default class. Recall for the default class measures the proportion of actual default cases that the model correctly identifies. With the model predicting 'non-default' for every instance, recall for the default class is 0%, because it fails to capture any true positives. This directly reveals the model's inability to detect any actual defaults, despite the high overall accuracy.

Answer

Precision for the default class

Answer

F1 score for the default class

Answer

Accuracy

Question 4

A news agency publishes hundreds of articles daily. They want to automatically determine the main topics discussed in each article, such as 'politics', 'economy', or 'sports', without manually tagging them. The agency has no labeled training data. Which built-in Azure AI Language feature should they use?

Accepted Answer

Key phrase extraction. Key phrase extraction is the correct choice because it automatically identifies the main topics or themes in a document without requiring any labeled training data. The news agency can use this built-in Azure AI Language feature to extract key phrases like 'politics', 'economy', or 'sports' from each article, enabling automatic topic categorization without manual tagging.

Answer

Sentiment analysis

Answer

Named entity recognition

Answer

Language detection

Question 5

What is 'batch inference' vs 'real-time inference' in Azure Machine Learning?

Accepted Answer

Real-time processes individual requests immediately; batch processes large datasets at scheduled intervals. Batch inference processes large datasets asynchronously at scheduled intervals, making it suitable for offline or periodic predictions, while real-time inference handles individual requests immediately with low latency for interactive applications. Azure Machine Learning supports both: real-time endpoints for synchronous scoring and batch endpoints for asynchronous, high-throughput processing.

Answer

Batch inference is more accurate; real-time is faster but less accurate

Answer

Batch requires GPU compute; real-time uses CPU only

Answer

Real-time inference is only available in Azure; batch works on-premises too

Question 6

What is 'model distillation' and why might you distill a large model to a small one?

Accepted Answer

Training a smaller model to mimic a larger model's behaviour for efficient deployment. Model distillation is a technique where a smaller 'student' model is trained to replicate the behavior of a larger 'teacher' model. This is done by using the teacher's softmax outputs (logits) as training targets, allowing the student to achieve similar accuracy with far fewer parameters, making it suitable for resource-constrained environments like edge devices or real-time inference.

Answer

Extracting the essential ideas from a model's outputs into a written summary

Answer

Removing duplicate or redundant parameters from a trained model

Answer

Concentrating training data into fewer, higher-quality examples

Question 7

A developer uses Azure OpenAI Service to generate product reviews for an e-commerce site. The developer notices that the model often repeats the same phrases within the same review, making the output sound unnatural. Which parameter should the developer adjust to reduce this repetition?

Accepted Answer

Frequency_penalty. The frequency_penalty parameter reduces the likelihood of the model repeating the same phrases by penalizing tokens that have already appeared in the generated text. A higher frequency_penalty value (e.g., 0.5 to 1.0) discourages the model from reusing the same words or phrases, making the output more diverse and natural. This directly addresses the issue of repetitive phrasing in product reviews.

Answer

Temperature

Answer

Top_p

Answer

Max_tokens

Question 8

What is 'tool calling' (function calling) in Azure OpenAI?

Accepted Answer

A feature allowing models to specify structured calls to external functions for real-world actions. Tool calling (function calling) in Azure OpenAI is a feature that allows the model to output structured JSON requests to invoke external functions or APIs, enabling it to perform real-world actions like querying databases or sending emails. This bridges the gap between the model's static knowledge and dynamic, up-to-date data or services.

Answer

The Azure OpenAI API endpoint URL used to call the model

Answer

Calling Azure support when the AI model returns incorrect results

Answer

A billing mechanism for counting API function calls per minute

Question 9

What is the Azure AI Vision Image Analysis 4.0's 'Florence' foundation model capable of?

Accepted Answer

Advanced image understanding including detailed captions, dense captioning, and multimodal embeddings. The Florence foundation model in Azure AI Vision Image Analysis 4.0 is a multimodal model designed for advanced image understanding. It can generate detailed image captions, produce dense captions (describing multiple regions within an image), and create multimodal embeddings that align visual and textual representations for tasks like image search and similarity.

Answer

Only detecting faces in images

Answer

Only processing medical imaging for diagnostic purposes

Answer

Converting images into 3D models

Question 10

What is GitHub Copilot and how does it use AI?

Accepted Answer

An AI-powered code assistant that generates code completions and suggestions in IDEs using LLMs. GitHub Copilot is an AI-powered code assistant developed by GitHub and OpenAI. It uses large language models (LLMs), specifically a version of OpenAI's Codex model, to analyze the context of the code a developer is writing in an IDE (like VS Code) and generate real-time code completions, suggestions, and even entire functions. This directly aligns with generative AI workloads on Azure, as Copilot leverages generative AI to produce new code content based on natural language prompts or existing code patterns.

Answer

An automated GitHub Actions workflow for running CI/CD pipelines

Answer

A bot that automatically reviews and merges GitHub pull requests

Answer

A GitHub feature for visualizing code repository history

Question 11

What is 'fine-tuning' a language model and when should you use it instead of prompt engineering?

Accepted Answer

Further training a model on domain-specific data to change its behaviour permanently for a task. Fine-tuning is the process of taking a pre-trained language model and further training it on a domain-specific dataset to adapt its behavior permanently for a particular task. This is used instead of prompt engineering when the task requires consistent, specialized outputs that cannot be reliably achieved through prompt instructions alone, such as classifying medical records or generating legal documents.

Answer

Fine-tuning repairs errors in a model's base training data

Answer

Adjusting the model's temperature setting to produce more consistent outputs

Answer

Selecting which pre-trained model from the Azure model catalogue best suits your task

Question 12

A company is developing a chatbot that can both answer customer questions in natural language and create images on demand (e.g., 'Generate a picture of a product prototype'). Which combination of Azure generative AI models should they integrate?

Accepted Answer

A. GPT-4 for text and DALL-E for images. GPT-4 is a generative AI model optimized for natural language understanding and generation, making it ideal for answering customer questions in a conversational manner. DALL-E is a generative AI model specifically designed to create images from textual descriptions, enabling the chatbot to generate product prototypes on demand. Together, they cover both text and image generation requirements.

Answer

B. GPT-3 for text and Custom Vision for images

Answer

C. BERT for text and OCR for images

Answer

D. Language Understanding (LUIS) and Face API

Question 13

What is 'mixture of experts' (MoE) architecture and how does it relate to efficient LLMs?

Accepted Answer

An architecture with many specialised sub-networks that only activates a few per token — enabling efficient large models. Mixture of Experts (MoE) architecture splits the model into multiple specialized sub-networks (experts) and uses a gating mechanism to activate only a small subset of experts per input token. This allows the model to have a very large total parameter count while keeping the computational cost per token low, making it highly efficient for scaling large language models (LLMs) without proportionally increasing inference cost.

Answer

A training approach using multiple human experts to annotate data for different domains

Answer

Combining predictions from multiple separately trained AI models at inference time

Answer

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

Question 14

What is the purpose of Azure AI Content Safety in the context of generative AI deployments?

Accepted Answer

To detect and filter harmful content in AI prompts and responses. Azure AI Content Safety is a service designed to detect and filter harmful content, such as hate speech, violence, self-harm, and sexually explicit material, in both user prompts and AI-generated responses. In generative AI deployments, this ensures that the model's outputs comply with safety policies and regulatory requirements, preventing the dissemination of offensive or dangerous content.

Answer

To compress generated content for faster delivery

Answer

To measure the quality and accuracy of AI-generated responses

Answer

To ensure AI content is written in the correct language

Question 15

A healthcare research organization uses an AI system to analyze patient medical records for pattern discovery. The organization must ensure that the AI system does not expose individual patient identities when reporting results. Which Microsoft responsible AI principle is most directly relevant?

Accepted Answer

Privacy and security. The scenario requires that individual patient identities are not exposed when reporting results from an AI system analyzing medical records. This directly aligns with the Privacy and security principle, which mandates that AI systems protect personal data through mechanisms such as de-identification, differential privacy, and access controls. The organization must ensure that pattern discovery outputs cannot be reverse-engineered to re-identify specific patients, making privacy the core concern.

Answer

Fairness

Answer

Reliability and safety

Answer

Inclusiveness