AI-900 Flashcards — Free Microsoft Azure AI Fundamentals AI-900 Study Cards

Fairness

Fairness ensures that AI systems treat all people and groups equitably, avoiding bias and discrimination by design. Accountability means taking responsibility for the system's outcomes, but the question specifically asks about preventing discrimination. Inclusiveness empowers everyone, and Reliability & Safety ensures systems work correctly under all conditions, but Fairness directly addresses bias.

Regression

Regression is used when the target variable is a continuous numeric value, such as price. Classification predicts categories (e.g., 'expensive' or 'cheap'), Clustering groups similar data without labels, and Anomaly Detection identifies unusual data points. Since house price is a continuous number, regression is correct.

Image classification

Image classification assigns a label (or multiple labels) to an entire image. Object detection identifies and locates multiple objects within an image with bounding boxes. OCR extracts text from images. Semantic segmentation classifies each pixel. Since the requirement is to give a single label per image, image classification is correct.

Text Analytics

Text Analytics provides Named Entity Recognition (NER) to identify and extract entities like people, dates, and quantities from text. LUIS is for intent and entity extraction from conversational utterances. Translator Text translates between languages. Speech handles audio transcription. The scenario requires extracting structured information from free-text notes, so Text Analytics is the correct service.

A. Azure OpenAI Service

Azure OpenAI Service provides access to generative language models like GPT-4 that can produce coherent and varied text from simple prompts, making it ideal for generating product descriptions at scale. Other options: Azure Computer Vision is for image analysis, LUIS is for language understanding, and Azure Machine Learning is a broader platform that would require building a custom model.

Conversational Language Understanding (CLU)

Conversational Language Understanding (CLU) is designed to extract intents and entities from conversational utterances. Custom Question Answering is for FAQ bots, Custom Text Classification assigns a single label to a document, and Named Entity Recognition only extracts entities without intents.

System message

A system message is a message at the start of a conversation that sets the context and behavior of the AI. By explicitly instructing the model not to reveal system instructions, ignore certain types of user inputs, or stay on topic, the developer can mitigate prompt injection attacks. Few-shot prompting uses examples but is not specifically a guard against injection. Temperature controls creativity. Content filtering blocks harmful content but does not prevent the model from revealing system instructions. System messages are the primary tool for defining constraints on model behavior.

Use Azure OpenAI On Your Data to connect to a product database so the model retrieves and references only the provided specifications.

The correct approach is to use Azure OpenAI On Your Data, which allows the model to ground its responses by retrieving relevant information from a provided data source (e.g., a product database or documents). This significantly reduces hallucinations by anchoring the output to the supplied content. Increasing the frequency penalty (A) reduces repetition but does not prevent hallucination. Decreasing the temperature (B) makes the model more deterministic and less creative, but it can still generate unsupported facts. Enabling content filtering (D) blocks harmful or offensive content but does not ensure factual accuracy based on the given specs.

Temperature

The Temperature parameter controls the randomness of output. Setting Temperature to 0 (the lowest value) makes the model highly deterministic, always choosing the most likely next token. This reduces creativity and speculation, which is ideal for factual, grounded responses. Frequency penalty and Presence penalty affect word repetition and topic diversity but do not directly control randomness. Top_p (nucleus sampling) also influences randomness, but Temperature is the primary parameter for controlling determinism and is commonly set to 0 for factual tasks.

Temperature

The temperature parameter controls randomness in generated text. Higher values (e.g., 0.8) make output more diverse and creative, while lower values (e.g., 0.2) make output more deterministic and repetitive. Increasing temperature directly addresses the lack of variety. Top_p can also influence diversity but is typically used in conjunction with temperature. Max_tokens limits length and repetition_penalty reduces repeated content but does not primarily increase diversity.

Temperature

Temperature controls randomness in generation. Lowering temperature (closer to 0) makes the model more deterministic and focused on the most likely tokens. Max tokens limits the length of the output. Top-p (nucleus sampling) also influences randomness but temperature is the primary parameter for controlling creativity. Frequency penalty discourages repetition of the same tokens.

Few-shot learning

Few-shot learning is a technique where a small number of labeled examples are included in the prompt to guide the model's output without updating its weights. This helps the model understand the desired format and content for classification tasks. Fine-tuning would require retraining the model on a large dataset, prompt engineering is a broader concept that may include few-shot, and retrieval-augmented generation involves fetching external data, not just examples in the prompt.

Frequency penalty

The frequency_penalty parameter penalizes tokens that have already appeared in the generated text, thereby reducing repetitive patterns. Lowering temperature (Option A) reduces randomness but does not directly target repetition. The max_tokens parameter (Option B) controls the length of the output, not repetition. The top_p parameter (Option D) controls nucleus sampling, which can affect diversity but is less direct than frequency_penalty for reducing repetition.

Few-shot learning

The developer is using few-shot learning, where a small number of examples are provided in the prompt to guide the model's output style and format. Zero-shot learning would not include any examples, and fine-tuning involves retraining the model on a custom dataset. Temperature scaling controls creativity, not style imitation.

Optical Character Recognition (OCR)

Optical Character Recognition (OCR) is the Azure Computer Vision capability that extracts printed or handwritten text from images, making it ideal for reading license plates from camera feeds.

Fairness

The Fairness principle requires that AI systems treat all people and groups equitably, avoiding biases that lead to discrimination. The disparate flagging rates indicate a potential bias, making fairness the most relevant principle.

Content Filtering

Azure OpenAI Service includes built-in content filtering that can be configured to block hate, violence, sexual, and self-harm content. This is the primary mechanism to prevent offensive output. Prompt engineering is a technique, not a built-in feature. Fine-tuning customizes the model but does not filter content. Token limit controls the maximum length of the output.

Implementing Retrieval Augmented Generation (RAG) with their catalog.

Retrieval Augmented Generation (RAG) combines a retrieval system that pulls relevant documents (e.g., from the product catalog) with a generative model. This allows the model to ground its responses in the provided documents, ensuring accuracy and relevance. Fine-tuning (A) could work but requires retraining for catalog updates. Few-shot learning (B) is not scalable for large catalogs. Adjusting temperature (D) controls randomness, not factual grounding.

A) Fairness

The responsible AI principle of Fairness ensures that AI systems treat all people fairly and avoid biases that could lead to discrimination. Inclusiveness focuses on empowering everyone, but fairness specifically targets discrimination. Reliability & Safety ensures systems perform reliably, and Transparency ensures users understand how decisions are made. For this scenario of avoiding discrimination, Fairness is the most relevant principle.

Custom Text Classification

Custom text classification is the correct choice because it allows you to train a model on your own labeled data (support tickets and their categories) to classify new text into custom-defined categories. Prebuilt Text Analytics offers sentiment analysis, key phrase extraction, and entity recognition, but not custom categorization. Language Understanding (LUIS) is for conversational intent and entity extraction, not for static document classification. Translator performs language translation, which is unrelated to categorization.

Principal Component Analysis (PCA)

Principal Component Analysis (PCA) is an unsupervised dimensionality reduction technique that transforms the data into a lower-dimensional space while preserving the maximum variance. Feature scaling standardizes the range of features but does not reduce dimensionality. Regularization (e.g., Lasso) can reduce feature impact but not count, and cross-validation is used for model evaluation, not feature reduction.

Key phrase extraction

Key phrase extraction identifies the main talking points in text, such as 'battery life' or 'customer service'. Sentiment analysis detects emotional tone, language detection identifies the language, and entity recognition extracts known entities (e.g., people, places) but not generic topics. Therefore, key phrase extraction is the correct choice for identifying frequently mentioned topics from unstructured text.

Regression

Regression is used when the target variable is a continuous numerical value. Predicting a traffic count is a classic regression problem, as opposed to classification (discrete categories) or clustering (grouping without labels).