Microsoft Azure AI Fundamentals AI-900 (AI-900) — Questions 526–600

Why this answer

How to eliminate wrong answers

Option A is wrong because key phrase extraction identifies important words or phrases in text but does not map them to predefined intents or entities, making it unsuitable for understanding user goals like 'cancel order'. Option B is wrong because sentiment analysis determines the emotional tone (positive, negative, neutral) of text, not the user's intent or specific data like order numbers. Option D is wrong because named entity recognition (NER) extracts entities such as people, places, or dates, but it does not classify the overall intent of a user's utterance, which is essential for a chatbot to determine what action to take.

Why this answer

How to eliminate wrong answers

Option A is wrong because Fairness focuses on ensuring AI systems do not discriminate against groups based on attributes like race or gender, not on explaining individual decisions. Option B is wrong because Reliability and Safety concerns the system's ability to perform consistently and without harm under expected conditions, not on providing post-hoc explanations for specific outputs. Option D is wrong because Privacy and Security deals with protecting patient data from unauthorized access or breaches, not with clarifying the rationale behind a model's recommendation.

Why this answer

How to eliminate wrong answers

Option A is wrong because NER does not identify grammatical parts of speech like nouns and verbs; that task is called part-of-speech (POS) tagging, which is a separate NLP feature. Option C is wrong because NER does not generate new names for products; that would be a generative or creative task, not a recognition or classification task. Option D is wrong because NER does not convert names into anonymous placeholders for privacy; that process is known as de-identification or anonymization, which may use NER as a step but is not the core definition of NER.

Why this answer

How to eliminate wrong answers

Option A is wrong because the content filter does not limit API response length for cost control; that is handled by the 'max_tokens' parameter in the API request. Option C is wrong because the content filter is not a spam filter for off-topic messages; it targets harmful content categories, not relevance or topic adherence. Option D is wrong because removing personally identifiable information (PII) is a separate feature, such as Azure AI Language's PII detection or data masking, not the content filter.

Why this answer

How to eliminate wrong answers

Option A is wrong because Fairness focuses on avoiding bias and ensuring equitable treatment across user groups, not on preventing harmful instructions or ensuring age-appropriate safety. Option C is wrong because Privacy and Security concerns data protection and unauthorized access, not the operational safety of the assistant's responses. Option D is wrong because Inclusiveness aims to design for diverse users and accessibility, but does not directly enforce safety constraints on the assistant's output.

Why this answer

How to eliminate wrong answers

Option A is wrong because standardising input features to the same scale is called feature scaling or normalisation, not regularisation; it addresses gradient descent convergence, not overfitting. Option C is wrong because applying government regulations for data privacy is a compliance or governance concern, not a machine learning regularisation technique; it relates to policies like GDPR, not model training. Option D is wrong because converting irregular training data shapes into a uniform format refers to data preprocessing or reshaping, which is unrelated to the penalty-based regularisation that controls model complexity.

Why this answer

How to eliminate wrong answers

Option A is wrong because fine-tuning updates the model's weights using the product catalog, which is expensive, time-consuming, and can lead to catastrophic forgetting or outdated information if the catalog changes; it does not dynamically retrieve the latest data at inference time. Option C is wrong because zero-shot prompting relies solely on the model's pre-existing knowledge, which cannot incorporate the agency's specific product catalog and risks hallucinating incorrect or generic descriptions. Option D is wrong because RLHF optimizes model behavior based on human preferences for helpfulness or safety, but it does not provide a mechanism to inject proprietary or real-time data from a vector database into the model's responses.

Why this answer

How to eliminate wrong answers

Option A is wrong because custom NER does not rename standard entity types; it creates entirely new entity types from scratch using your own labeled data. Option C is wrong because custom NER is not a faster or lighter version; it uses the same machine learning pipeline as pre-built NER, requiring training and inference. Option D is wrong because filtering NER outputs is a post-processing step, not a training process; custom NER involves training a model to recognize new entities, not just selecting which pre-built entities to return.

Why this answer

How to eliminate wrong answers

Option A is wrong because zero-shot object detection does not imply zero latency; latency depends on model architecture, hardware, and optimization, not on the zero-shot capability. Option C is wrong because zero-shot refers to the absence of training examples for a class, not to the number of color channels in the input image; models can process grayscale or color images regardless. Option D is wrong because zero-shot object detection makes no claim about false positive rate; a model can have false positives even in a zero-shot setting, and achieving zero false positives is an unrealistic performance metric.

Why this answer

How to eliminate wrong answers

Option A is wrong because face detection and identification are distinct operations with different purposes and API endpoints; detection uses the 'Detect' operation, while identification uses the 'Identify' operation against a PersonGroup. Option C is wrong because both detection and identification work on still images and video frames; Azure AI Vision supports both modes for each, with no restriction that detection is only for live video or identification only for still images. Option D is wrong because both face detection and identification require a paid (S0) tier of the Face API; the free (F0) tier is limited to a low number of transactions per month and does not support identification at all.

Why this answer

How to eliminate wrong answers

Option A is wrong because Sentiment Analysis determines the emotional tone (positive, negative, neutral) of text, not the presence of sensitive entities like names or SSNs. Option B is wrong because Key Phrase Extraction identifies main talking points or topics in a document, not specific sensitive data elements. Option D is wrong because Language Detection identifies the language in which text is written (e.g., English, Spanish), not any form of entity recognition or redaction.

Why this answer

How to eliminate wrong answers

Option A is wrong because background removal in Azure AI Vision operates on images, not audio; removing background noise from audio is a speech or audio processing task, not a computer vision feature. Option C is wrong because deleting metadata (e.g., EXIF data) is a file management or privacy operation, not a computer vision capability; Azure AI Vision does not remove metadata as part of its image analysis. Option D is wrong because removing blurry or out-of-focus areas is an image enhancement or deblurring task, not the foreground/background segmentation that background removal performs.

Why this answer

How to eliminate wrong answers

Option A (Precision) is wrong because precision measures the proportion of predicted positives that are actually positive (TP/(TP+FP)), which focuses on avoiding false alarms, not on catching all fraud. Option C (Accuracy) is wrong because accuracy is (TP+TN)/(total) = (8+942)/1000 = 0.95, which is misleadingly high due to class imbalance (990 legitimate vs 10 fraudulent) and does not reflect the cost of missing fraud. Option D (Specificity) is wrong because specificity measures the proportion of actual negatives correctly identified (TN/(TN+FP)), which is about correctly classifying legitimate transactions, not about minimizing missed fraudulent transactions.

Why this answer

How to eliminate wrong answers

Option A is wrong because zero-shot learning involves generating output without any examples in the prompt, relying solely on the model's pre-trained knowledge, whereas the developer explicitly provides three examples. Option C is wrong because fine-tuning requires retraining the model on a custom dataset to adjust its weights, which is a more resource-intensive process not used here; the developer is simply crafting a prompt. Option D is wrong because temperature scaling controls the randomness of token selection (higher values increase creativity, lower values make output more deterministic), not the inclusion of examples in the prompt.

Why this answer

How to eliminate wrong answers

Option A is wrong because multi-modal AI is not about processing data in multiple programming languages; that describes polyglot programming or multi-language support, not data modality. Option C is wrong because deploying AI models across multiple Azure regions for global availability is a geo-redundancy or high-availability strategy, not a characteristic of multi-modal AI. Option D is wrong because using multiple AI models in sequence where each processes a different step describes a pipeline or chained architecture, not the simultaneous processing and relating of multiple data types that defines multi-modal AI.

Why this answer

How to eliminate wrong answers

Option A is wrong because Image Classification assigns a single label or category to an entire image (e.g., 'box' or 'package'), not extracting specific text characters. Option C is wrong because Object Detection identifies and locates objects (e.g., boxes, people) within an image using bounding boxes, but it does not read or interpret text content. Option D is wrong because Image Analysis (captioning and tagging) generates descriptive captions or tags about the image's content (e.g., 'a box on a conveyor belt'), not extracting alphanumeric strings.

Why this answer

How to eliminate wrong answers

Option A is wrong because it describes entities, not intents; entities extract specific pieces of information like dates, amounts, or names from user messages. Option C is wrong because it describes predefined bot responses or a knowledge base, which are separate from intents; intents classify user goals, not store answers. Option D is wrong because confidence levels are a property of the model's prediction (e.g., a score for intent classification), not the definition of an intent itself.

Why this answer

How to eliminate wrong answers

Option A is wrong because classification requires labeled data to predict discrete class labels, but this dataset has no labels. Option B is wrong because regression predicts continuous numerical values (e.g., income amount) from labeled data, not groupings. Option D is wrong because reinforcement learning involves an agent learning from rewards and punishments in an environment, not from static unlabeled data.

Why this answer

How to eliminate wrong answers

Option A is wrong because error analysis does not log Python exceptions or training errors; it focuses on model prediction errors on test data, not code-level failures. Option C is wrong because counting total incorrect predictions is a basic aggregate metric (e.g., error rate), not the subgroup-level analysis that error analysis provides. Option D is wrong because error analysis evaluates model predictions, not infrastructure or pipeline run errors; diagnosing failed runs is a separate operational concern.

Why this answer

How to eliminate wrong answers

Option A is wrong because SMOTE does not involve collecting real examples from external sources; it generates synthetic data from the existing minority class. Option C is wrong because that describes random undersampling, not SMOTE, which oversamples the minority class rather than removing majority examples. Option D is wrong because adjusting confidence thresholds is a post-training decision boundary technique, not a data-level method like SMOTE for handling imbalanced classification.

Why this answer

How to eliminate wrong answers

Option B is wrong because fine-tuning the base model on a small set of example descriptions is inefficient and may lead to overfitting or catastrophic forgetting, as fine-tuning requires a larger, diverse dataset and modifies model weights, which is unnecessary when few-shot prompting can achieve the same result. Option C is wrong because increasing the temperature parameter to the maximum value would make the output highly random and creative, which is the opposite of consistently matching a specific brand voice. Option D is wrong because training a new model using Azure Machine Learning is overkill for this task, as it involves building and training a custom model from scratch, which is resource-intensive and not required when the existing Azure OpenAI model can be adapted via prompt engineering.

Why this answer

How to eliminate wrong answers

Option A is wrong because reliability and safety address the system's ability to perform consistently and without harmful errors, not the comprehensibility of its decision-making process. Option C is wrong because fairness concerns avoiding bias and ensuring equitable outcomes across patient groups, which is not the primary issue when doctors cannot understand the system's reasoning. Option D is wrong because privacy and security focus on protecting patient data and preventing unauthorized access, which is unrelated to the interpretability of the AI's recommendations.

Why this answer

How to eliminate wrong answers

Option A (Few-shot learning) is wrong because it provides a few examples in the prompt to guide the model's output style or format, but it does not supply new factual data like current pricing or catalog updates; the model still relies on its pre-existing knowledge. Option B (Fine-tuning) is wrong because it retrains the model on a custom dataset, which is costly, time-consuming, and still results in a static model that cannot reflect real-time changes to the catalog or pricing without repeated retraining. Option D (Prompt engineering) is wrong because it involves crafting the input text to influence the model's response, but it cannot inject new, external data; the model remains limited to its original training cutoff.

Why this answer

How to eliminate wrong answers

Option A is wrong because CLU does not translate chatbot responses; translation is handled by the Azure Translator service, not by CLU. Option C is wrong because converting speech to text is the function of Azure Speech-to-Text (part of Azure Speech Services), not CLU, which works on text input. Option D is wrong because answering FAQ questions automatically is typically done using Azure Cognitive Search with a QnA Maker or Azure AI Language's pre-built question answering feature, not by training a custom CLU model for intent and entity extraction.

Why this answer

How to eliminate wrong answers

Option A is wrong because a scalable cluster for running distributed training jobs across many nodes describes an Azure Machine Learning compute cluster, not a compute instance. Option C is wrong because a virtual machine that automatically scales to run batch predictions describes an Azure Machine Learning inference cluster or a managed online endpoint, not a compute instance. Option D is wrong because a serverless execution environment for ML inference requests describes Azure Machine Learning serverless inference endpoints or Azure Functions, not a compute instance.

Why this answer

How to eliminate wrong answers

Option A is wrong because custom summarization does not involve custom fonts or formatting; it focuses on the content and accuracy of the summary, not visual presentation. Option C is wrong because while you can set a maximum length for summaries (e.g., via parameters like 'maxLength'), custom summarization is about adapting the model to a domain, not just setting a character limit. Option D is wrong because custom summarization is not an automation feature for all documents in a storage account; it requires training a model on labeled data and does not automatically process all documents without explicit configuration.

Why this answer

How to eliminate wrong answers

Option A is wrong because it describes reducing compute resources, not the data efficiency technique of few-shot learning. Option C is wrong because it describes parallel training experiments, which is a resource optimization strategy unrelated to few-shot learning. Option D is wrong because it describes early stopping based on iteration count, which is a training termination heuristic, not a method for achieving accuracy with minimal labeled data.

Why this answer

How to eliminate wrong answers

Option B (Accuracy) is wrong because accuracy is a classification metric that measures the proportion of correct predictions out of total predictions, not applicable to regression tasks predicting continuous values like house prices. Option C (F1 Score) is wrong because F1 Score is a harmonic mean of precision and recall used for evaluating classification models, particularly for imbalanced datasets, and does not measure prediction error magnitude in regression. Option D (Precision) is wrong because precision is a classification metric that measures the proportion of true positive predictions among all positive predictions, irrelevant for regression error analysis.

Why this answer

How to eliminate wrong answers

Option A is wrong because Image Analysis describes the content of an image (objects, scenes, tags) but does not extract text, especially handwritten text. Option C is wrong because Face API is dedicated to detecting, recognizing, and analyzing human faces, not text. Option D is wrong because Custom Vision is used to train custom image classifiers or object detectors on specific visual features, not for general-purpose text extraction from varied handwriting.

Why this answer

How to eliminate wrong answers

Option A is wrong because `temperature` controls the randomness or creativity of the output, not the length; a lower temperature makes the model more deterministic, but does not limit token count. Option C is wrong because `top_p` (nucleus sampling) controls the cumulative probability threshold for token selection, affecting diversity but not the maximum number of tokens generated. Option D is wrong because `frequency_penalty` reduces repetition by penalizing tokens that have already appeared, but it does not impose a hard limit on the length of the generated code.

Why this answer

How to eliminate wrong answers

Option A is wrong because transcribing spoken audio into text is the purpose of Azure AI Speech's speech-to-text feature, not speaker recognition. Option C is wrong because detecting whether audio contains speech or background noise is handled by the voice activity detection (VAD) component, which is a preprocessing step, not a speaker recognition capability. Option D is wrong because improving audio quality by removing background noise is the function of audio enhancement or noise suppression features, such as those in Azure AI Speech's custom audio processing, not speaker recognition.

Why this answer

How to eliminate wrong answers

Option B (Object detection) is wrong because it identifies and locates objects within an image (e.g., cars, pedestrians) but does not extract text characters from those objects. Option C (Image classification) is wrong because it assigns a single label or category to an entire image (e.g., 'car' or 'truck') and cannot read specific alphanumeric sequences like license plates. Option D (Facial recognition) is wrong because it detects and identifies human faces, not vehicle license plates, and is designed for biometric identification rather than text extraction.

Why this answer

How to eliminate wrong answers

Option A is wrong because it describes a learning management system (LMS) for human education, not a machine learning training paradigm. Option C is wrong because while cloud compute may be used, online learning does not require an internet connection; it refers to incremental data processing, not network connectivity. Option D is wrong because it describes interactive or active learning where humans correct the model, which is a different concept from automated incremental weight updates based on new data.

Why this answer

How to eliminate wrong answers

Option A is wrong because object detection identifies and locates multiple objects within an image using bounding boxes, outputting multiple labels and positions, not a single label for the entire image. Option C is wrong because Optical Character Recognition (OCR) extracts text from images, not vehicle types. Option D is wrong because semantic segmentation assigns a class label to every pixel in the image, creating a pixel-level mask rather than a single image-level label.

Why this answer

How to eliminate wrong answers

Option A is wrong because RAG does not skip or accelerate the training process; the underlying LLM remains fully trained, and RAG is a retrieval-augmented inference technique. Option C is wrong because RAG typically increases API costs due to the additional retrieval step (e.g., vector search queries) and does not batch requests for cost reduction. Option D is wrong because RAG is primarily a text-based retrieval mechanism; multimodal capabilities (e.g., image processing) are separate features of models like GPT-4V, not a benefit of RAG.

Why this answer

How to eliminate wrong answers

Option A is wrong because precision measures the proportion of positive predictions that are correct; since the model never predicts positive, precision is undefined or 0, but precision does not directly reveal the failure to find actual cases. Option C is wrong because the F1 score is the harmonic mean of precision and recall; with recall at 0%, the F1 score is also 0, but it does not isolate the failure as clearly as recall does. Option D is wrong because mean absolute error (MAE) is a regression metric used for continuous values, not for binary classification tasks like disease prediction.

Why this answer

How to eliminate wrong answers

Option A is wrong because it describes affordability or pricing models, not the core purpose of AI for accessibility, which is about removing barriers for people with disabilities—not making AI cheap for small businesses. Option C is wrong because it focuses on API documentation and developer experience, which is a general best practice for any service, not the specific goal of using AI to assist individuals with disabilities. Option D is wrong because it addresses low-bandwidth connectivity in developing regions, which is a network infrastructure concern, not the targeted use of AI to aid people with disabilities through speech, vision, or language capabilities.

Why this answer

How to eliminate wrong answers

Option A is wrong because it describes data augmentation via translation, not transfer learning; cross-lingual transfer learning does not require explicit translation of training data—it relies on shared embeddings learned during pre-training. Option C is wrong because it confuses cross-lingual transfer with multimodal learning; multilingual NLP models are specific to text across languages, not cross-domain transfer between NLP and computer vision. Option D is wrong because it refers to model portability between cloud providers, which is a deployment concern unrelated to the linguistic transfer of knowledge within a model.

Why this answer

How to eliminate wrong answers

Option A is wrong because Fairness addresses bias and equitable treatment across user groups, not operational correctness or safety in physical tasks like item picking. Option C is wrong because Privacy and Security concerns data protection and unauthorized access, not the robot's functional accuracy or physical safety in warehouse operations. Option D is wrong because Transparency focuses on explainability and user understanding of AI decisions, not the system's reliable performance under expected environmental conditions.

Why this answer

How to eliminate wrong answers

Option A is wrong because key phrase extraction identifies important words or phrases but does not associate them with sentiment or distinguish between aspects and general topics. Option B is wrong because named entity recognition identifies entities like people, places, or organizations, not product aspects or opinions. Option D is wrong because language detection only identifies the language of the text and provides no information about aspects or sentiment.

Why this answer

How to eliminate wrong answers

Option A is wrong because custom text classification does not involve applying CSS classes to web page text; that is a front-end styling task unrelated to Azure AI Language's NLP capabilities. Option C is wrong because classifying text files by their file type (PDF, Word, TXT) is a file format identification task, not a semantic or content-based classification, and Azure AI Language focuses on analyzing the textual content, not the file extension. Option D is wrong because custom text classification is not a pre-built classifier; it requires you to provide labeled examples to train a model for your own categories, whereas a pre-built classifier for 10 universal topics would be a built-in feature like the pre-configured sentiment analysis or key phrase extraction, not a custom one.

Why this answer

How to eliminate wrong answers

Option A is wrong because guardrails are not physical barriers in data centers; they are software-based safety mechanisms applied to model inputs and outputs. Option C is wrong because rate limiting controls API usage and prevents overuse, but it does not address content safety or responsible AI concerns. Option D is wrong because version control systems manage model updates and rollbacks, not the real-time filtering of harmful or inappropriate outputs.

Why this answer

How to eliminate wrong answers

Option A is wrong because it describes the mathematical operations (e.g., matrix multiplications) performed during model training, which is a computational process, not the infrastructure that runs it. Option C is wrong because it refers to FLOPS (floating-point operations per second), a performance metric for measuring computational throughput, not the managed compute service itself. Option D is wrong because it describes the Azure Pricing Calculator or TCO calculator, which estimates costs but does not execute ML workloads.

Why this answer

How to eliminate wrong answers

Option A is wrong because fairness is about ensuring AI systems do not discriminate against groups or individuals based on attributes like race or gender; the scenario does not mention any bias or disparate impact. Option B is wrong because reliability and safety focus on the system performing consistently and safely under expected conditions; the scenario does not describe failures, errors, or unsafe behavior. Option D is wrong because inclusiveness is about designing AI to empower and engage everyone, including people with disabilities; the scenario does not involve accessibility or representation issues.

Why this answer

How to eliminate wrong answers

Option A is wrong because Mean Absolute Error (MAE) treats all errors equally by taking the absolute value of residuals, so it does not penalize large errors more than small ones. Option C is wrong because R-squared (R²) measures the proportion of variance explained by the model, not the magnitude of prediction errors, and it does not directly penalize large errors. Option D is wrong because Mean Absolute Percentage Error (MAPE) uses absolute percentage errors, which are not squared and thus do not disproportionately penalize large errors; it also has issues with division by zero when actual values are zero.

Why this answer

How to eliminate wrong answers

Option A is wrong because liveness detection does not measure vital signs like heart rate or blood pressure; it relies on visual cues to assess liveness, not biometric health indicators. Option C is wrong because counting people in a live video stream is a separate capability of the Azure Video Indexer or Computer Vision service, not a function of Face liveness detection. Option D is wrong because monitoring whether a person remains present during a video call is a feature of Azure Communication Services or presence detection, not the Face service's liveness detection, which focuses on spoof prevention at the moment of capture.

Why this answer

How to eliminate wrong answers

Option B is wrong because Named Entity Recognition (NER) identifies specific, pre-defined categories like people, organizations, locations, and dates, not general recurring issue topics or multi-word phrases. Option C is wrong because Sentiment Analysis detects the emotional tone (positive, negative, neutral) of text, not the specific subject matter or recurring issues. Option D is wrong because Language Detection identifies the language of the text (e.g., English, Spanish), which is irrelevant to finding common support topics within a single language.

Why this answer

How to eliminate wrong answers

Option B is wrong because reliability and safety focus on ensuring the system performs consistently and without harm, not on explaining decisions. Option C is wrong because fairness addresses bias and equitable treatment across groups, but the scenario does not mention any discriminatory outcomes or biased data. Option D is wrong because privacy and security concern data protection and unauthorized access, not the ability to explain model decisions.

Why this answer

How to eliminate wrong answers

Option A is wrong because unsupervised learning uses unlabeled data and finds hidden patterns or groupings without any correct output provided. Option B is wrong because reinforcement learning learns through trial-and-error interactions with an environment using rewards and penalties, not from pre-labeled input-output pairs. Option D is wrong because transfer learning is a technique that reuses a pre-trained model on a new but related task, not a distinct learning paradigm that uses labeled training data directly.

Why this answer

How to eliminate wrong answers

Option A is wrong because classification predicts discrete class labels (e.g., 'hot' or 'cold'), not a continuous numeric value. Option C is wrong because clustering groups unlabeled data into clusters based on similarity, without predicting a specific numeric output. Option D is wrong because object detection identifies and locates objects within images or video, producing bounding boxes and labels, not a single numeric temperature value.

Why this answer

How to eliminate wrong answers

Option A is wrong because transfer learning and training from scratch do not produce identical results; transfer learning typically converges faster and may achieve higher accuracy with limited data, while training from scratch requires more data and compute to reach comparable performance. Option C is wrong because transfer learning is a machine learning technique involving model weights, not a mechanism for copying model weights between Azure subscriptions; Azure subscriptions are unrelated to the concept. Option D is wrong because transfer learning can be used even when original training data is available; it is chosen to save resources and improve performance, not solely due to data unavailability.

Why this answer

How to eliminate wrong answers

Option A is wrong because deep learning models do not use simple mathematical formulas; they involve complex, non-linear transformations across many layers, making transparency difficult, not easy. Option C is wrong because transparency matters for all AI systems, especially in high-stakes domains like healthcare, finance, and criminal justice, not just consumer products. Option D is wrong because showing training data does not explain how a model processes that data to reach decisions; transparency requires understanding the model's internal logic, not just the data it was trained on.

Why this answer

How to eliminate wrong answers

Option A is wrong because entity linking does not create hyperlinks in a document; it maps textual mentions to knowledge base entries for disambiguation, not for navigation. Option C is wrong because entity linking resolves a single mention to a knowledge base entry, not tracking the same entity across multiple documents over time (that would be coreference resolution or entity resolution). Option D is wrong because entity linking is a standalone disambiguation step, not a mechanism to trigger downstream API calls for data enrichment.

Why this answer

How to eliminate wrong answers

Option A is wrong because feature engineering (creating or transforming input variables) can improve model performance but does not inherently address generalization across different data segments; it may even exacerbate overfitting if features are tailored to the historical segment. Option C is wrong because increasing model complexity (e.g., adding more layers or parameters) typically worsens generalization by increasing the risk of overfitting to the training data, making the model less adaptable to new segments. Option D is wrong because using a larger learning rate can cause the model to converge too quickly to a suboptimal solution or diverge, but it does not directly improve generalization and may harm performance on both historical and new data.

Why this answer

How to eliminate wrong answers

Option A is wrong because sentiment analysis detects emotional tone (positive/negative/neutral) and key phrase extraction identifies general important phrases, neither of which can route emails to departments or extract specific account numbers and order IDs. Option B is wrong because language detection identifies the language of the text and translation converts text between languages, which is irrelevant to routing or extracting customer-specific data. Option D is wrong because named entity recognition (NER) extracts only prebuilt entity types (e.g., person, organization, location) and cannot extract custom fields like account numbers or order IDs, while summarization condenses text and does not perform routing or extraction.

Why this answer

How to eliminate wrong answers

Option A is wrong because fine-tuning requires retraining the model on a curated dataset, which contradicts the requirement to avoid retraining. Option C is wrong because reducing the temperature parameter to zero makes the model deterministic and may reduce creativity but does not inherently prevent harmful stereotypes; it can still generate biased or stereotypical text if the training data contains such patterns. Option D is wrong because increasing the maximum output length only allows the model to generate longer responses; it does not influence the content quality or reduce harmful outputs.

Why this answer

How to eliminate wrong answers

Option A is wrong because MoE does not involve human experts annotating data; it is a neural network architectural pattern, not a data annotation methodology. Option C is wrong because MoE activates different experts within a single model per token, not combining predictions from multiple separately trained models at inference time (that would be ensemble learning). Option D is wrong because MoE is not a training technique where ML experts review outputs; it is a static architectural design with learned routing, not a human-in-the-loop validation process.

Why this answer

How to eliminate wrong answers

Option A is wrong because training large language models from scratch involves massive datasets, specialized hardware, and fine-tuning of model parameters—this is a separate process called pre-training or fine-tuning, not prompt engineering. Option C is wrong because compressing AI models to run on smaller devices refers to techniques like quantization, pruning, or distillation (e.g., using ONNX Runtime or TensorFlow Lite), which are unrelated to designing input prompts. Option D is wrong because fixing bugs in AI software is a software engineering or debugging task (e.g., fixing code errors in model inference pipelines), not a method for crafting inputs to guide model outputs.

Why this answer

How to eliminate wrong answers

Option A is wrong because 'Azure OpenAI on your data' does not involve training or fine-tuning a custom model; it uses an existing Azure OpenAI model (e.g., GPT-4) with your data as a retrieval source. Option C is wrong because the feature does not restrict the model to only your Azure subscription data; it can still access its pre-trained knowledge, but responses are grounded in your connected data sources. Option D is wrong because it is not a billing option; it is a feature that incurs standard token-based charges plus costs for the underlying data storage and search services.

Why this answer

How to eliminate wrong answers

Option B (Object detection) is wrong because it identifies and locates objects (e.g., boxes, vehicles) within an image, not text characters or words. Option C (Image classification) is wrong because it assigns a single label or category to an entire image (e.g., 'shipping label'), but does not extract specific textual details like addresses or weights. Option D (Facial recognition) is wrong because it detects and identifies human faces, which is irrelevant to reading handwritten text on labels.

Why this answer

How to eliminate wrong answers

Option A is wrong because Image Analysis dense captioning generates descriptive captions for images, not real-time spatial tracking of people's movement. Option B is wrong because Facial recognition identifies or verifies individuals by their face, not tracking movement or direction in a physical space. Option D is wrong because Optical Character Recognition (OCR) extracts text from images, not people detection or motion analysis.

Why this answer

How to eliminate wrong answers

Option B (Top_p) is wrong because Top_p (nucleus sampling) controls the cumulative probability threshold for token selection, not the overall randomness; reducing Top_p can also make output more deterministic, but the question specifically asks about decreasing a parameter to reduce randomness, and Temperature is the primary control for randomness. Option C (Frequency_penalty) is wrong because frequency_penalty reduces the likelihood of repeating the same tokens or phrases, which affects diversity but does not directly control the randomness of token selection; it is used to avoid repetitive outputs, not to fix syntax errors caused by high randomness. Option D (Max_tokens) is wrong because max_tokens limits the length of the generated output, not the randomness or determinism of the token choices; it cannot reduce syntax errors caused by overly random sampling.

Why this answer

How to eliminate wrong answers

Option A is wrong because scanning product barcodes to look up inventory information relies on barcode scanning technology, not computer vision-based product recognition; Azure AI Vision product recognition identifies products visually without requiring barcodes. Option C is wrong because generating product descriptions from images for e-commerce listings is a feature of Azure AI Vision's image captioning or content moderation, not the specialized product recognition API for retail. Option D is wrong because detecting counterfeit or damaged products in a manufacturing quality line falls under anomaly detection or custom vision models, not the prebuilt product recognition capability designed for retail shelf analysis.

Why this answer

How to eliminate wrong answers

Option A is wrong because k=10 does not always produce a better model; it provides a more reliable estimate of model performance, but the actual model quality depends on the algorithm and data, not the cross-validation fold count. Option C is wrong because k=5 and k=10 produce different results due to different training/validation splits and variance in estimates; they are not identical. Option D is wrong because k-fold cross-validation does not require more labelled data; it uses the same dataset, just partitioned differently.

Why this answer

How to eliminate wrong answers

Option A is wrong because Azure AI Language is not a programming language; it is a managed cloud service, whereas Microsoft's AI-focused programming languages include tools like ML.NET or Python SDKs. Option C is wrong because while Azure AI Language includes translation capabilities via the Translator service, it is not solely a machine translation service; it offers a broader suite of NLP features beyond translation. Option D is wrong because Azure AI Language is not a language learning application; it is an enterprise-grade NLP service for developers, not an end-user educational tool.

Why this answer

How to eliminate wrong answers

Option A is wrong because precision measures the proportion of predicted positive cases that are actually positive; if the model predicts no positives, precision is undefined (division by zero) or 0/0, which does not clearly reveal the failure to detect fraud. Option C is wrong because the F1 score is the harmonic mean of precision and recall; with recall at 0%, the F1 score will also be 0%, but it does not directly highlight the failure as intuitively as recall alone. Option D is wrong because specificity measures the proportion of actual negative cases (legitimate transactions) correctly identified; the model correctly classifies all legitimate transactions, so specificity would be 100%, masking the failure to detect fraud.

Why this answer

How to eliminate wrong answers

Option A is wrong because temperature controls the randomness of token selection by scaling the logits before softmax, but it does not directly penalize repetition; higher temperature can increase diversity but may also lead to incoherence. Option B is wrong because top_p (nucleus sampling) limits the cumulative probability mass of tokens considered for sampling, which can reduce repetition indirectly but is not designed to specifically penalize repeated concepts. Option D is wrong because max tokens only sets the maximum length of the generated output and has no effect on the model's tendency to repeat concepts within that output.

Why this answer

How to eliminate wrong answers

Option A is wrong because object detection only identifies and locates objects with bounding boxes, not pixel-level delineation, so it cannot separate walls from doors or furniture at the granularity required. Option C is wrong because image classification assigns a single label to the entire image (e.g., 'kitchen'), not per-pixel segmentation of multiple elements. Option D is wrong because OCR extracts text from images, which is irrelevant to identifying walls, doors, windows, or furniture in a room.

Why this answer

How to eliminate wrong answers

Option A is wrong because temperature in generative AI is a model parameter, not a hardware metric; GPU operating temperature is a physical measurement unrelated to model output randomness. Option C is wrong because the time required to generate a response is determined by factors like model size, sequence length, and hardware, not by the temperature parameter. Option D is wrong because temperature does not set a confidence threshold; confidence thresholds are typically handled via top-k or top-p (nucleus) sampling, or by logit filtering, not by temperature scaling.

Why this answer

How to eliminate wrong answers

Option A is wrong because extractive summarization works on text, not images, and abstractive summarization also works on text; image summarization falls under computer vision, not NLP. Option C is wrong because both extractive and abstractive summarization can be applied to documents of any length; the choice depends on the desired output style, not document length. Option D is wrong because accuracy depends on the specific use case and model quality; abstractive summarization can introduce errors or hallucinations, while extractive summarization is often more faithful to the original text.

Why this answer

How to eliminate wrong answers

Option B is wrong because undersampling the majority class discards the vast majority of normal operation data, which can lead to loss of valuable information and reduced model generalization, especially when the majority class is 99.5% of the data. Option C is wrong because using precision as the optimization metric does not directly address the class imbalance at the data level; it is a model evaluation metric that can be used after rebalancing, but it does not change the underlying skewed distribution. Option D is wrong because reducing model complexity does not fix the class imbalance; it may help prevent overfitting but will not enable the model to learn from the rare failure class when it is vastly underrepresented in the training data.

Why this answer

How to eliminate wrong answers

Option A is wrong because it reverses the definitions: extractive summarization copies existing sentences, not produces new words, and abstractive summarization generates new sentences, not copies sentences. Option C is wrong because abstractive and extractive summarization produce fundamentally different outputs—abstractive creates novel phrasing, while extractive outputs verbatim excerpts—not identical results. Option D is wrong because abstractive summarization is not limited to legal documents; it works across various domains, and extractive summarization is also used for general text, not exclusively for general text.

Why this answer

How to eliminate wrong answers

Option A is wrong because supervised learning requires labeled input-output pairs, whereas game-playing agents learn from rewards, not from pre-labeled correct moves. Option C is wrong because clustering is an unsupervised learning technique that groups similar data points without any reward signal, not suitable for training an agent to play games. Option D is wrong because regression predicts continuous numerical values (e.g., final score), but it does not involve a reward-based learning loop where the agent takes actions to maximize rewards.

Why this answer

How to eliminate wrong answers

Option A is wrong because it describes color depth (bits per pixel), which is a property of image encoding, not a computer vision technique for measuring spatial distance. Option C is wrong because it confuses depth estimation with semantic segmentation or object detection in cluttered scenes; 'depth' here refers to physical distance, not how deeply a subject is embedded in a background. Option D is wrong because it refers to photographic detail determined by lens quality (optical resolution), which is unrelated to the algorithmic inference of object distances from camera data.

Why this answer

How to eliminate wrong answers

Option A is wrong because Temperature controls the randomness of token selection by scaling logits before softmax, not content safety. Option B is wrong because Top-p (nucleus sampling) selects from the smallest set of tokens whose cumulative probability exceeds p, affecting output diversity, not filtering harmful content. Option C is wrong because the system message sets the assistant's behavior and tone via instructions, but it cannot enforce content safety rules; it relies on the model's adherence and does not provide a hard filter.