Practice Exam

B) Supervised learning - regression. The company has labeled historical data (input: dates/prices, output: quantities) and wants to predict a continuous number (units sold). This is regression. Classification (A) predicts categories. Clustering (C) requires no labels. RL (D) involves agents and rewards.

C) Network configuration. While infrastructure is needed, network configuration is not a phase of the ML lifecycle. The lifecycle consists of: problem definition, data collection/preparation, feature engineering, model training, evaluation, deployment, and monitoring/maintenance.

B) Overfitting. High training accuracy but low test accuracy is the textbook definition of overfitting. The model memorized the training data instead of learning general patterns. Solutions include regularization, dropout, more training data, or a simpler model.

C) Recall. When missing positives (false negatives) is costly, optimize for recall. Recall measures what percentage of actual positives the model catches. For fraud detection, missing a fraud (false negative) is worse than flagging a legitimate transaction (false positive). Accuracy (A) is misleading for imbalanced data. Precision (B) minimizes false positives. RMSE (D) is for regression.

B) A subset of machine learning using multi-layered neural networks. Deep learning uses neural networks with multiple hidden layers ("deep" networks) to learn complex patterns. It excels at unstructured data (images, text, audio), not only structured data (D). It is not limited to RL (C) or defined by the Turing test (A).

D) Reinforcement learning. The robot (agent) interacts with the warehouse (environment) and learns from rewards and penalties. This is the defining pattern of RL: agent + environment + rewards. AWS DeepRacer is a similar RL use case.

B) Few-shot prompting. Providing examples (shots) in the prompt to show the model the expected output format is few-shot prompting. 3 examples = few-shot. Zero examples = zero-shot (A). Chain-of-thought (C) adds "think step by step" reasoning. Fine-tuning (D) changes the model itself, not the prompt.

C) RAG can be updated with new information without retraining the model. The key advantage of RAG is that you simply update the knowledge base (add/remove documents), and the model immediately has access to the new information. Fine-tuning requires retraining. RAG does need an FM (B), does not guarantee higher quality (A), and reduces but does not eliminate hallucination (D).

C) Guardrails. Bedrock Guardrails includes a "denied topics" feature that prevents the model from responding about specific subjects you define. Knowledge Bases (A) is for RAG. Agents (B) is for multi-step task automation. Model evaluation (D) compares model performance.

C) Temperature = 0.8. Higher temperature (0.7-1.0) increases randomness and creativity in the output, which is ideal for marketing copy. Low temperature (A, B) produces deterministic, factual output. Low top-p with low temperature (D) would make output even more constrained.

B) A large model pre-trained on broad data that can be adapted to many tasks. Foundation models are trained on massive, diverse datasets and serve as a "foundation" that can be adapted (via prompting, fine-tuning, or RAG) for many different downstream tasks. They are general-purpose, not task-specific (A, C), and use ML, not rules (D).

B) To represent text as numerical vectors for similarity search. In RAG, documents are converted to vector embeddings (numerical representations) so that when a user asks a question, the system can find the most semantically similar document chunks using vector similarity search. Amazon Titan Embeddings and Amazon Bedrock Knowledge Bases handle this process.

C) Fine-tuning with domain-specific data. When the model needs to adopt a specific writing style or terminology that differs from its training, fine-tuning is appropriate. RAG (B) adds knowledge but does not change the model's style. Zero-shot (A) uses the model as-is. Temperature (D) affects randomness, not style.

B) Amazon Comprehend. Comprehend provides sentiment analysis that classifies text as positive, negative, neutral, or mixed. It is the go-to service for analyzing text meaning. Rekognition (A) is for images. Translate (C) is for language translation. Lex (D) is for building chatbots.

B) Amazon Transcribe. Transcribe converts speech to text and supports speaker identification (speaker diarization). Polly (A) does the opposite: text to speech. Comprehend (C) analyzes text meaning. Translate (D) translates between languages.

C) Amazon SageMaker. SageMaker is the comprehensive ML platform for building custom models. The team wants full control over the ML process (model selection, training, deployment), which is exactly what SageMaker provides. Bedrock (A) is for generative AI. Comprehend (B) is for NLP. Rekognition (D) is for images.

B) Amazon Polly. Polly converts text to lifelike spoken audio (text-to-speech). This is the correct service for accessibility features that read text aloud. Transcribe (A) does the opposite: speech to text. Lex (C) is for chatbots. Translate (D) is for language translation.

B) Amazon Comprehend Medical. Comprehend Medical is specifically designed to extract medical entities (medications, dosages, conditions, procedures) from clinical text. Textract (A) extracts text from scanned documents but does not understand medical terminology. Rekognition (C) analyzes images. Translate (D) translates languages.

C) SageMaker Canvas. Canvas is the no-code ML interface designed for business users. It provides visual, point-and-click model building. Studio (A) is a full IDE for data scientists. Autopilot (B) automates ML but still requires some technical knowledge. JumpStart (C) provides pre-trained models but is accessed through Studio.

B) Amazon Personalize. Personalize is the recommendation engine that uses user interaction data (viewing history, ratings) to provide personalized recommendations. Kendra (A) is for enterprise search. Comprehend (C) is for text analysis. Forecast (D) is for time-series prediction.

B) Amazon Forecast. Forecast is the time-series forecasting service designed for exactly this use case: predicting future values based on historical time-series data. Personalize (A) is for recommendations. Comprehend (C) is for NLP. Bedrock (D) is for generative AI.

C) Amazon Rekognition. Rekognition provides face comparison capabilities that can compare a face in one image against a collection of faces in another database. This is exactly what identity verification requires. Textract (A) is for documents. Comprehend (B) is for text. Bedrock (D) is for generative AI.

A) Amazon Kendra. Kendra is the intelligent enterprise search service that understands natural language queries and finds answers across documents, wikis, and knowledge bases. Textract (B) extracts text from scanned documents. Translate (C) translates languages. Polly (D) converts text to speech.

B) Amazon Bedrock. Bedrock is the fully managed service that provides access to foundation models from multiple providers through a unified API. SageMaker JumpStart also offers some FMs, but the "single API, multiple providers" description specifically matches Bedrock.

B) Bias / fairness issue. When a model produces systematically different outcomes for different groups despite similar qualifications, this is a bias/fairness problem. The model may have learned biased patterns from historical data. SageMaker Clarify can detect this type of disparate impact.

B) Amazon SageMaker Clarify. Clarify provides model explainability through SHAP values, which show how much each feature contributed to a specific prediction. Model Monitor (A) detects data drift. Guardrails (C) filters generative AI content. CloudWatch (D) monitors infrastructure metrics.

B) Guardrails with content filters. Bedrock Guardrails allows you to configure content filters that block harmful content categories including violence, hate speech, sexual content, and more. You can also define denied topics for additional specificity.

B) Prompt injection. Prompt injection is when an attacker crafts inputs designed to override the model's system prompt and instructions. The input explicitly tries to make the model "ignore previous instructions." Mitigations include input validation, Bedrock Guardrails, and robust system prompts.

A) Amazon SageMaker Model Cards. Model Cards provide a structured framework for documenting ML models, including their intended use, limitations, performance metrics, and bias evaluation results. This supports governance and transparency. CloudFormation (B) manages infrastructure. S3 (C) stores files. CloudTrail (D) logs API calls.

C) Data drift. Data drift occurs when the statistical properties of production data change over time compared to the training data. Customer behavior evolving is a classic example. The model was trained on historical patterns that no longer apply. SageMaker Model Monitor detects data drift automatically.