AWS AI Practitioner (AIF-C01) — Practice Questions and Mock Exam

Amazon EMR (Elastic MapReduce) is the correct answer because it is AWS's managed big data platform specifically designed to run distributed processing frameworks like Apache Spark. As stated in the documentation, Amazon EMR provides features and functionality for "processing and analyzing big data on AWS." EMR natively supports Apache Spark and its MLlib library, handling cluster provisioning, configuration, and scaling so that data science teams can focus on building and training distributed ML models.

Why each incorrect option is wrong:

• A. AWS Glue — While AWS Glue does use Apache Spark under the hood for ETL (Extract, Transform, Load) jobs, it is primarily designed for data integration and ETL workloads, not for running full Spark MLlib distributed training workflows. It does not expose the same level of cluster control and Spark configuration flexibility that EMR provides for ML training.

• C. Amazon ECS — Amazon Elastic Container Service (ECS) is a container orchestration service for running Docker containers. While you could theoretically containerize Spark, ECS does not provide native managed Spark cluster provisioning or built-in support for Spark MLlib. It is not designed for distributed Spark workloads.

• D. AWS Batch — AWS Batch is designed for running batch computing jobs at scale, typically for HPC or simple parallel workloads. It does not natively support Apache Spark frameworks or provide managed Spark cluster provisioning needed for Spark MLlib distributed training.

(Source: https://docs.aws.amazon.com/emr/latest/ManagementGuide/emr-what-is-emr.html)

Sensitive information filters is the correct answer. Amazon Bedrock Guardrails provides sensitive information filters that can detect and redact personally identifiable information (PII) such as Social Security numbers, credit card numbers, and email addresses from both user inputs and model responses. This capability is essential for compliance with privacy regulations when building customer-facing generative AI applications.

According to AWS documentation, Amazon Bedrock Guardrails allows you to "detect and filter harmful content" and one of the key filter types is sensitive information filters, which specifically targets PII detection and redaction to protect end-users from exposure to sensitive data.

Why each incorrect option is wrong:

• A. Content filters — Content filters are designed to detect and block harmful content across categories such as hate speech, insults, sexual content, violence, and misconduct. They do not specifically target PII like Social Security numbers or credit card numbers.

• B. Denied topics — Denied topics allow you to define specific topics that the model should not engage with (e.g., investment advice, competitor discussions). This is a topic-level restriction, not a mechanism for detecting or redacting specific PII data patterns.

• D. Contextual grounding checks — Contextual grounding checks are used to detect hallucinations by verifying whether model responses are grounded in the provided source material and are relevant to the user's query. They have nothing to do with PII detection or redaction.

(Source: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html)

Data augmentation is the correct answer. Data augmentation is a technique that creates modified versions of existing training images — such as rotations, flips, color adjustments, and crops — to artificially increase the effective size and diversity of the training dataset. This is particularly valuable when limited additional training data is available, as it helps the model generalize better by exposing it to realistic variations of existing samples. A rotated or flipped image of a cat is still a cat, so the model learns to recognize objects regardless of orientation, scale, or lighting conditions.

Why the other options are incorrect:

• A. Regularization — Regularization is a technique used to prevent overfitting by adding a penalty term to the model's loss function (e.g., L1/L2 regularization) or by using methods like dropout. It does not involve creating modified versions of training images; rather, it constrains the model's complexity.

• C. Feature engineering — Feature engineering involves selecting, transforming, or creating input features to improve model performance. While it manipulates the representation of data, it does not involve creating modified copies of training images through geometric or color transformations.

• D. Transfer learning — Transfer learning involves taking a model pre-trained on a large dataset (e.g., ImageNet) and fine-tuning it on a smaller, task-specific dataset. While it is useful when training data is limited, it does not involve creating modified versions of existing training images.

Data augmentation is a standard and widely-used technique in computer vision that directly addresses the problem of limited training data by generating realistic variations of existing samples.

(Source: https://docs.aws.amazon.com/prescriptive-guidance/latest/image-classification/examples.html)

Chain-of-thought (CoT) prompting is the correct technique for instructing a model to reason through complex problems step by step before arriving at a final answer. AWS documentation describes chain of thought reasoning as taking a larger, complex task and breaking it down into smaller, simpler steps, and notes that it can often improve model accuracy and make the reasoning process more transparent. This technique involves either explicitly instructing the model to 'think step by step' or providing examples that demonstrate intermediate reasoning steps. It is most useful for tasks such as multi-step analysis, math problems, and complex reasoning.

Why each incorrect option is wrong:

• A. Negative prompting — Negative prompting is a technique where you tell the model what NOT to do or generate (e.g., 'Do not include personal opinions'). It constrains outputs but does not encourage step-by-step reasoning.

• B. Zero-shot prompting — Zero-shot prompting provides a task instruction without any examples, relying entirely on the model's pre-trained knowledge. While it can be combined with CoT (zero-shot CoT), on its own it does not specifically encourage the model to show intermediate reasoning steps.

• D. Prompt caching — Prompt caching is an optimization feature that caches reusable prompt prefixes to reduce inference response latency and input token costs. It is designed for efficiency, not to make the model reason through problems step by step.

(Source: https://docs.aws.amazon.com/bedrock/latest/userguide/inference-reasoning.html; https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html)